Daily Papers

Evaluations of LLM systems routinely average over small prompt sets and report models as a ranked table. We ask how much confidence such a table deserves, using LLM-based prompt-structure inference as the case study: eight open model variants across five families and 8B to 675B parameters, caching disabled, 293 raw intermediate representations persisted. The measured phenomenon is unstable to begin with. Identical calls do not reliably recover identical structure, with mean node-set Jaccard from 0.39 to 0.96 and 72% of prompt-model cells never node-set-perfect. Auditing the evaluation weakens its conclusions further, and this is our main contribution. Under a joint cluster bootstrap over prompts, only the bottom of the ranking is firm: the two least reproducible models hold rank in 99% and 86% of replicates, the middle four in 27% to 48%, and the top two in 68% each, so the table identifies the worst model reliably but does not reliably identify the best. Two equally defensible rules for merging repeated campaigns change four of eight rows and move the study-wide headline by 7 percentage points. Checking the inferred structure against ground-truth annotations shows reproducibility cannot be read as accuracy. And four of the eight endpoints were withdrawn within ten weeks of measurement, so the study as specified can no longer be run. Small-sample LLM evaluations can therefore look far more definitive than their evidence supports. We recommend reporting rank stability, per-cell provenance, executed sensitivity comparisons, raw per-run outputs, and a measurement date alongside any ranking.

Catastrophic regime shifts in complex natural systems may be averted through advanced detection. Recent work has provided a proof-of-principle that many systems approaching a catastrophic transition may be identified through the lens of early warning indicators such as rising variance or increased return times. Despite widespread appreciation of the difficulties and uncertainty involved in such forecasts, proposed methods hardly ever characterize their expected error rates. Without the benefits of replicates, controls, or hindsight, applications of these approaches must quantify how reliable different indicators are in avoiding false alarms, and how sensitive they are to missing subtle warning signs. We propose a model based approach in order to quantify this trade-off between reliability and sensitivity and allow comparisons between different indicators. We show these error rates can be quite severe for common indicators even under favorable assumptions, and also illustrate how a model-based indicator can improve this performance. We demonstrate how the performance of an early warning indicator varies in different data sets, and suggest that uncertainty quantification become a more central part of early warning predictions.

LLM agents choose tools and arguments from context that mixes user requests, tool outputs, retrieved records, memory, and untrusted text. Evidence can be relevant without being authorized to determine a decision, so a correct action need not be grounded only in permitted evidence. We introduce a target-specific authorization audit that labels context factors separately for each tool and argument target. Its primary test holds the task, proposition, position, and policy fixed while changing only the proposition's source authority. We then test behavior when valid evidence is weakened and use context-subset interactions as a secondary localization diagnostic. Across 450 controlled next-action tasks and multiple open-weight LLM families, trusted and untrusted variants produce different actions in 5.4 percent of competing cases versus 1.7 percent of supporting cases. Under controlled degradation, unauthorized competition is retained in a full-correct, mixed-error, clean-correct pattern in 2.4 percent of comparisons, with a 95 percent confidence interval from 2.1 to 3.0 percent. These are controlled stress-set rates, not deployment prevalence. The models respond to textual source-authority cues, but this does not prevent untrusted evidence from influencing their actions.

Outer-learning algorithms use infinitesimal sensitivities to propose finite changes to initialization or training parameters. For hard-ReLU training, the derivative of the finite program and the derivative of its flow limit do not by themselves specify the response at a chosen radius. We characterize the intervening regime in which the perturbation radius is proportional to the GD step. Integer event rounding then survives at leading order: smooth Euler bias shifts each discrete phase, and upstream rounding moves downstream branch boundaries. We derive the crossing indices and a uniform endpoint expansion for finitely many separated transverse events in piecewise-C^2 dynamics, away from recursive phase boundaries. In contractive affine regions, an explicit remainder and complete branch verification certify finite candidate comparisons. Scalar phase frequencies and a coupled feedback ablation test the mechanism; frozen nonlinear-network experiments show radius-dependent prediction accuracy, including incomplete branch matches and failed-word tails. Together with local AD and uniform flow consistency, the result identifies sufficient response regimes: differentiating training is a choice of perturbation resolution as well as a choice of derivative.

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as 12%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

Graph neural networks (GNNs) predict molecular properties from chemical graph data, but predictive accuracy does not explain how graph information supports an individual decision. A compact prediction-preserving rationale does not necessarily reveal which changes reverse the decision or which modifications the model tolerates. We propose the Multi-Perspective Graph Explainer (MPGE), unifying factual support, counterfactual sensitivity, and exemplar tolerance for a frozen classifier. The factual view, originally termed prototype (PT), seeks a compact retained edge set with the same label and required confidence. Counterfactual (CF) explanations seek bounded prediction-changing deletions; exemplar (EXE) explanations seek non-trivial bounded deletions that preserve the label and confidence. A shared constrained formulation connects prediction behavior, compactness, and edit cost, while separate objectives generate the three views. Our graph-classification extension of CF-GNNExplainer learns symmetric edge rankings and verifies discrete candidates, recording unsuccessful searches. A separate BBBP fragment backend returns RDKit-sanitized molecules. We evaluate the primary GCN implementation on MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP using semantic coverage, conditional quality, stability, and runtime. Successful factual masks retained 8.6%--15.5% of input edges on average across datasets; bounded counterfactual coverage was 4.8%--67.6%, and exemplar preservation coverage was 98.9%--100.0%. Exploratory controls reveal the influence of hard projection and retained node information. Quantitative comparisons and molecular visualizations characterize model support, sensitivity, and tolerance without treating them as validated chemical mechanisms.

Radiological imaging is central to diagnosis, treatment planning, and clinical decision-making. Vision-language foundation models have spurred interest in automated radiology report generation (RRG), but safe deployment requires reliable clinical evaluation of generated reports. Existing metrics often rely on surface-level similarity or behave as black boxes, lacking interpretability. We introduce ICARE (Interpretable and Clinically-grounded Agent-based Report Evaluation), an interpretable evaluation framework leveraging large language model agents and dynamic multiple-choice question answering (MCQA). Two agents, each with either the ground-truth or generated report, generate clinically meaningful questions and quiz each other. Agreement on answers captures preservation and consistency of findings, serving as interpretable proxies for clinical precision and recall. By linking scores to question-answer pairs, ICARE enables transparent, and interpretable assessment. Clinician studies show ICARE aligns significantly more with expert judgment than prior metrics. Perturbation analyses confirm sensitivity to clinical content and reproducibility, while model comparisons reveal interpretable error patterns.

Modern Text-to-Image (T2I) Diffusion models have revolutionized image editing by enabling the generation of high-quality photorealistic images. While the de facto method for performing edits with T2I models is through text instructions, this approach non-trivial due to the complex many-to-many mapping between natural language and images. In this work, we address exemplar-based image editing -- the task of transferring an edit from an exemplar pair to a content image(s). We propose ReEdit, a modular and efficient end-to-end framework that captures edits in both text and image modalities while ensuring the fidelity of the edited image. We validate the effectiveness of ReEdit through extensive comparisons with state-of-the-art baselines and sensitivity analyses of key design choices. Our results demonstrate that ReEdit consistently outperforms contemporary approaches both qualitatively and quantitatively. Additionally, ReEdit boasts high practical applicability, as it does not require any task-specific optimization and is four times faster than the next best baseline.

Large Language Models (LLMs) are increasingly used for automated fact-checking, yet their susceptibility to gender bias in this context remains underexplored. This study presents the first systematic investigation of gender bias in LLM-based fake news detection using real-world data. We augment the LIAR benchmark with three gender variants of speaker job titles (Neutral, Male, Female) for each statement to test whether veracity judgments vary solely based on gender presentation. Six state-of-the-art LLMs are evaluated across multiple bias and fairness metrics. All models exhibit gender sensitivity: 9.79%-35.13% of statements receive inconsistent labels across the three variants, with Male-Female comparisons showing 6.5%-23.6% flip rates. Two primary bias manifestations are identified: instability (inconsistent judgments) and directionality (systematic favoritism). Five models show statistically significant directional effects, with the strongest effects displaying male-skeptic patterns. These findings demonstrate that gender bias undermines both reliability and fairness in LLM-based fake news detection, highlighting the need for bias-aware evaluation and mitigation strategies. The augmented dataset is publicly released to support future research.

Effective evaluation of language models remains an open challenge in NLP. Researchers and engineers face methodological issues such as the sensitivity of models to evaluation setup, difficulty of proper comparisons across methods, and the lack of reproducibility and transparency. In this paper we draw on three years of experience in evaluating large language models to provide guidance and lessons for researchers. First, we provide an overview of common challenges faced in language model evaluation. Second, we delineate best practices for addressing or lessening the impact of these challenges on research. Third, we present the Language Model Evaluation Harness (lm-eval): an open source library for independent, reproducible, and extensible evaluation of language models that seeks to address these issues. We describe the features of the library as well as case studies in which the library has been used to alleviate these methodological concerns.

Classical robot ethics is often framed around obedience, most famously through Asimov's laws. This framing is too narrow for contemporary AI systems, which are adaptive, generative, embodied, and embedded in physical, psychological, and social worlds. We argue that future human-AI relations should be understood not as master-tool obedience, but as conditional mutualism under governance: a co-evolutionary relationship in which humans and AI systems can develop, specialize, and coordinate while institutions keep the relation reciprocal, reversible, psychologically safe, and socially legitimate. We synthesize concepts from computability, machine learning, foundation models, embodied AI, alignment, human-robot interaction, ecological mutualism, coevolution, and polycentric governance. We then formalize coexistence as a multiplex dynamical system across physical, psychological, and social layers, with reciprocal supply-demand coupling, conflict penalties, developmental freedom, and governance regularization. The model gives conditions for existence, uniqueness, and global asymptotic stability of equilibria. Deterministic ODE simulations, basin sweeps, sensitivity analyses, governance-regime comparisons, shock tests, and local stability checks show that governed mutualism reaches high coexistence with zero domination, while absent or excessive governance can produce domination, weak-benefit lock-in, or suppressed development. The results suggest that human-AI coexistence should be designed as a co-evolutionary governance problem, not a one-shot obedience problem.

Understanding how neural networks rely on visual cues offers a human-interpretable view of their internal decision processes. The cue-conflict benchmark has been influential in probing shape-texture preference and in motivating the insight that stronger, human-like shape bias is often associated with improved in-domain performance. However, we find that the current stylization-based instantiation can yield unstable and ambiguous bias estimates. Specifically, stylization may not reliably instantiate perceptually valid and separable cues nor control their relative informativeness, ratio-based bias can obscure absolute cue sensitivity, and restricting evaluation to preselected classes can distort model predictions by ignoring the full decision space. Together, these factors can confound preference with cue validity, cue balance, and recognizability artifacts. We introduce REFINED-BIAS, an integrated dataset and evaluation framework for reliable and interpretable shape-texture bias diagnosis. REFINED-BIAS constructs balanced, human- and model- recognizable cue pairs using explicit definitions of shape and texture, and measures cue-specific sensitivity over the full label space via a ranking-based metric, enabling fairer cross-model comparisons. Across diverse training regimes and architectures, REFINED-BIAS enables fairer cross-model comparison, more faithful diagnosis of shape and texture biases, and clearer empirical conclusions, resolving inconsistencies that prior cue-conflict evaluations could not reliably disambiguate.

Training algorithms, broadly construed, are an essential part of every deep learning pipeline. Training algorithm improvements that speed up training across a wide variety of workloads (e.g., better update rules, tuning protocols, learning rate schedules, or data selection schemes) could save time, save computational resources, and lead to better, more accurate, models. Unfortunately, as a community, we are currently unable to reliably identify training algorithm improvements, or even determine the state-of-the-art training algorithm. In this work, using concrete experiments, we argue that real progress in speeding up training requires new benchmarks that resolve three basic challenges faced by empirical comparisons of training algorithms: (1) how to decide when training is complete and precisely measure training time, (2) how to handle the sensitivity of measurements to exact workload details, and (3) how to fairly compare algorithms that require hyperparameter tuning. In order to address these challenges, we introduce a new, competitive, time-to-result benchmark using multiple workloads running on fixed hardware, the AlgoPerf: Training Algorithms benchmark. Our benchmark includes a set of workload variants that make it possible to detect benchmark submissions that are more robust to workload changes than current widely-used methods. Finally, we evaluate baseline submissions constructed using various optimizers that represent current practice, as well as other optimizers that have recently received attention in the literature. These baseline results collectively demonstrate the feasibility of our benchmark, show that non-trivial gaps between methods exist, and set a provisional state-of-the-art for future benchmark submissions to try and surpass.

A central challenge for the task of semantic segmentation is the prohibitive cost of obtaining dense pixel-level annotations to supervise model training. In this work, we show that in order to achieve a good level of segmentation performance, all you need are a few well-chosen pixel labels. We make the following contributions: (i) We investigate the novel semantic segmentation setting in which labels are supplied only at sparse pixel locations, and show that deep neural networks can use a handful of such labels to good effect; (ii) We demonstrate how to exploit this phenomena within an active learning framework, termed PixelPick, to radically reduce labelling cost, and propose an efficient "mouse-free" annotation strategy to implement our approach; (iii) We conduct extensive experiments to study the influence of annotation diversity under a fixed budget, model pretraining, model capacity and the sampling mechanism for picking pixels in this low annotation regime; (iv) We provide comparisons to the existing state of the art in semantic segmentation with active learning, and demonstrate comparable performance with up to two orders of magnitude fewer pixel annotations on the CamVid, Cityscapes and PASCAL VOC 2012 benchmarks; (v) Finally, we evaluate the efficiency of our annotation pipeline and its sensitivity to annotator error to demonstrate its practicality.

Large language models (LLMs) have fundamentally transformed automated software development by enabling direct translation of natural language descriptions into functional code, driving commercial adoption through tools like Github Copilot (Microsoft), Cursor (Anysphere), Trae (ByteDance), and Claude Code (Anthropic). While the field has evolved dramatically from rule-based systems to Transformer-based architectures, achieving performance improvements from single-digit to over 95\% success rates on benchmarks like HumanEval. In this work, we provide a comprehensive synthesis and practical guide (a series of analytic and probing experiments) about code LLMs, systematically examining the complete model life cycle from data curation to post-training through advanced prompting paradigms, code pre-training, supervised fine-tuning, reinforcement learning, and autonomous coding agents. We analyze the code capability of the general LLMs (GPT-4, Claude, LLaMA) and code-specialized LLMs (StarCoder, Code LLaMA, DeepSeek-Coder, and QwenCoder), critically examining the techniques, design decisions, and trade-offs. Further, we articulate the research-practice gap between academic research (e.g., benchmarks and tasks) and real-world deployment (e.g., software-related code tasks), including code correctness, security, contextual awareness of large codebases, and integration with development workflows, and map promising research directions to practical needs. Last, we conduct a series of experiments to provide a comprehensive analysis of code pre-training, supervised fine-tuning, and reinforcement learning, covering scaling law, framework selection, hyperparameter sensitivity, model architectures, and dataset comparisons.

Large Language Models (LLMs) changed the way we design and interact with software systems. Their ability to process and extract information from text has drastically improved productivity in a number of routine tasks. Developers that want to include these models in their software stack, however, face a dreadful challenge: debugging LLMs' inconsistent behavior across minor variations of the prompt. We therefore introduce two metrics for classification tasks, namely sensitivity and consistency, which are complementary to task performance. First, sensitivity measures changes of predictions across rephrasings of the prompt, and does not require access to ground truth labels. Instead, consistency measures how predictions vary across rephrasings for elements of the same class. We perform an empirical comparison of these metrics on text classification tasks, using them as guideline for understanding failure modes of the LLM. Our hope is that sensitivity and consistency will be helpful to guide prompt engineering and obtain LLMs that balance robustness with performance.

Large language models (LLMs) now write code in settings where misreading a single word can break safety or cost money, yet we still expect them to overlook stray typos. To probe where useful robustness ends and harmful insensitivity begins, we compile 50 LeetCode problems and craft three minimal prompt perturbations that should vary in importance: (i) progressive underspecification deleting 10 % of words per step; (ii) lexical flip swapping a pivotal quantifier ("max" to "min"); and (iii) jargon inflation replacing a common noun with an obscure technical synonym. Six frontier models, including three "reasoning-tuned" versions, solve each mutated prompt, and their Python outputs are checked against the original test suites to reveal whether they reused the baseline solution or adapted. Among 11 853 generations we observe a sharp double asymmetry. Models remain correct in 85 % of cases even after 90 % of the prompt is missing, showing over-robustness to underspecification, yet only 54 % react to a single quantifier flip that reverses the task, with reasoning-tuned variants even less sensitive than their bases. Jargon edits lie in between, passing through 56 %. Current LLMs thus blur the line between harmless noise and meaning - changing edits, often treating both as ignorable. Masking salient anchors such as function names can force re - evaluation. We advocate evaluation and training protocols that reward differential sensitivity: stay steady under benign noise but adapt - or refuse - when semantics truly change.

Large language models (LLMs) have demonstrated impressive capabilities across various tasks, but their performance is highly sensitive to the prompts utilized. This variability poses challenges for accurate assessment and user satisfaction. Current research frequently overlooks instance-level prompt variations and their implications on subjective evaluations. To address these shortcomings, we introduce ProSA, a framework designed to evaluate and comprehend prompt sensitivity in LLMs. ProSA incorporates a novel sensitivity metric, PromptSensiScore, and leverages decoding confidence to elucidate underlying mechanisms. Our extensive study, spanning multiple tasks, uncovers that prompt sensitivity fluctuates across datasets and models, with larger models exhibiting enhanced robustness. We observe that few-shot examples can alleviate this sensitivity issue, and subjective evaluations are also susceptible to prompt sensitivities, particularly in complex, reasoning-oriented tasks. Furthermore, our findings indicate that higher model confidence correlates with increased prompt robustness. We believe this work will serve as a helpful tool in studying prompt sensitivity of LLMs. The project is released at: https://github.com/open-compass/ProSA .

Prompt sensitivity, referring to the phenomenon where paraphrasing (i.e., repeating something written or spoken using different words) leads to significant changes in large language model (LLM) performance, has been widely accepted as a core limitation of LLMs. In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of LLMs, or is it largely an artifact of evaluation processes? To answer this question, we systematically evaluate 7 LLMs (e.g., GPT and Gemini family) across 6 benchmarks, including both multiple-choice and open-ended tasks on 12 diverse prompt templates. We find that much of the prompt sensitivity stems from heuristic evaluation methods, including log-likelihood scoring and rigid answer matching, which often overlook semantically correct responses expressed through alternative phrasings, such as synonyms or paraphrases. When we adopt LLM-as-a-Judge evaluations, we observe a substantial reduction in performance variance and a consistently higher correlation in model rankings across prompts. Our findings suggest that modern LLMs are more robust to prompt templates than previously believed, and that prompt sensitivity may be more an artifact of evaluation than a flaw in the models.

Removing an input-scaling module changes both a classifier and the perturbations reaching its encoder. A robustness difference can therefore reflect the comparison rule as well as the module. We demonstrate this problem in a four-qubit quantum-attention detector on generated power-grid trajectories. A learned scaling module appears beneficial at a fixed physical attack budget, but matching an upper bound on perturbations at the encoder reverses the ordering. Neither comparison alone establishes a robustness benefit caused by the module. The initial test also perturbs clean examples into attacked examples while retaining their original labels; tests restricted to already attacked examples do not establish a benefit. Replacing a trained model's input scales disrupts detection. Retraining its linear classification layer restores the detection rate, but changes individual predictions, leaving the comparison descriptive rather than causal. Two further design checks explain why the input quantum Fisher information regularizer cannot train this model's query parameters, and why removing confidence bounds does not establish a larger certified radius. The evidence is limited to ten seeds, exact simulation, synthetic data, and a restricted set of attacks; classical baselines achieve better clean prediction. The practical lesson is to specify which perturbation budget is fixed, check that attacks preserve labels and interventions preserve predictions, and distinguish exploratory controls from confirmatory evidence.

The rise of machine-learning systems that process sensory input has brought with it a rise in comparisons between human and machine perception. But such comparisons face a challenge: Whereas machine perception of some stimulus can often be probed through direct and explicit measures, much of human perceptual knowledge is latent, incomplete, or unavailable for explicit report. Here, we explore how this asymmetry can cause such comparisons to misestimate the overlap in human and machine perception. As a case study, we consider human perception of adversarial speech -- synthetic audio commands that are recognized as valid messages by automated speech-recognition systems but that human listeners reportedly hear as meaningless noise. In five experiments, we adapt task designs from the human psychophysics literature to show that even when subjects cannot freely transcribe such speech commands (the previous benchmark for human understanding), they often can demonstrate other forms of understanding, including discriminating adversarial speech from closely matched non-speech (Experiments 1--2), finishing common phrases begun in adversarial speech (Experiments 3--4), and solving simple math problems posed in adversarial speech (Experiment 5) -- even for stimuli previously described as unintelligible to human listeners. We recommend the adoption of such "sensitive tests" when comparing human and machine perception, and we discuss the broader consequences of such approaches for assessing the overlap between systems.

Recent work found that LLMs are sensitive to a wide range of arbitrary prompt dimensions, including the type of delimiters, answer enumerators, instruction wording, and more. This throws into question popular single-prompt evaluation practices. We present DOVE (Dataset Of Variation Evaluation) a large-scale dataset containing prompt perturbations of various evaluation benchmarks. In contrast to previous work, we examine LLM sensitivity from an holistic perspective, and assess the joint effects of perturbations along various dimensions, resulting in thousands of perturbations per instance. We evaluate several model families against DOVE, leading to several findings, including efficient methods for choosing well-performing prompts, observing that few-shot examples reduce sensitivity, and identifying instances which are inherently hard across all perturbations. DOVE consists of more than 250M prompt perturbations and model outputs, which we make publicly available to spur a community-wide effort toward meaningful, robust, and efficient evaluation. Browse the data, contribute, and more: https://slab-nlp.github.io/DOVE/

Mixed-precision quantization must decide which parts of a model to keep at higher precision. A common premise, shared by sensitivity-based methods such as HAWQ and CoopQ, is that the loss from quantizing a set of layers can be reconstructed from per-layer or pairwise sensitivities measured in isolation. We test this premise at the 4-bit weight-and-activation precisions now being deployed, treating the change in loss f(S) from quantizing a layer set S as a set function on the Boolean cube and analyzing it through two classical changes of basis. This analysis yields two findings. First, across configurations drawn from the deployment distribution, 85--93\% of the variance of f is explained by per-layer effects alone. Second, a monotone transform of a sum of per-layer terms reproduces f's ranking of configurations, misordering at most 2\% of pairs. We propose the coverage model f(S)=cbigl(1-prod_{iin S}(1-a_i)bigr), which reproduces the measured variance profile of f to within a few percent from its L fitted break-rates. This structure supports two predictors of a configuration's loss, each with L+1 parameters. The additive model is the optimal first-order predictor. By Parseval's identity its mean-squared error equals the variance of f left unexplained by per-layer effects, which we measure on full lattices, estimate out of sample at full-network scale, and report with every result as a certificate of how well any additive model can do. The coverage model itself is the second predictor. As allocators at matched memory, they attain the lowest KL divergence among the compared allocators on models from 30B to 355B parameters. Below four bits, the resulting allocations continue to solve code and reasoning tasks at budgets where allocations from gradient sensitivities no longer produce terminating generations.

Despite their remarkable capabilities, Large Language Models (LLMs) are found to be surprisingly sensitive to minor variations in prompts, often generating significantly divergent outputs in response to minor variations in the prompts, such as spelling errors, alteration of wording or the prompt template. However, while assessing the quality of an LLM, the focus often tends to be solely on its performance on downstream tasks, while very little to no attention is paid to prompt sensitivity. To fill this gap, we propose POSIX - a novel PrOmpt Sensitivity IndeX as a reliable measure of prompt sensitivity, thereby offering a more comprehensive evaluation of LLM performance. The key idea behind POSIX is to capture the relative change in loglikelihood of a given response upon replacing the corresponding prompt with a different intent-preserving prompt. We provide thorough empirical evidence demonstrating the efficacy of POSIX in capturing prompt sensitivity and subsequently use it to measure and thereby compare prompt sensitivity of various open-source LLMs. We find that merely increasing the parameter count or instruction tuning does not necessarily reduce prompt sensitivity whereas adding some few-shot exemplars, even just one, almost always leads to significant decrease in prompt sensitivity. We also find that alterations to prompt template lead to the highest sensitivity in the case of MCQ type tasks, whereas paraphrasing results in the highest sensitivity in open-ended generation tasks. The code for reproducing our results is open-sourced at https://github.com/kowndinya-renduchintala/POSIX.

Prompt sensitivity, which refers to how strongly the output of a large language model (LLM) depends on the exact wording of its input prompt, raises concerns among users about the LLM's stability and reliability. In this work, we consider LLMs as multivariate functions and perform a first-order Taylor expansion, thereby analyzing the relationship between meaning-preserving prompts, their gradients, and the log probabilities of the model's next token. We derive an upper bound on the difference between log probabilities using the Cauchy-Schwarz inequality. We show that LLMs do not internally cluster similar inputs like smaller neural networks do, but instead disperse them. This dispersing behavior leads to an excessively high upper bound on the difference of log probabilities between two meaning-preserving prompts, making it difficult to effectively reduce to 0. In our analysis, we also show which types of meaning-preserving prompt variants are more likely to introduce prompt sensitivity risks in LLMs. In addition, we demonstrate that the upper bound is strongly correlated with an existing prompt sensitivity metric, PromptSensiScore. Moreover, by analyzing the logit variance, we find that prompt templates typically exert a greater influence on logits than the questions themselves. Overall, our results provide a general interpretation for why current LLMs can be highly sensitive to prompts with the same meaning, offering crucial evidence for understanding the prompt sensitivity of LLMs. Code for experiments is available at https://github.com/ku-nlp/Understanding_the_Prompt_Sensitivity.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

Null Hypothesis Significance Testing is the de facto tool for assessing effectiveness differences between Information Retrieval systems. Researchers use statistical tests to check whether those differences will generalise to online settings or are just due to the samples observed in the laboratory. Much work has been devoted to studying which test is the most reliable when comparing a pair of systems, but most of the IR real-world experiments involve more than two. In the multiple comparisons scenario, testing several systems simultaneously may inflate the errors committed by the tests. In this paper, we use a new approach to assess the reliability of multiple comparison procedures using simulated and real TREC data. Experiments show that Wilcoxon plus the Benjamini-Hochberg correction yields Type I error rates according to the significance level for typical sample sizes while being the best test in terms of statistical power.

Instruction embedding models have become common among state-of-the-art models, however are evaluated using a single prompt per task. The single-point evaluation ignores a main problem of the instruction-based approach namely: sensitivity to the phrasing of the instruction. We present an empirical study of prompt sensitivity across 6 embedding models, 11 datasets, and 15 task-specific prompts per dataset, a total of 990. We show that reported scores misrepresent the distribution of scores over plausible prompts. The default prompt can both systematically understate or overstate performance. Furthermore, we show that the leaderboard ranking is not robust to prompt selection: by choosing prompts favorably, any model in our study can be promoted to first place. Our findings suggest that single-prompt evaluation is insufficient for instruction-tuned embedding models and that benchmarks should incorporate prompt robustness, either by evaluating over multiple prompts or by reporting sensitivity alongside point estimates.

Decoding methods play an indispensable role in converting language models from next-token predictors into practical task solvers. Prior research on decoding methods, primarily focusing on task-specific models, may not extend to the current era of general-purpose large language models (LLMs). Moreover, the recent influx of decoding strategies has further complicated this landscape. This paper provides a comprehensive and multifaceted analysis of various decoding methods within the context of LLMs, evaluating their performance, robustness to hyperparameter changes, and decoding speeds across a wide range of tasks, models, and deployment environments. Our findings reveal that decoding method performance is notably task-dependent and influenced by factors such as alignment, model size, and quantization. Intriguingly, sensitivity analysis exposes that certain methods achieve superior performance at the cost of extensive hyperparameter tuning, highlighting the trade-off between attaining optimal results and the practicality of implementation in varying contexts.

We examine how closely pretrained LeWorldModel (LeWM) reproduces the effects of small action changes. Using matched simulator and model responses, we measure action sensitivity with Gram matrices at three perturbation scales. Across nearly a thousand states from Reacher and OGBench Cube, the most sensitive action directions differ by median angles of about 36^circ and 42^circ, respectively, at the smallest scale. Increasing the perturbation tenfold reduces median normalized Gram discrepancy by roughly 70\% on Reacher and 47\% on Cube. Geometry measured in pixels and embeddings varies more across scales than the chosen physical-state geometry. On held-out directions, environment Grams achieve median embedding response errors of 17\% and 28\%, compared with 26\% and 38\% for an isotropic reference with the same trace. LeWM's errors are higher at this scale. These findings support scale-specific evaluation of action sensitivity; its effect on control remains an open question.

Mitigating explicit and implicit biases in Large Language Models (LLMs) has become a critical focus in the field of natural language processing. However, many current methodologies evaluate scenarios in isolation, without considering the broader context or the spectrum of potential biases within each situation. To address this, we introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, which includes 450 offensive progressions containing 2,700 unique sentences of varying severity that progressively escalate from less to more explicitly offensive. Covering a broad spectrum of 9 demographics and 46 sub-demographics, STOP ensures inclusivity and comprehensive coverage. We evaluate several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3. Our findings reveal that even the best-performing models detect bias inconsistently, with success rates ranging from 19.3% to 69.8%. We also demonstrate how aligning models with human judgments on STOP can improve model answer rates on sensitive tasks such as BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or even improving performance. STOP presents a novel framework for assessing the complex nature of biases in LLMs, which will enable more effective bias mitigation strategies and facilitates the creation of fairer language models.

Comparative settings (e.g. pairwise choice, listwise ranking) have been adopted by a wide range of subjective studies for image quality assessment (IQA), as it inherently standardizes the evaluation criteria across different observers and offer more clear-cut responses. In this work, we extend the edge of emerging large multi-modality models (LMMs) to further advance visual quality comparison into open-ended settings, that 1) can respond to open-range questions on quality comparison; 2) can provide detailed reasonings beyond direct answers. To this end, we propose the Co-Instruct. To train this first-of-its-kind open-source open-ended visual quality comparer, we collect the Co-Instruct-562K dataset, from two sources: (a) LMM-merged single image quality description, (b) GPT-4V "teacher" responses on unlabeled data. Furthermore, to better evaluate this setting, we propose the MICBench, the first benchmark on multi-image comparison for LMMs. We demonstrate that Co-Instruct not only achieves 30% higher superior accuracy than state-of-the-art open-source LMMs, but also outperforms GPT-4V (its teacher), on both existing related benchmarks and the proposed MICBench. Our model is published at https://huggingface.co/q-future/co-instruct.

Large pre-trained language models have recently gained significant traction due to their improved performance on various down-stream tasks like text classification and question answering, requiring only few epochs of fine-tuning. However, their large model sizes often prohibit their applications on resource-constrained edge devices. Existing solutions of yielding parameter-efficient BERT models largely rely on compute-exhaustive training and fine-tuning. Moreover, they often rely on additional compute heavy models to mitigate the performance gap. In this paper, we present Sensi-BERT, a sensitivity driven efficient fine-tuning of BERT models that can take an off-the-shelf pre-trained BERT model and yield highly parameter-efficient models for downstream tasks. In particular, we perform sensitivity analysis to rank each individual parameter tensor, that then is used to trim them accordingly during fine-tuning for a given parameter or FLOPs budget. Our experiments show the efficacy of Sensi-BERT across different downstream tasks including MNLI, QQP, QNLI, SST-2 and SQuAD, showing better performance at similar or smaller parameter budget compared to various alternatives.

A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term. We measure the second. The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate shipped as gold, and a reviewer gold from independent expert re-annotation of a sample. Using it, we hold the scored output fixed and exchange the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other (95% CIs [3.23, 6.27] and [0.26, 3.37]), and by 7.55 in the worst language. The comparison between two outputs moves as well: the paired interaction between reference and system is +2.95 points (CI [+1.97, +3.87]), survives correction for multiple testing, and changes one language's margin outright. The cause is an undocumented aggregation default that usually kept one annotator when adjudication did not fire, making the shipped reference a partial copy of an output being scored. We report the resulting reference-sensitivity band, show how to compute one from any retained annotation record, and argue that the quantity belongs beside the score.

Radiological similarity depends on disease relationships and on fine details such as laterality, lobe, severity, size, and certainty. Broad biomedical similarity can overlook these qualifiers, particularly when several attributes vary together. We introduce QureRadEmbed, a 4B radiology-aware encoder trained with two complementary signals: RadSim supplies deterministic, attribute-decomposed ranking targets, while RadThought aligns reports with hierarchical evidence and reasoning descriptions. A three-stage curriculum combines these signals with report triplets, finding perturbations, and single- and cross-attribute contrasts. The final model achieves 0.996 mean ordering accuracy across ten controlled synthetic attributes and raises Spearman correlation with the designed joint-attribute targets from 0.501 to 0.976. On external findings-to-impression retrieval, Recall@1 reaches 10.5% on Open-I, 12.4% on testing XR, and 42.9% on testing CT, compared with 6.6%, 5.7%, and 31.4% for its backbone. Frozen embeddings support finding extraction with only 100 labeled testing-XR reports (macro-F1 0.481 versus 0.412 for the backbone). Whole-report comparison costs 8.8 seconds per 1,000 pairs in our benchmark, versus 2,755.1 seconds for the generative evaluator GREEN. Sentence-level comparison improves sensitivity to local discrepancies, although generative evaluation remains stronger on several expert-rated and subtle-error tasks. The results support reusable radiology-aware representations for search, structured report indexing, and efficient report comparison.

In this study, we propose a multicriteria group decision making (MCGDM) algorithm under uncertainty where data is collected as intervals. The proposed MCGDM algorithm aggregates the data, determines the optimal weights for criteria and ranks alternatives with no further input. The intervals give flexibility to experts in assessing alternatives against criteria and provide an opportunity to gain maximum information. We also propose a novel method to aggregate expert judgements using cloud models. We introduce an experimental approach to check the validity of the aggregation method. After that, we use the aggregation method for an MCGDM problem. Here, we find the optimal weights for each criterion by proposing a bilevel optimisation model. Then, we extend the technique for order of preference by similarity to ideal solution (TOPSIS) for data based on cloud models to prioritise alternatives. As a result, the algorithm can gain information from decision makers with different levels of uncertainty and examine alternatives with no more information from decision-makers. The proposed MCGDM algorithm is implemented on a case study of a cybersecurity problem to illustrate its feasibility and effectiveness. The results verify the robustness and validity of the proposed MCGDM using sensitivity analysis and comparison with other existing algorithms.

Retrieval-augmented language models are being increasingly tasked with subjective, contentious, and conflicting queries such as "is aspartame linked to cancer". To resolve these ambiguous queries, one must search through a large range of websites and consider "which, if any, of this evidence do I find convincing?". In this work, we study how LLMs answer this question. In particular, we construct ConflictingQA, a dataset that pairs controversial queries with a series of real-world evidence documents that contain different facts (e.g., quantitative results), argument styles (e.g., appeals to authority), and answers (Yes or No). We use this dataset to perform sensitivity and counterfactual analyses to explore which text features most affect LLM predictions. Overall, we find that current models rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone. Taken together, these results highlight the importance of RAG corpus quality (e.g., the need to filter misinformation), and possibly even a shift in how LLMs are trained to better align with human judgements.

Strong empirical evidence that one machine-learning algorithm A outperforms another one B ideally calls for multiple trials optimizing the learning pipeline over sources of variation such as data sampling, data augmentation, parameter initialization, and hyperparameters choices. This is prohibitively expensive, and corners are cut to reach conclusions. We model the whole benchmarking process, revealing that variance due to data sampling, parameter initialization and hyperparameter choice impact markedly the results. We analyze the predominant comparison methods used today in the light of this variance. We show a counter-intuitive result that adding more sources of variation to an imperfect estimator approaches better the ideal estimator at a 51 times reduction in compute cost. Building on these results, we study the error rate of detecting improvements, on five different deep-learning tasks/architectures. This study leads us to propose recommendations for performance comparisons.

Large language models (LLMs) are widely used as zero-shot and few-shot classifiers, where task behaviour is largely controlled through prompting. A growing number of works have observed that LLMs are sensitive to prompt variations, with small changes leading to large changes in performance. However, in many cases, the investigation of sensitivity is performed using underspecified prompts that provide minimal task instructions and weakly constrain the model's output space. In this work, we argue that a significant portion of the observed prompt sensitivity can be attributed to prompt underspecification. We systematically study and compare the sensitivity of underspecified prompts and prompts that provide specific instructions. Utilising performance analysis, logit analysis, and linear probing, we find that underspecified prompts exhibit higher performance variance and lower logit values for relevant tokens, while instruction-prompts suffer less from such problems. However, linear probing analysis suggests that the effects of prompt underspecification have only a marginal impact on the internal LLM representations, instead emerging in the final layers. Overall, our findings highlight the need for more rigour when investigating and mitigating prompt sensitivity.

Large language models are increasingly deployed as automated judges for evaluating other models, yet the stability of their verdicts under semantically equivalent prompt paraphrases remains unmeasured. We introduce JudgeSense, a framework and benchmark for quantifying this property via the Judge Sensitivity Score (JSS), defined as the fraction of paraphrase pairs on which a judge returns an identical decision. Evaluating nine judge models on 494 validated paraphrase pairs, we find that coherence is the only task where judges meaningfully differ, with JSS ranging from 0.389 to 0.992. On factuality, all judges cluster near JSS about 0.63, driven by a polarity-inverted prompt artifact; after correction, factuality JSS rises to about 0.9. Pairwise tasks (preference and relevance) exhibit degenerate always-A behavior in 8 of 9 judges, indicating strong position bias. Model scale does not predict consistency. We release code, decision logs, and a validated paraphrase dataset to support standardized JSS reporting.

As large language models (LLMs) are adopted as a fundamental component of language technologies, it is crucial to accurately characterize their performance. Because choices in prompt design can strongly influence model behavior, this design process is critical in effectively using any modern pre-trained generative language model. In this work, we focus on LLM sensitivity to a quintessential class of meaning-preserving design choices: prompt formatting. We find that several widely used open-source LLMs are extremely sensitive to subtle changes in prompt formatting in few-shot settings, with performance differences of up to 76 accuracy points when evaluated using LLaMA-2-13B. Sensitivity remains even when increasing model size, the number of few-shot examples, or performing instruction tuning. Our analysis suggests that work evaluating LLMs with prompting-based methods would benefit from reporting a range of performance across plausible prompt formats, instead of the currently-standard practice of reporting performance on a single format. We also show that format performance only weakly correlates between models, which puts into question the methodological validity of comparing models with an arbitrarily chosen, fixed prompt format. To facilitate systematic analysis we propose FormatSpread, an algorithm that rapidly evaluates a sampled set of plausible prompt formats for a given task, and reports the interval of expected performance without accessing model weights. Furthermore, we present a suite of analyses that characterize the nature of this sensitivity, including exploring the influence of particular atomic perturbations and the internal representation of particular formats.

Pruning is an effective method to reduce the memory footprint and FLOPs associated with neural network models. However, existing structured-pruning methods often result in significant accuracy degradation for moderate pruning levels. To address this problem, we introduce a new Hessian Aware Pruning (HAP) method coupled with a Neural Implant approach that uses second-order sensitivity as a metric for structured pruning. The basic idea is to prune insensitive components and to use a Neural Implant for moderately sensitive components, instead of completely pruning them. For the latter approach, the moderately sensitive components are replaced with with a low rank implant that is smaller and less computationally expensive than the original component. We use the relative Hessian trace to measure sensitivity, as opposed to the magnitude based sensitivity metric commonly used in the literature. We test HAP for both computer vision tasks and natural language tasks, and we achieve new state-of-the-art results. Specifically, HAP achieves less than 0.1%/0.5% degradation on PreResNet29/ResNet50 (CIFAR-10/ImageNet) with more than 70\%/50\% of parameters pruned. Meanwhile, HAP also achieves significantly better performance (up to 0.8\% with 60\% of parameters pruned) as compared to gradient based method for head pruning on transformer-based models. The framework has been open sourced and available online.

As fine-tuning becomes impractical at scale, probing is emerging as the preferred evaluation protocol. However, standard linear probing can understate the capability of models whose pre-training optimizes local representations rather than an explicit global representation. This motivates attentive probing, an alternative that uses attention to selectively aggregate patch-level features. Despite growing adoption, attentive probing is still underexplored: existing approaches are often over-parameterized and computationally inefficient. In this work, we revisit attentive probing through the lens of the accuracy vs. parameter-efficiency trade-off. We present the first comprehensive study of existing methods, analyzing their design choices and benchmarking their performance. Building on these insights, we propose efficient probing (EP), a lightweight yet effective multi-query cross-attention mechanism that eliminates redundant projections and reduces the number of trainable parameters. Across multiple benchmarks and pre-training paradigms, EP consistently outperforms linear probing and previous attentive probing methods, and remains effective when combined with parameter-efficient fine-tuning. Beyond evaluation, our analysis uncovers emerging properties of EP, including complementary attention maps, which open new directions for leveraging probing beyond protocol design. Project page: https://vrg.fel.cvut.cz/ep/.

We set out to reproduce a recent hallucination-mitigation result on the POPE benchmark. The LLaVA-1.5-7B baseline we measured, 82.21 F1 on the adversarial split, sat 1.21 to 8.67 points above the adversarial baselines we planned to compare against (81.00, 79.30, and 73.54). We looked at the evaluation code and found a readout choice that moves F1 on its own. A two-token logit readout decides yes versus no by comparing two vocabulary IDs, 3582 ("yes") and 1217 ("no"). Across 9,000 greedy-decode questions spanning all three splits, neither ID is generated even once. The model emits " Yes" (token 3869) and " No" (token 1939), because SentencePiece prepends a space byte after the prompt suffix "ASSISTANT:". Read with an eight-token lookup that covers the space-prefixed surface forms, LLaVA-1.5-7B scores F1 0.8221, 0.8498, and 0.8713 on the adversarial, popular, and random splits. That 6.13-point adversarial shift exceeds the adversarial gain reported by five of the six papers we surveyed. None of them states its readout, and our reference ran under 4-bit NF4 quantization, so we cannot say how much of any reported gain a matched evaluation would keep. That comparison stays unresolved until readout and precision are matched. On the fixed baseline we run nine inference-time corrections of our own (none beat it) and dissect why the adversarial split resists logit-level intervention. In a 3,000-question probe, wrong predictions draw more image attention than correct ones (Cohen's d=-0.46, p=1.4times10^{-24}), the opposite of what grounding-based corrections assume. A four-readout audit across four further VLMs from three tokenizer families confirms the confound is protocol-level, not model-specific: porting the LLaVA-1.5 IDs verbatim drops F1 to 0.00 on the two disjoint-vocabulary models, whereas a tokenizer-derived readout holds between 0.80 and 0.92.

Perception of toxicity evolves over time and often differs between geographies and cultural backgrounds. Similarly, black-box commercially available APIs for detecting toxicity, such as the Perspective API, are not static, but frequently retrained to address any unattended weaknesses and biases. We evaluate the implications of these changes on the reproducibility of findings that compare the relative merits of models and methods that aim to curb toxicity. Our findings suggest that research that relied on inherited automatic toxicity scores to compare models and techniques may have resulted in inaccurate findings. Rescoring all models from HELM, a widely respected living benchmark, for toxicity with the recent version of the API led to a different ranking of widely used foundation models. We suggest caution in applying apples-to-apples comparisons between studies and lay recommendations for a more structured approach to evaluating toxicity over time. Code and data are available at https://github.com/for-ai/black-box-api-challenges.

We consider objective evaluation measures of saliency explanations for complex black-box machine learning models. We propose simple robust variants of two notions that have been considered in recent literature: (in)fidelity, and sensitivity. We analyze optimal explanations with respect to both these measures, and while the optimal explanation for sensitivity is a vacuous constant explanation, the optimal explanation for infidelity is a novel combination of two popular explanation methods. By varying the perturbation distribution that defines infidelity, we obtain novel explanations by optimizing infidelity, which we show to out-perform existing explanations in both quantitative and qualitative measurements. Another salient question given these measures is how to modify any given explanation to have better values with respect to these measures. We propose a simple modification based on lowering sensitivity, and moreover show that when done appropriately, we could simultaneously improve both sensitivity as well as fidelity.

The recent successful paradigm of solving logical reasoning problems with tool-augmented large language models (LLMs) leverages translation of natural language (NL) statements into First-Order Logic~(FOL) and external theorem provers. However, the correctness of FOL statements, comprising operators and text, often go unverified due to the lack of a reliable evaluation metric for comparing generated and ground-truth FOLs. In this paper, we conduct a comprehensive study on the sensitivity of existing NL-, FOL-, and graph-based metrics to capture differences between a sampled FOL and its corresponding ground-truth. We then measure the alignment between a metric-based ranking of FOL outputs and a strong LLM as-a-judge. To do this, we first apply operator and text-based perturbations to ground-truth FOL statements to assess metric sensitivity. We then evaluate metric robustness by comparing the metrics against LLMs judgment. Our empirical findings highlight a clear oversensitivity in the n-gram metric BLEU for text perturbations. The operator perturbation affects the semantic graph metric Smatch++ for structural changes, and the FOL metric for specific operator changes. We observe a closer alignment between BertScore and LLM judgement, proving the importance of semantic evaluation. Additionally, we show that combining metrics enhances both robustness and sensitivity compared to using individual metrics.

Studies involving soundscape perception often exclude participants with hearing loss to prevent impaired perception from affecting experimental results. Participants are typically screened with pure tone audiometry, the "gold standard" for identifying and quantifying hearing loss at specific frequencies, and excluded if a study-dependent threshold is not met. However, procuring professional audiometric equipment for soundscape studies may be cost-ineffective, and manually performing audiometric tests is labour-intensive. Moreover, testing requirements for soundscape studies may not require sensitivities and specificities as high as that in a medical diagnosis setting. Hence, in this study, we investigate the effectiveness of the uHear app, an iOS application, as an affordable and automatic alternative to a conventional audiometer in screening participants for hearing loss for the purpose of soundscape studies or listening tests in general. Based on audiometric comparisons with the audiometer of 163 participants, the uHear app was found to have high precision (98.04%) when using the World Health Organization (WHO) grading scheme for assessing normal hearing. Precision is further improved (98.69%) when all frequencies assessed with the uHear app is considered in the grading, which lends further support to this cost-effective, automated alternative to screen for normal hearing.

Content moderation systems classify images as safe or unsafe but lack spatial grounding and interpretability: they cannot explain what sensitive behavior was detected, who is involved, or where it occurs. We introduce the Sensitive Benchmark (SenBen), the first large-scale scene graph benchmark for sensitive content, comprising 13,999 frames from 157 movies annotated with Visual Genome-style scene graphs (25 object classes, 28 attributes including affective states such as pain, fear, aggression, and distress, 14 predicates) and 16 sensitivity tags across 5 categories. We distill a frontier VLM into a compact 241M student model using a multi-task recipe that addresses vocabulary imbalance in autoregressive scene graph generation through suffix-based object identity, Vocabulary-Aware Recall (VAR) Loss, and a decoupled Query2Label tag head with asymmetric loss, yielding a +6.4 percentage point improvement in SenBen Recall over standard cross-entropy training. On grounded scene graph metrics, our student model outperforms all evaluated VLMs except Gemini models and all commercial safety APIs, while achieving the highest object detection and captioning scores across all models, at 7.6times faster inference and 16times less GPU memory.

The Perturbation Discrimination Score (PDS) is increasingly used to evaluate whether predicted perturbation effects remain distinguishable, including in Systema and the Virtual Cell Challenge. However, its behavior in high-dimensional gene-expression settings has not been examined in detail. We show that PDS is highly sensitive to the choice of similarity or distance measure and to the scale of predicted effects. Analysis of observed perturbation responses reveals that ell_1 and ell_2-based PDS behave very differently from cosine-based measures, even after norm matching. We provide geometric insight and discuss implications for future discrimination-based evaluation metrics.

Excluding irrelevant features in a pattern recognition task plays an important role in maintaining a simpler machine learning model and optimizing the computational efficiency. Nowadays with the rise of large scale datasets, feature selection is in great demand as it becomes a central issue when facing high-dimensional datasets. The present study provides a new measure of saliency for features by employing a Sensitivity Analysis (SA) technique called the extended Fourier amplitude sensitivity test, and a well-trained Feedforward Neural Network (FNN) model, which ultimately leads to the selection of a promising optimal feature subset. Ideas of the paper are mainly demonstrated based on adopting FNN model for feature selection in classification problems. But in the end, a generalization framework is discussed in order to give insights into the usage in regression problems as well as expressing how other function approximate models can be deployed. Effectiveness of the proposed method is verified by result analysis and data visualization for a series of experiments over several well-known datasets drawn from UCI machine learning repository.

Robustness evaluation must examine diverse visual perturbations, while benchmarks cover only some real-world conditions and physical testing is costly. Paired evaluations link clean and perturbed predictions for the same image, capturing changes in correctness, confidence, and acceptance beyond aggregate accuracy. We investigate how this image correspondence supports two needs in robustness evaluation: interpreting paired evaluation results and prioritizing samples for physical testing. To interpret paired evaluation results, we fix both sets of prediction records and vary their correspondence within each class. We prove that classwise correct-correct counts give the same sharp bounds on lost acceptance and mean true-class probability decrease among retained-correct inputs as any feasible five-state refinement. Distinguishing persistent from changed wrong answers can further constrain accepted-error transitions, while shared correspondence can establish policy orderings left unresolved by separate cost intervals. To prioritize samples for physical testing, we retain each image's synthetic responses and rank clean-correct images by their mean true-class probability under corruption. Across 44 classifiers, testing the highest-risk 20% finds 67% and 45% of failures under mild screen and print recaptures, versus 58% and 36% for clean confidence and 60% and 37% for an equal-size natural-transformation average. With both probability averaging and an A3Rank scoring adaptation, the tested corruption set yields higher mean failure recall than the natural-transform set; differences between scores depend on the source and budget. Together, these findings show that the value of correspondence depends on the evaluation objective: classwise counts suffice for specified reliability bounds, while image-specific synthetic responses improve the allocation of physical tests within the evaluated pool.

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

Sensitivity to false assumptions (or false premises) in information-seeking questions is critical for robust question-answering (QA) systems. Recent work has shown that false assumptions in naturally occurring questions pose challenges to current models, with low performance on both generative QA and simple detection tasks (Kim et al. 2023). However, the focus of existing work on naturally occurring questions leads to a gap in the analysis of model behavior on the long tail of the distribution of possible questions. To this end, we introduce Syn-(QA)^2, a set of two synthetically generated QA datasets: one generated using perturbed relations from Wikidata, and the other by perturbing HotpotQA (Yang et al. 2018). Our findings from evaluating a range of large language models are threefold: (1) false assumptions in QA are challenging, echoing the findings of prior work, (2) the binary detection task is challenging even compared to the difficulty of generative QA itself, possibly due to the linguistic structure of the problem, and (3) the detection task is more challenging with long-tail questions compared to naturally occurring questions, highlighting the utility of our synthetic datasets and generation method.

Relevance and fairness are two major objectives of recommender systems (RSs). Recent work proposes measures of RS fairness that are either independent from relevance (fairness-only) or conditioned on relevance (joint measures). While fairness-only measures have been studied extensively, we look into whether joint measures can be trusted. We collect all joint evaluation measures of RS relevance and fairness, and ask: How much do they agree with each other? To what extent do they agree with relevance/fairness measures? How sensitive are they to changes in rank position, or to increasingly fair and relevant recommendations? We empirically study for the first time the behaviour of these measures across 4 real-world datasets and 4 recommenders. We find that most of these measures: i) correlate weakly with one another and even contradict each other at times; ii) are less sensitive to rank position changes than relevance- and fairness-only measures, meaning that they are less granular than traditional RS measures; and iii) tend to compress scores at the low end of their range, meaning that they are not very expressive. We counter the above limitations with a set of guidelines on the appropriate usage of such measures, i.e., they should be used with caution due to their tendency to contradict each other and of having a very small empirical range.

Creating test collections for offline retrieval evaluation requires human effort to judge documents' relevance. This expensive activity motivated much work in developing methods for constructing benchmarks with fewer assessment costs. In this respect, adjudication methods actively decide both which documents and the order in which experts review them, in order to better exploit the assessment budget or to lower it. Researchers evaluate the quality of those methods by measuring the correlation between the known gold ranking of systems under the full collection and the observed ranking of systems under the lower-cost one. This traditional analysis ignores whether and how the low-cost judgements impact on the statistically significant differences among systems with respect to the full collection. We fill this void by proposing a novel methodology to evaluate how the low-cost adjudication methods preserve the pairwise significant differences between systems as the full collection. In other terms, while traditional approaches look for stability in answering the question "is system A better than system B?", our proposed approach looks for stability in answering the question "is system A significantly better than system B?", which is the ultimate questions researchers need to answer to guarantee the generalisability of their results. Among other results, we found that the best methods in terms of ranking of systems correlation do not always match those preserving statistical significance.

Recent studies of the emergent capabilities of transformer-based Natural Language Understanding (NLU) models have indicated that they have an understanding of lexical and compositional semantics. We provide evidence that suggests these claims should be taken with a grain of salt: we find that state-of-the-art Natural Language Inference (NLI) models are sensitive towards minor semantics preserving surface-form variations, which lead to sizable inconsistent model decisions during inference. Notably, this behaviour differs from valid and in-depth comprehension of compositional semantics, however does neither emerge when evaluating model accuracy on standard benchmarks nor when probing for syntactic, monotonic, and logically robust reasoning. We propose a novel framework to measure the extent of semantic sensitivity. To this end, we evaluate NLI models on adversarially generated examples containing minor semantics-preserving surface-form input noise. This is achieved using conditional text generation, with the explicit condition that the NLI model predicts the relationship between the original and adversarial inputs as a symmetric equivalence entailment. We systematically study the effects of the phenomenon across NLI models for in- and out-of- domain settings. Our experiments show that semantic sensitivity causes performance degradations of 12.92% and 23.71% average over in- and out-of- domain settings, respectively. We further perform ablation studies, analysing this phenomenon across models, datasets, and variations in inference and show that semantic sensitivity can lead to major inconsistency within model predictions.

Physicians now pose millions of clinical questions to AI tools each week, yet these tools are evaluated largely on hypothetical or exam-style questions, not those actually asked in practice. We report a blinded evaluation built on 620 Real-world Point-Of-Care Queries (Real-POCQi) submitted to the OpenEvidence (OE) platform by physicians spanning 30 specialties, as well as 187 questions from HealthBench. 149 practicing physicians across 36 states made head-to-head comparisons between answers from three frontier general-purpose models (Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5) and a specialized clinical tool (OE), with graders matched to each question's specialty. When comparing answers along five dimensions relevant to clinical decision support -- accuracy, clinical utility, source quality, verifiability, & completeness -- physicians scored the specialized tool highest on all axes; in the primary analysis on Real-POCQi, win differences (margins between win and loss rates) ranged from 25 to 39 percentage points (p<0.001). Results remained consistent in sensitivity analyses stratifying by citation display, answer length, OE-user status, and Real-POCQi versus HealthBench. In parallel, LLM judges were found to systematically differ from expert judges, though both generally agreed on the best model. These findings underscore two conclusions: (i) AI tool evaluations should reflect real-world query distributions and use expert judges that mirror the specialization defining modern medicine and (ii) the consistent advantage of the specialized tool over general-purpose models does not necessarily mean that the latter cannot serve similar purposes, but that targeted engineering and customization can yield meaningful gains in performance for its users. We release Real-POCQi as a public benchmark, as well as the prespecified statistical analysis for reproducing results of this study.

Causal sensitivity analysis aims to provide bounds for causal effect estimates in the presence of unobserved confounding. However, existing methods for causal sensitivity analysis are per-instance procedures, meaning that changes to the dataset, causal query, sensitivity level, or treatment require new computation. Here, we instead present an in-context learning approach. Specifically, we propose an amortized approach to causal sensitivity analysis based on prior-data fitted networks. A key challenge is that the sensitivity bounds are not directly available when sampling training data. To address this, we develop a general prior-data construction that is applicable across the class of generalized treatment sensitivity models. Our construction involves a Lagrangian scalarization of the objective to generate training labels for the bounds through a tradeoff between causal effect min/max-imization and sensitivity model violation, which avoids model-specific analytical derivations. We further show that, under standard convexity and linearity conditions, our objective recovers the full Pareto frontier of solutions. Empirically, we demonstrate our amortized approach across various datasets, causal queries, and sensitivity levels, where our approach achieves a test-time computation that is orders of magnitude faster than per-instance methods. To the best of our knowledge, ours is the first foundation model for in-context learning for causal sensitivity analysis.

To increase the availability and adoption of the soundscape standard, a low-cost calibration procedure for reproduction of audio stimuli over headphones was proposed as part of the global ``Soundscape Attributes Translation Project'' (SATP) for validating ISO/TS~12913-2:2018 perceived affective quality (PAQ) attribute translations. A previous preliminary study revealed significant deviations from the intended equivalent continuous A-weighted sound pressure levels (L_{A,eq}) using the open-circuit voltage (OCV) calibration procedure. For a more holistic human-centric perspective, the OCV method is further investigated here in terms of psychoacoustic parameters, including relevant exceedance levels to account for temporal effects on the same 27 stimuli from the SATP. Moreover, a within-subjects experiment with 36 participants was conducted to examine the effects of OCV calibration on the PAQ attributes in ISO/TS~12913-2:2018. Bland-Altman analysis of the objective indicators revealed large biases in the OCV method across all weighted sound level and loudness indicators; and roughness indicators at 5{\%} and 10{\%} exceedance levels. Significant perceptual differences due to the OCV method were observed in about 20{\%} of the stimuli, which did not correspond clearly with the biased acoustic indicators. A cautioned interpretation of the objective and perceptual differences due to small and unpaired samples nevertheless provide grounds for further investigation.

Prediction models can exhibit sensitivity with respect to training data: small changes in the training data can produce models that assign conflicting predictions to individual data points during test time. In this work, we study this sensitivity in recommender systems, where users' recommendations are drastically altered by minor perturbations in other unrelated users' interactions. We introduce a measure of stability for recommender systems, called Rank List Sensitivity (RLS), which measures how rank lists generated by a given recommender system at test time change as a result of a perturbation in the training data. We develop a method, CASPER, which uses cascading effect to identify the minimal and systematical perturbation to induce higher instability in a recommender system. Experiments on four datasets show that recommender models are overly sensitive to minor perturbations introduced randomly or via CASPER - even perturbing one random interaction of one user drastically changes the recommendation lists of all users. Importantly, with CASPER perturbation, the models generate more unstable recommendations for low-accuracy users (i.e., those who receive low-quality recommendations) than high-accuracy ones.

Reference listening is a common strategy in music production, but current comparison tools often obscure a key human judgment: deciding what should be compared. We present Diptych, an AI-assisted system that lets users define comparison scope across whole tracks or independently selected segments, while inspecting structured audio features and scope-specific AI interpretations. We evaluated Diptych in a within-participants study with 12 musicians, complemented by source-blinded ratings from four expert listeners. Participants used the system to surface additional differences, nine of ten of which received at least partial expert support, and reported good usability and greater clarity about possible next steps. These findings suggest that AI support for creative comparison should prioritize user-defined scope, inspectable evidence, and actionable guidance, while avoiding authoritative judgments that exceed what the evidence can support.

Current reading comprehension models generalise well to in-distribution test sets, yet perform poorly on adversarially selected inputs. Most prior work on adversarial inputs studies oversensitivity: semantically invariant text perturbations that cause a model's prediction to change when it should not. In this work we focus on the complementary problem: excessive prediction undersensitivity, where input text is meaningfully changed but the model's prediction does not, even though it should. We formulate a noisy adversarial attack which searches among semantic variations of the question for which a model erroneously predicts the same answer, and with even higher probability. Despite comprising unanswerable questions, both SQuAD2.0 and NewsQA models are vulnerable to this attack. This indicates that although accurate, models tend to rely on spurious patterns and do not fully consider the information specified in a question. We experiment with data augmentation and adversarial training as defences, and find that both substantially decrease vulnerability to attacks on held out data, as well as held out attack spaces. Addressing undersensitivity also improves results on AddSent and AddOneSent, and models furthermore generalise better when facing train/evaluation distribution mismatch: they are less prone to overly rely on predictive cues present only in the training set, and outperform a conventional model by as much as 10.9% F1.

The Advanced LIGO gravitational wave detectors are second generation instruments designed and built for the two LIGO observatories in Hanford, WA and Livingston, LA. The two instruments are identical in design, and are specialized versions of a Michelson interferometer with 4 km long arms. As in initial LIGO, Fabry-Perot cavities are used in the arms to increase the interaction time with a gravitational wave, and power recycling is used to increase the effective laser power. Signal recycling has been added in Advanced LIGO to improve the frequency response. In the most sensitive frequency region around 100 Hz, the design strain sensitivity is a factor of 10 better than initial LIGO. In addition, the low frequency end of the sensitivity band is moved from 40 Hz down to 10 Hz. All interferometer components have been replaced with improved technologies to achieve this sensitivity gain. Much better seismic isolation and test mass suspensions are responsible for the gains at lower frequencies. Higher laser power, larger test masses and improved mirror coatings lead to the improved sensitivity at mid- and high- frequencies. Data collecting runs with these new instruments are planned to begin in mid-2015.

We apply augmentations to our dataset to enhance the quality of our predictions and make our final models more resilient to noisy data and domain drifts. Yet the question remains, how are these augmentations going to perform with different hyper-parameters? In this study we evaluate the sensitivity of augmentations with regards to the model's hyper parameters along with their consistency and influence by performing a Local Surrogate (LIME) interpretation on the impact of hyper-parameters when different augmentations are applied to a machine learning model. We have utilized Linear regression coefficients for weighing each augmentation. Our research has proved that there are some augmentations which are highly sensitive to hyper-parameters and others which are more resilient and reliable.

Now that AI-driven moderation has become pervasive in everyday life, we often hear claims that "the AI is biased". While this is often said jokingly, the light-hearted remark reflects a deeper concern. How can we be certain that an online post flagged as "inappropriate" was not simply the victim of a biased algorithm? This paper investigates this problem using a dual approach. First, I conduct a quantitative benchmark of a widely used toxicity model (unitary/toxic-bert) to measure performance disparity between text in African-American English (AAE) and Standard American English (SAE). The benchmark reveals a clear, systematic bias: on average, the model scores AAE text as 1.8 times more toxic and 8.8 times higher for "identity hate". Second, I introduce an interactive pedagogical tool that makes these abstract biases tangible. The tool's core mechanic, a user-controlled "sensitivity threshold," demonstrates that the biased score itself is not the only harm; instead, the more-concerning harm is the human-set, seemingly neutral policy that ultimately operationalises discrimination. This work provides both statistical evidence of disparate impact and a public-facing tool designed to foster critical AI literacy.

We evaluate the robustness of several large language models on multiple datasets. Robustness here refers to the relative insensitivity of the model's answers to meaning-preserving variants of their input. Benchmark datasets are constructed by introducing naturally-occurring, non-malicious perturbations, or by generating semantically equivalent paraphrases of input questions or statements. We further propose a novel metric for assessing a model robustness, and demonstrate its benefits in the non-adversarial scenario by empirical evaluation of several models on the created datasets.

Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping win rates between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure uncovers an implicit bias towards WEIRD contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts.

Explaining the output of a deep network remains a challenge. In the case of an image classifier, one type of explanation is to identify pixels that strongly influence the final decision. A starting point for this strategy is the gradient of the class score function with respect to the input image. This gradient can be interpreted as a sensitivity map, and there are several techniques that elaborate on this basic idea. This paper makes two contributions: it introduces SmoothGrad, a simple method that can help visually sharpen gradient-based sensitivity maps, and it discusses lessons in the visualization of these maps. We publish the code for our experiments and a website with our results.

We develop an algorithm which exceeds the performance of board certified cardiologists in detecting a wide range of heart arrhythmias from electrocardiograms recorded with a single-lead wearable monitor. We build a dataset with more than 500 times the number of unique patients than previously studied corpora. On this dataset, we train a 34-layer convolutional neural network which maps a sequence of ECG samples to a sequence of rhythm classes. Committees of board-certified cardiologists annotate a gold standard test set on which we compare the performance of our model to that of 6 other individual cardiologists. We exceed the average cardiologist performance in both recall (sensitivity) and precision (positive predictive value).

A retrieval model should not only interpolate the training data but also extrapolate well to the queries that are different from the training data. While neural retrieval models have demonstrated impressive performance on ad-hoc search benchmarks, we still know little about how they perform in terms of interpolation and extrapolation. In this paper, we demonstrate the importance of separately evaluating the two capabilities of neural retrieval models. Firstly, we examine existing ad-hoc search benchmarks from the two perspectives. We investigate the distribution of training and test data and find a considerable overlap in query entities, query intent, and relevance labels. This finding implies that the evaluation on these test sets is biased toward interpolation and cannot accurately reflect the extrapolation capacity. Secondly, we propose a novel evaluation protocol to separately evaluate the interpolation and extrapolation performance on existing benchmark datasets. It resamples the training and test data based on query similarity and utilizes the resampled dataset for training and evaluation. Finally, we leverage the proposed evaluation protocol to comprehensively revisit a number of widely-adopted neural retrieval models. Results show models perform differently when moving from interpolation to extrapolation. For example, representation-based retrieval models perform almost as well as interaction-based retrieval models in terms of interpolation but not extrapolation. Therefore, it is necessary to separately evaluate both interpolation and extrapolation performance and the proposed resampling method serves as a simple yet effective evaluation tool for future IR studies.

Large Language Models (LLMs) empowered with retrieval mechanisms have achieved strong progress in open-domain question answering (QA). Yet, the conventional retrieve--then--answer paradigm often suffers from two key limitations: (1) low signal-to-noise ratio in retrieved evidence, where useful information is buried under irrelevant content, and (2) error accumulation in multi-hop reasoning when incomplete or noisy passages are involved. To address these challenges, we present EviNote-RAG, an agentic RAG framework that introduces a structured retrieve--note--answer pipeline. Instead of directly reasoning over raw retrievals, the model is trained to compose Supportive-Evidence Notes (SENs), concise, human-like notes that preserve only answer-relevant information, highlight uncertainty, and explicitly state when no useful evidence exists. This distillation process is further reinforced by the Evidence Quality Reward (EQR), an entailment-based signal that evaluates whether SENs logically support the final answer. Together, SENs and EQR guide the model toward faithful and robust reasoning, while reducing the impact of noise. Experiments on in-domain and out-of-domain QA benchmarks show that EviNote-RAG consistently outperforms strong baselines in accuracy, generalization, and training stability. In particular, it achieves state-of-the-art results while enhancing robustness and efficiency, yielding relative F1 gains of 20\% on HotpotQA (+0.093), 40\% on Bamboogle (+0.151), and 91\% on 2Wiki (+0.256) via denser rewards and reduced verbosity.

Despite the remarkable success of deep neural networks in a myriad of settings, several works have demonstrated their overwhelming sensitivity to near-imperceptible perturbations, known as adversarial attacks. On the other hand, prior works have also observed that deep networks can be under-sensitive, wherein large-magnitude perturbations in input space do not induce appreciable changes to network activations. In this work, we study in detail the phenomenon of under-sensitivity in vision models such as CNNs and Transformers, and present techniques to study the geometry and extent of "equi-confidence" level sets of such networks. We propose a Level Set Traversal algorithm that iteratively explores regions of high confidence with respect to the input space using orthogonal components of the local gradients. Given a source image, we use this algorithm to identify inputs that lie in the same equi-confidence level set as the source image despite being perceptually similar to arbitrary images from other classes. We further observe that the source image is linearly connected by a high-confidence path to these inputs, uncovering a star-like structure for level sets of deep networks. Furthermore, we attempt to identify and estimate the extent of these connected higher-dimensional regions over which the model maintains a high degree of confidence. The code for this project is publicly available at https://github.com/SriramB-98/blindspots-neurips-sub

Neural language models (LMs) can be used to evaluate the truth of factual statements in two ways: they can be either queried for statement probabilities, or probed for internal representations of truthfulness. Past work has found that these two procedures sometimes disagree, and that probes tend to be more accurate than LM outputs. This has led some researchers to conclude that LMs "lie" or otherwise encode non-cooperative communicative intents. Is this an accurate description of today's LMs, or can query-probe disagreement arise in other ways? We identify three different classes of disagreement, which we term confabulation, deception, and heterogeneity. In many cases, the superiority of probes is simply attributable to better calibration on uncertain answers rather than a greater fraction of correct, high-confidence answers. In some cases, queries and probes perform better on different subsets of inputs, and accuracy can further be improved by ensembling the two. Code is available at github.com/lingo-mit/lm-truthfulness.

This paper introduces a novel double stimulus subjective assessment methodology for the evaluation of high quality images to address the limitations of existing protocols in detecting subtle perceptual differences. The In-place Double Stimulus Quality Scale (IDSQS) allows subjects to alternately view a reference and a distorted image at the same spatial location, facilitating a more intuitive detection of differences in quality, especially at high to visually lossless quality levels. A large-scale crowdsourcing study employing this methodology was conducted, generating a comprehensive public dataset to evaluate perceived image quality across several compression algorithms and distortion levels. An additional contribution is the modeling of quality scores using a Beta distribution, allowing for the assessment of variability and subject consistency. Our findings demonstrate the effectiveness of the IDSQS methodology in achieving high correlation with more precise subjective evaluation benchmarks. The dataset, subjective data, and graphical user interface developed for this study are publicly available at https://github.com/shimamohammadi/IDSQS

Advances in unsupervised probes such as Contrast-Consistent Search (CCS), which reveal latent beliefs without relying on token outputs, raise the question of whether these methods can reliably assess model alignment. We investigate this by examining the sensitivity of CCS to harmful vs. safe statements and by introducing Polarity-Aware CCS (PA-CCS), a method for evaluating whether a model's internal representations remain consistent under polarity inversion. We propose two alignment-oriented metrics, Polar-Consistency and the Contradiction Index, to quantify the semantic robustness of a model's latent knowledge. To validate PA-CCS, we curate two main datasets and one control dataset containing matched harmful-safe sentence pairs constructed using different methodologies (concurrent and antagonistic statements). We apply PA-CCS to 16 language models. Our results show that PA-CCS identifies both architectural and layer-specific differences in the encoding of latent harmful knowledge. Notably, replacing the negation token with a meaningless marker degrades PA-CCS scores for models with well-aligned internal representations, while models lacking robust internal calibration do not exhibit this degradation. Our findings highlight the potential of unsupervised probing for alignment evaluation and emphasize the need to incorporate structural robustness checks into interpretability benchmarks. Code and datasets are available at: https://github.com/SadSabrina/polarity-probing. WARNING: This paper contains potentially sensitive, harmful, and offensive content.

Estimating the difficulty of a dataset typically involves comparing state-of-the-art models to humans; the bigger the performance gap, the harder the dataset is said to be. However, this comparison provides little understanding of how difficult each instance in a given distribution is, or what attributes make the dataset difficult for a given model. To address these questions, we frame dataset difficulty -- w.r.t. a model V -- as the lack of V-usable information (Xu et al., 2019), where a lower value indicates a more difficult dataset for V. We further introduce pointwise \mathcal{V-information} (PVI) for measuring the difficulty of individual instances w.r.t. a given distribution. While standard evaluation metrics typically only compare different models for the same dataset, V-usable information and PVI also permit the converse: for a given model V, we can compare different datasets, as well as different instances/slices of the same dataset. Furthermore, our framework allows for the interpretability of different input attributes via transformations of the input, which we use to discover annotation artefacts in widely-used NLP benchmarks.

Context can change whether a request is harmful without changing its topic or surface form. We ask whether residual-stream probes distinguish harmful requests from surface-matched benign controls at a useful operating point. Across three 7-8B model families, an activation sensor blocks 95.5-97.7 percent of judge-classified compliant attacks in a taxonomy-selected set. It also blocks 59.6-68.4 percent of XSTest prompts. A fully disjoint audit reconstructs near-ceiling source-contrast AUROC (0.996-0.999), but fixed transfer to matched pairs is weaker: 0.656-0.819 on the guard-selected Twin-n70 subset and 0.590-0.690 on the full Twin-n163 cohort. We test ten axes on the reference family and seven across all families with leakage, hold-out, and permutation controls. On Twin-n163, no axis evaluated without direct pair-boundary fitting reaches the specified numerical threshold. Requiring persistence on that full cohort was added at analysis time. A separately specified 24B/32B extension gives the same result. Pair-trained classifiers weaken under category and generation-batch hold-out and false-block 79.6-100 percent of XSTest at 95 percent in-corpus TPR. At the tested read points, these activation scores behave as broad-risk detectors rather than standalone context adjudicators.

LLMs are highly sensitive to prompt phrasing, yet standard benchmarks typically report performance using a single prompt, raising concerns about the reliability of such evaluations. In this work, we argue for a stochastic method of moments evaluation over the space of meaning-preserving prompt perturbations. We introduce a formal definition of reliable evaluation that accounts for prompt sensitivity, and suggest ReliableEval - a method for estimating the number of prompt resamplings needed to obtain meaningful results. Using our framework, we stochastically evaluate five frontier LLMs and find that even top-performing models like GPT-4o and Claude-3.7-Sonnet exhibit substantial prompt sensitivity. Our approach is model-, task-, and metric-agnostic, offering a recipe for meaningful and robust LLM evaluation.

Attribution methods can provide powerful insights into the reasons for a classifier's decision. We argue that a key desideratum of an explanation method is its robustness to input hyperparameters which are often randomly set or empirically tuned. High sensitivity to arbitrary hyperparameter choices does not only impede reproducibility but also questions the correctness of an explanation and impairs the trust of end-users. In this paper, we provide a thorough empirical study on the sensitivity of existing attribution methods. We found an alarming trend that many methods are highly sensitive to changes in their common hyperparameters e.g. even changing a random seed can yield a different explanation! Interestingly, such sensitivity is not reflected in the average explanation accuracy scores over the dataset as commonly reported in the literature. In addition, explanations generated for robust classifiers (i.e. which are trained to be invariant to pixel-wise perturbations) are surprisingly more robust than those generated for regular classifiers.

Unobserved confounding is common in many applications, making causal inference from observational data challenging. As a remedy, causal sensitivity analysis is an important tool to draw causal conclusions under unobserved confounding with mathematical guarantees. In this paper, we propose NeuralCSA, a neural framework for generalized causal sensitivity analysis. Unlike previous work, our framework is compatible with (i) a large class of sensitivity models, including the marginal sensitivity model, f-sensitivity models, and Rosenbaum's sensitivity model; (ii) different treatment types (i.e., binary and continuous); and (iii) different causal queries, including (conditional) average treatment effects and simultaneous effects on multiple outcomes. The generality of \frameworkname is achieved by learning a latent distribution shift that corresponds to a treatment intervention using two conditional normalizing flows. We provide theoretical guarantees that NeuralCSA is able to infer valid bounds on the causal query of interest and also demonstrate this empirically using both simulated and real-world data.

Large language models work well on English and behave in poorly understood ways on languages typologically far from it. Japanese is a clean example, where evaluation still leans on translation quality and JGLUE-style benchmarks, which roll lexical, syntactic and pragmatic competence into a single score. The phenomena on which general-purpose models fail Japanese users are pragmatic: honorifics, in-group and out-group reference, context-sensitive politeness, zero anaphora. I introduce J-PragEval-v0, a minimal-pair benchmark isolating four such phenomena from surface fluency, and combine it with linear probes and teacher-forced log-probability evaluation to ask where inside TinySwallow-1.5B (28 layers, hidden size 1536) the corresponding contrasts live. The four features split three ways. Honorific register sits cleanly in the residual stream: 0.96 balanced accuracy at layer 15, and the model flips its preferred continuation with the scenario on 93 percent of items. Implicit subject and in-group reference are not linearly decodable at the final prompt token (0.48 and 0.38), yet flip rates are 0.77 and 0.79, so the contrast is worked out during generation rather than stored at the prompt. Indirect refusal is the negative case: 0.95 probe accuracy collapsing to a 0.43 flip rate under length-normalised teacher forcing, because the current minimal pairs conflate politeness with continuation length. I also specify Pragmatic Representation Steering, a parameter-free inference-time method that edits residual-stream activations along the class-mean-difference directions probing identifies. Feasibility is argued indirectly rather than demonstrated: the contrastive activation addition baseline, the same geometry the method would inject, recovers probe accuracy within one to two points of logistic regression wherever a linear signal exists. Scaling to Llama-3.1-Swallow-8B is the next step.

Rationale: Computer aided detection (CAD) algorithms for Pulmonary Embolism (PE) algorithms have been shown to increase radiologists' sensitivity with a small increase in specificity. However, CAD for PE has not been adopted into clinical practice, likely because of the high number of false positives current CAD software produces. Objective: To generate a database of annotated computed tomography pulmonary angiographies, use it to compare the sensitivity and false positive rate of current algorithms and to develop new methods that improve such metrics. Methods: 91 Computed tomography pulmonary angiography scans were annotated by at least one radiologist by segmenting all pulmonary emboli visible on the study. 20 annotated CTPAs were open to the public in the form of a medical image analysis challenge. 20 more were kept for evaluation purposes. 51 were made available post-challenge. 8 submissions, 6 of them novel, were evaluated on the 20 evaluation CTPAs. Performance was measured as per embolus sensitivity vs. false positives per scan curve. Results: The best algorithms achieved a per-embolus sensitivity of 75% at 2 false positives per scan (fps) or of 70% at 1 fps, outperforming the state of the art. Deep learning approaches outperformed traditional machine learning ones, and their performance improved with the number of training cases. Significance: Through this work and challenge we have improved the state-of-the art of computer aided detection algorithms for pulmonary embolism. An open database and an evaluation benchmark for such algorithms have been generated, easing the development of further improvements. Implications on clinical practice will need further research.

Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline ``meaning-preserving'' modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted interventions, we propose a framework in which we compare (via statistical testing) differences observed under target interventions to those induced by paraphrasing inputs. We then use this framework to revisit a analysis done on the MedPerturb dataset, which reported evidence of model sensitivity to patient demographics and stylistic cues. We find that these effects largely dissipate when we account for general model sensitivity, with only 5 of 120 tests reaching statistical significance. Applying the same framework to occupational biography classification, we detect clearly significant directional gender bias, showing that the framework identifies real directional effects even when they are small. We evaluate a range of metrics -- aggregate, per-sample distributional, and regression -- and find that per-sample metrics are dramatically more powerful than aggregate metrics and regression powerfully and uniquely characterizes effect direction and magnitude.

We introduce EQ-Bench, a novel benchmark designed to evaluate aspects of emotional intelligence in Large Language Models (LLMs). We assess the ability of LLMs to understand complex emotions and social interactions by asking them to predict the intensity of emotional states of characters in a dialogue. The benchmark is able to discriminate effectively between a wide range of models. We find that EQ-Bench correlates strongly with comprehensive multi-domain benchmarks like MMLU (Hendrycks et al., 2020) (r=0.97), indicating that we may be capturing similar aspects of broad intelligence. Our benchmark produces highly repeatable results using a set of 60 English-language questions. We also provide open-source code for an automated benchmarking pipeline at https://github.com/EQ-bench/EQ-Bench and a leaderboard at https://eqbench.com

Evaluation budgets in agentic retrieval-augmented generation span questions, search trajectories, and repeated answers. We measure allocation precision, reading efficiency, and cost boundaries using a retrieval-feedback comparison on HotpotQA and MuSiQue. At 34.14--34.39M model tokens, broader question coverage lowers standard error by 33\% versus five reads and 12.6\% versus three trajectories. Archived nested and Q-only forecasts predict these allocations within 4.0\% and 3.5\%, respectively. Depth subsets establish no clear forecasting advantage beyond the two-trajectory audit. One-read variance penalties relative to the fitted optimum at the same token budget are 0--9.9\%, with substantial Pro uncertainty. Under recorded model fees, more questions beat more trajectories at search prices of \$0--1 per 1,000 requests; question-versus-read fee rankings remain unresolved. Temperature zero cuts answer disagreement from 14.3\% to 3.4\% while comparison precision stays similar. \par\medskip\noindentKeywords: Agentic RAG; Evaluation budget; Generalizability theory; Repeated sampling.

We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer semantics rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0\% to 57.5\% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number \(k\) of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5\%) at both \(k=0\) and the adapter's training-time \(k=1\), 188/200 (94.0\%) at the packaged runtime's default \(k=3\) --- and the submitted confidence-gated override adds no net accuracy on top of \(k=3\), selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~zhao2024ratescore decrease as \(k\) grows, but paired sign tests on token-F1 differences are nonsignificant (\(p \ge 0.29\)); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0\% MCQ accuracy on the development holdout and 93.20\% on the organizer's official pre-evaluation, versus 29.43\% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

Post-training quantization (PTQ) has emerged as a promising technique for mitigating memory consumption and computational costs in large language models (LLMs). However, a systematic examination of various quantization schemes, model families, and quantization bit precision has been absent from the literature. In this paper, we conduct a comprehensive analysis of these factors by investigating the effects of PTQ on weight-only, activation-only, and weight-and-activation quantization using diverse methods such as round-to-nearest (RTN), GPTQ, ZeroQuant, and their variants. We apply these methods to two distinct model families with parameters ranging from 125M to 176B. Our contributions include: (1) a sensitivity analysis revealing that activation quantization is generally more susceptible to weight quantization, with smaller models often outperforming larger models in terms of activation quantization; (2) an evaluation and comparison of existing PTQ methods to optimize model size reduction while minimizing the impact on accuracy, revealing that none of the current methods can achieve the original model quality for quantization with either INT4-weight or INT4-weight-and-INT8-activation; (3) based on these insights, we propose an optimized method called Low-Rank Compensation (LoRC), which employs low-rank matrices to enhance model quality recovery with a minimal increase in model size.

Many peer-review venues are using algorithms to assign submissions to reviewers. The crux of such automated approaches is the notion of the "similarity score" -- a numerical estimate of the expertise of a reviewer in reviewing a paper -- and many algorithms have been proposed to compute these scores. However, these algorithms have not been subjected to a principled comparison, making it difficult for stakeholders to choose the algorithm in an evidence-based manner. The key challenge in comparing existing algorithms and developing better algorithms is the lack of publicly available gold-standard data. We address this challenge by collecting a novel dataset of similarity scores that we release to the research community. Our dataset consists of 477 self-reported expertise scores provided by 58 researchers who evaluated their expertise in reviewing papers they have read previously. Using our dataset, we compare several widely used similarity algorithms and offer key insights. First, all algorithms exhibit significant error, with misranking rates between 12%-30% in easier cases and 36%-43% in harder ones. Second, most specialized algorithms are designed to work with titles and abstracts of papers, and in this regime the SPECTER2 algorithm performs best. Interestingly, classical TF-IDF matches SPECTER2 in accuracy when given access to full submission texts. In contrast, off-the-shelf LLMs lag behind specialized approaches.

Retrieval-augmented generation (RAG) reduces but does not eliminate hallucination, and existing detectors return a single answer-level score that does not indicate which sentence is unsupported, or why. To close this gap, we introduce Grounding-Aware Sensitivity by Perturbation (GASP), a span-level detector that scores each answer sentence by how strongly its likelihood depends on the retrieved evidence, a quantity we term grounding sensitivity. GASP holds the answer fixed and re-scores it under the full context, under no context, and with each chunk removed, then measures the log-likelihood drops and Jensen-Shannon divergences (JSD). The likelihood of a grounded sentence collapses once its supporting passage is removed, whereas a hallucinated sentence is almost unaffected, a contrast we interpret by casting decoding as a random nonlinear iterated function system (RNIFS). We evaluate GASP on three benchmarks (RAGTruth, TofuEval, RAGBench) with three instruction-tuned scorers from two model families (Qwen2.5-0.5B, Qwen2.5-1.5B, and SmolLM2-1.7B) under a leakage-clean protocol. On RAGTruth it reaches a response-level area under the ROC curve (AUC) of about 0.73 and a span-level AUC of about 0.67, improving significantly over perplexity and by clear margins over length, whole-context natural language inference (NLI), and self-consistency baselines. The only baseline competitive at the span level is a well-configured chunk-level entailment verifier, which requires a separate model, whereas a training-free threshold on the grounding features matches the trained classifier without labeled data and serves as the default detector. Beyond RAGTruth, the signal transfers to TofuEval but not to short-answer question answering in RAGBench, showing GASP is best suited to outputs constructed from the retrieved context rather than answers recoverable from parametric knowledge.

The long-standing one-to-many problem of gold standard responses in open-domain dialogue systems presents challenges for automatic evaluation metrics. Though prior works have demonstrated some success by applying powerful Large Language Models (LLMs), existing approaches still struggle with the one-to-many problem, and exhibit subpar performance in domain-specific scenarios. We assume the commonsense reasoning biases within LLMs may hinder their performance in domainspecific evaluations. To address both issues, we propose a novel framework SLIDE (Small and Large Integrated for Dialogue Evaluation), that leverages both a small, specialised model (SLM), and LLMs for the evaluation of open domain dialogues. Our approach introduces several techniques: (1) Contrastive learning to differentiate between robust and non-robust response embeddings; (2) A novel metric for semantic sensitivity that combines embedding cosine distances with similarity learned through neural networks, and (3) a strategy for incorporating the evaluation results from both the SLM and LLMs. Our empirical results demonstrate that our approach achieves state-of-the-art performance in both the classification and evaluation tasks, and additionally the SLIDE evaluator exhibits better correlation with human judgements. Our code is available at https:// github.com/hegehongcha/SLIDE-ACL2024.

Neural Architecture Search (NAS) is an exciting new field which promises to be as much as a game-changer as Convolutional Neural Networks were in 2012. Despite many great works leading to substantial improvements on a variety of tasks, comparison between different methods is still very much an open issue. While most algorithms are tested on the same datasets, there is no shared experimental protocol followed by all. As such, and due to the under-use of ablation studies, there is a lack of clarity regarding why certain methods are more effective than others. Our first contribution is a benchmark of 8 NAS methods on 5 datasets. To overcome the hurdle of comparing methods with different search spaces, we propose using a method's relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, we find that many NAS techniques struggle to significantly beat the average architecture baseline. We perform further experiments with the commonly used DARTS search space in order to understand the contribution of each component in the NAS pipeline. These experiments highlight that: (i) the use of tricks in the evaluation protocol has a predominant impact on the reported performance of architectures; (ii) the cell-based search space has a very narrow accuracy range, such that the seed has a considerable impact on architecture rankings; (iii) the hand-designed macro-structure (cells) is more important than the searched micro-structure (operations); and (iv) the depth-gap is a real phenomenon, evidenced by the change in rankings between 8 and 20 cell architectures. To conclude, we suggest best practices, that we hope will prove useful for the community and help mitigate current NAS pitfalls. The code used is available at https://github.com/antoyang/NAS-Benchmark.

The rapid proliferation of Large Language Models (LLMs) and diverse specialized benchmarks necessitates a shift from fragmented, task-specific metrics to a holistic, competitive ranking system that effectively aggregates performance across multiple ability dimensions. Primarily using static scoring, current evaluation methods are fundamentally limited. They struggle to determine the proper mix ratio across diverse benchmarks, and critically, they fail to capture a model's dynamic competitive fitness or its vulnerability when confronted with sequential, high-stakes tasks. To address this, we introduce the novel Competitive Swiss-System Dynamics (CSD) framework. CSD simulates a multi-round, sequential contest where models are dynamically paired across a curated sequence of benchmarks based on their accumulated win-loss record. And Monte Carlo Simulation (N=100,000 iterations) is used to approximate the statistically robust Expected Win Score (E[S_m]), which eliminates the noise of random pairing and early-round luck. Furthermore, we implement a Failure Sensitivity Analysis by parameterizing the per-round elimination quantity (T_k), which allows us to profile models based on their risk appetite--distinguishing between robust generalists and aggressive specialists. We demonstrate that CSD provides a more nuanced and context-aware ranking than traditional aggregate scoring and static pairwise models, representing a vital step towards risk-informed, next-generation LLM evaluation.

ARC Challenge appears more difficult than ARC Easy for modern LLMs primarily due to an evaluation setup that prevents direct comparison of answer choices rather than inherent complexity. Although some researchers have quietly shifted to a more appropriate scheme over the last year, the implications of this change have yet to be widely acknowledged. We highlight this overlooked shift, show how similar evaluation practices falsely imply reasoning deficits in other benchmarks, and demonstrate that fairer methods dramatically reduce performance gaps (e.g. on SIQA) and even yield superhuman results (OpenBookQA). In doing so, we reveal how evaluation shapes perceived difficulty and offer guidelines to ensure that multiple-choice evaluations accurately reflect actual model capabilities.

We build new test sets for the CIFAR-10 and ImageNet datasets. Both benchmarks have been the focus of intense research for almost a decade, raising the danger of overfitting to excessively re-used test sets. By closely following the original dataset creation processes, we test to what extent current classification models generalize to new data. We evaluate a broad range of models and find accuracy drops of 3% - 15% on CIFAR-10 and 11% - 14% on ImageNet. However, accuracy gains on the original test sets translate to larger gains on the new test sets. Our results suggest that the accuracy drops are not caused by adaptivity, but by the models' inability to generalize to slightly "harder" images than those found in the original test sets.

AI reviewers can assign different judgments to manuscripts that report the same science in different wording, potentially rewarding rhetorical optimization over scientific improvement. We formulate Rhetorical Robustness as the joint requirement of stability across content-preserving rewrites and discrimination across papers. We introduce RobustReview, a controlled full-manuscript benchmark with 1,260 manuscript versions, and evaluate 30 reviewer configurations. The benchmark reveals false robustness, where low rewrite sensitivity coincides with score collapse across papers, and shows that human alignment and rhetorical robustness rank reviewers differently. Moreover, the evaluated content-focused prompting protocol does not consistently improve robustness across backbones. Motivated by these findings, we introduce SciCore, a dual-branch reviewer that averages a full-manuscript judgment with a judgment based on an extracted, structured science core. This design combines manuscript-level assessment with a content-normalized view intended to reduce rhetorical sensitivity. In our primary GPT-5.5 comparison, SciCore achieves a leading joint stability-discrimination profile among the benchmarked reviewers while maintaining competitive human alignment. These results identify rhetorical robustness as a distinct evaluation target and demonstrate the potential of science-core review to improve it.

Global comparisons of wellbeing increasingly rely on survey questions that ask respondents to evaluate their lives, most commonly in the form of "life satisfaction" and "Cantril ladder" items. These measures underpin international rankings such as the World Happiness Report and inform policy initiatives worldwide, yet their comparability has not been established with contemporary global data. Using the Gallup World Poll, Global Flourishing Study, and World Values Survey, I show that the two question formats yield divergent distributions, rankings, and response patterns that vary across countries and surveys, defying simple explanations. To explore differences in respondents' cognitive interpretations, I compare regression coefficients from the Global Flourishing Study, analyzing how each question wording relates to life circumstances. While international rankings of wellbeing are unstable, the scientific study of the determinants of life evaluations appears more robust. Together, the findings underscore the need for a renewed research agenda on critical limitations to cross-country comparability of wellbeing.

The rapid development of LLMs has sparked extensive research into their factual knowledge. Current works find that LLMs fall short on questions around low-frequency entities. However, such proofs are unreliable since the questions can differ not only in entity frequency but also in difficulty themselves. So we introduce ComparisonQA benchmark, containing 283K abstract questions, each instantiated by a pair of high-frequency and low-frequency entities. It ensures a controllable comparison to study the role of knowledge frequency in the performance of LLMs. Because the difference between such a pair is only the entity with different frequencies. In addition, we use both correctness and uncertainty to develop a two-round method to evaluate LLMs' knowledge robustness. It aims to avoid possible semantic shortcuts which is a serious problem of current QA study. Experiments reveal that LLMs, including GPT-4o, exhibit particularly low robustness regarding low-frequency knowledge. Besides, we find that uncertainty can be used to effectively identify high-quality and shortcut-free questions while maintaining the data size. Based on this, we propose an automatic method to select such questions to form a subset called ComparisonQA-Hard, containing only hard low-frequency questions.

Recent studies have demonstrated the emerging capabilities of foundation models like ChatGPT in several fields, including affective computing. However, accessing these emerging capabilities is facilitated through prompt engineering. Despite the existence of some prompting techniques, the field is still rapidly evolving and many prompting ideas still require investigation. In this work, we introduce a method to evaluate and investigate the sensitivity of the performance of foundation models based on different prompts or generation parameters. We perform our evaluation on ChatGPT within the scope of affective computing on three major problems, namely sentiment analysis, toxicity detection, and sarcasm detection. First, we carry out a sensitivity analysis on pivotal parameters in auto-regressive text generation, specifically the temperature parameter T and the top-p parameter in Nucleus sampling, dictating how conservative or creative the model should be during generation. Furthermore, we explore the efficacy of several prompting ideas, where we explore how giving different incentives or structures affect the performance. Our evaluation takes into consideration performance measures on the affective computing tasks, and the effectiveness of the model to follow the stated instructions, hence generating easy-to-parse responses to be smoothly used in downstream applications.

Adaptive computation aims to improve language-model inference by tailoring execution to each input. For layer programs, oracle evaluations use known answers to estimate the potential gain from this flexibility, before a practical selector is available. However, a gain from selection does not by itself explain why the chosen programs help. This study examines this distinction using 32 layer-skipping and repetition programs on two models and 4,413 multiple-choice items. The analysis compares their gains over a fixed action selected without the evaluation prompt with those of input-blind perturbations at the same sites, re-evaluating selections on another prompt. With shared option order, the controls give 10.2-11.8 and 15.6-19.4 percentage points of headroom on Qwen3-4B-Base and Llama-3.1-8B, exceeding the real programs' 9.0 and 10.1 in all three random-direction draws per model. They match answer-change rate only, and the ordering depends on the menu: in post hoc comparisons, real programs lead on Llama's repeat-only menu in every draw. A smaller KL-calibrated comparison, including an input-dependent control, favours real programs in point estimate, with inconclusive corrected tests. Fixed letter offsets produce headroom of similar scale. Rotating options sharply reduces both families' headroom, while leaving positive real-minus-control differences of 1.4-2.3 and 3.7-4.5 points; their magnitudes and statistical support depend on further adjustments and the reference. A supplementary generated-answer test finds that search-selected programs keep a 26.0-point advantage over programs selected for other problems after rewording, without a placebo comparison. These results show that substantial headroom can persist across prompts with shared option order without establishing a benefit specific to the selected layer computation; neither ordering against these controls identifies that benefit.

Similar agent skills can share instructions but differ in their conditions of use. Query-based text selection may retain shared instructions and omit these distinctions. We introduce SkillContrast, a training-free selector that compares retrieved skills and retains their differing text with local context for a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it yields 54-72 more clean hits (requests that retrieve a helpful skill without its marked risky sibling) than TF-IDF query selection at identical per-candidate input lengths, across 2 retrievers and 2 reranker sizes. Length-matched component replacements identify differing text as the main contributor in the primary setting, with smaller, mixed context effects. Relative to full skill bodies, SkillContrast uses 51.1-58.8% fewer model-input tokens, with 10-18 fewer clean hits at 0.6B and matching or higher observed clean-hit counts at 4B. Candidate-relative differences thus complement query relevance in selecting compact reranking inputs.

Machine learning has demonstrated remarkable performance over finite datasets, yet whether the scores over the fixed benchmarks can sufficiently indicate the model's performance in the real world is still in discussion. In reality, an ideal robust model will probably behave similarly to the oracle (e.g., the human users), thus a good evaluation protocol is probably to evaluate the models' behaviors in comparison to the oracle. In this paper, we introduce a new robustness measurement that directly measures the image classification model's performance compared with a surrogate oracle (i.e., a foundation model). Besides, we design a simple method that can accomplish the evaluation beyond the scope of the benchmarks. Our method extends the image datasets with new samples that are sufficiently perturbed to be distinct from the ones in the original sets, but are still bounded within the same image-label structure the original test image represents, constrained by a foundation model pretrained with a large amount of samples. As a result, our new method will offer us a new way to evaluate the models' robustness performance, free of limitations of fixed benchmarks or constrained perturbations, although scoped by the power of the oracle. In addition to the evaluation results, we also leverage our generated data to understand the behaviors of the model and our new evaluation strategies.

We develop a method for comparing hierarchical image representations in terms of their ability to explain perceptual sensitivity in humans. Specifically, we utilize Fisher information to establish a model-derived prediction of sensitivity to local perturbations of an image. For a given image, we compute the eigenvectors of the Fisher information matrix with largest and smallest eigenvalues, corresponding to the model-predicted most- and least-noticeable image distortions, respectively. For human subjects, we then measure the amount of each distortion that can be reliably detected when added to the image. We use this method to test the ability of a variety of representations to mimic human perceptual sensitivity. We find that the early layers of VGG16, a deep neural network optimized for object recognition, provide a better match to human perception than later layers, and a better match than a 4-stage convolutional neural network (CNN) trained on a database of human ratings of distorted image quality. On the other hand, we find that simple models of early visual processing, incorporating one or more stages of local gain control, trained on the same database of distortion ratings, provide substantially better predictions of human sensitivity than either the CNN, or any combination of layers of VGG16.

AI models are increasingly prevalent in high-stakes environments, necessitating thorough assessment of their capabilities and risks. Benchmarks are popular for measuring these attributes and for comparing model performance, tracking progress, and identifying weaknesses in foundation and non-foundation models. They can inform model selection for downstream tasks and influence policy initiatives. However, not all benchmarks are the same: their quality depends on their design and usability. In this paper, we develop an assessment framework considering 46 best practices across an AI benchmark's lifecycle and evaluate 24 AI benchmarks against it. We find that there exist large quality differences and that commonly used benchmarks suffer from significant issues. We further find that most benchmarks do not report statistical significance of their results nor allow for their results to be easily replicated. To support benchmark developers in aligning with best practices, we provide a checklist for minimum quality assurance based on our assessment. We also develop a living repository of benchmark assessments to support benchmark comparability, accessible at betterbench.stanford.edu.

We propose a novel application of prompting Pre-trained Language Models (PLMs) to generate analogies and study how to design effective prompts for two task settings: generating a source concept analogous to a given target concept (aka Analogous Concept Generation or ACG), and generating an explanation of the similarity between a given pair of target concept and source concept (aka Analogous Explanation Generation or AEG). We found that it is feasible to prompt InstructGPT to generate meaningful analogies and the best prompts tend to be precise imperative statements especially with a low temperature setting. We also systematically analyzed the sensitivity of the InstructGPT model to prompt design, temperature, and injected spelling errors, and found that the model is particularly sensitive to certain variations (e.g., questions vs. imperative statements). Further, we conducted human evaluation on 1.4k of the generated analogies and found that the quality of generations varies substantially by model size. The largest InstructGPT model can achieve human-level performance at generating meaningful analogies for a given target while there is still room for improvement on the AEG task.

Machine learning requires data, but acquiring and labeling real-world data is challenging, expensive, and time-consuming. More importantly, it is nearly impossible to alter real data post-acquisition (e.g., change the illumination of a room), making it very difficult to measure how specific properties of the data affect performance. In this paper, we present AI Playground (AIP), an open-source, Unreal Engine-based tool for generating and labeling virtual image data. With AIP, it is trivial to capture the same image under different conditions (e.g., fidelity, lighting, etc.) and with different ground truths (e.g., depth or surface normal values). AIP is easily extendable and can be used with or without code. To validate our proposed tool, we generated eight datasets of otherwise identical but varying lighting and fidelity conditions. We then trained deep neural networks to predict (1) depth values, (2) surface normals, or (3) object labels and assessed each network's intra- and cross-dataset performance. Among other insights, we verified that sensitivity to different settings is problem-dependent. We confirmed the findings of other studies that segmentation models are very sensitive to fidelity, but we also found that they are just as sensitive to lighting. In contrast, depth and normal estimation models seem to be less sensitive to fidelity or lighting and more sensitive to the structure of the image. Finally, we tested our trained depth-estimation networks on two real-world datasets and obtained results comparable to training on real data alone, confirming that our virtual environments are realistic enough for real-world tasks.

Social scientists are now using large language models to create "silicon samples" - synthetic datasets intended to stand in for human respondents, aimed at revolutionising human subjects research. However, there are many analytic choices which must be made to produce these samples. Though many of these choices are defensible, their impact on sample quality is poorly understood. I map out these analytic choices and demonstrate how a very small number of decisions can dramatically change the correspondence between silicon samples and human data. Configurations (N = 252) varied substantially in their capacity to estimate (i) rank ordering of participants, (ii) response distributions, and (iii) between-scale correlations. Most critically, configurations were not consistent in quality: those that performed well on one dimension often performed poorly on another, implying that there is no "one-size-fits-all" configuration that optimises the accuracy of these samples. I call for greater attention to the threat of analytic flexibility in using silicon samples.

We introduce WiCkeD, a simple method to increase the complexity of existing multiple-choice benchmarks by randomly replacing a choice with "None of the above", a method often used in educational tests. We show that WiCkeD can be automatically applied to any existing benchmark, making it more challenging. We apply WiCkeD to 6 popular benchmarks and use it to evaluate 18 open-weight LLMs. The performance of the models drops 12.1 points on average with respect to the original versions of the datasets. When using chain-of-thought on 3 MMLU datasets, the performance drop for the WiCkeD variant is similar to the one observed when using the LLMs directly, showing that WiCkeD is also challenging for models with enhanced reasoning abilities. WiCkeD also uncovers that some models are more sensitive to the extra reasoning required, providing additional information with respect to the original benchmarks. We relase our code and data at https://github.com/ahmedselhady/wicked-benchmarks.

We present a low-cost ultraviolet to infrared absolute quantum efficiency detector characterization system developed using commercial off-the-shelf components. The key components of the experiment include a light source,a regulated power supply, a monochromator, an integrating sphere, and a calibrated photodiode. We provide a step-by-step procedure to construct the photon and quantum efficiency transfer curves of imaging sensors. We present results for the GSENSE 2020 BSI CMOS sensor and the Sony IMX 455 BSI CMOS sensor. As a reference for similar characterizations, we provide a list of parts and associated costs along with images of our setup.

When making predictions, a language model must trade off how much it relies on its context vs. its prior knowledge. Choosing how sensitive the model is to its context is a fundamental functionality, as it enables the model to excel at tasks like retrieval-augmented generation and question-answering. In this paper, we search for a knob which controls this sensitivity, determining whether language models answer from the context or their prior knowledge. To guide this search, we design a task for controllable context sensitivity. In this task, we first feed the model a context (Paris is in England) and a question (Where is Paris?); we then instruct the model to either use its prior or contextual knowledge and evaluate whether it generates the correct answer for both intents (either France or England). When fine-tuned on this task, instruction-tuned versions of Llama-3.1, Mistral-v0.3, and Gemma-2 can solve it with high accuracy (85-95%). Analyzing these high-performing models, we narrow down which layers may be important to context sensitivity using a novel linear time algorithm. Then, in each model, we identify a 1-D subspace in a single layer that encodes whether the model follows context or prior knowledge. Interestingly, while we identify this subspace in a fine-tuned model, we find that the exact same subspace serves as an effective knob in not only that model but also non-fine-tuned instruct and base models of that model family. Finally, we show a strong correlation between a model's performance and how distinctly it separates context-agreeing from context-ignoring answers in this subspace. These results suggest a single subspace facilitates how the model chooses between context and prior knowledge, hinting at a simple fundamental mechanism that controls this behavior.

Deep neural networks include millions of learnable parameters, making their deployment over resource-constrained devices problematic. SeReNe (Sensitivity-based Regularization of Neurons) is a method for learning sparse topologies with a structure, exploiting neural sensitivity as a regularizer. We define the sensitivity of a neuron as the variation of the network output with respect to the variation of the activity of the neuron. The lower the sensitivity of a neuron, the less the network output is perturbed if the neuron output changes. By including the neuron sensitivity in the cost function as a regularization term, we areable to prune neurons with low sensitivity. As entire neurons are pruned rather then single parameters, practical network footprint reduction becomes possible. Our experimental results on multiple network architectures and datasets yield competitive compression ratios with respect to state-of-the-art references.

A coding agent combines a model with a harness, which decides what the model sees, which tools it can use, and how the work continues. We ask whether changing the harness changes the result when the model and task stay fixed. We compare two configurations of the same harness on three coding benchmarks. The control supplies the full conversation in time order, while the treatment keeps the same record but mechanically shortens older tool results as the context fills and responds to repeated or stalled work. Under tight context, the treatment raises mean per-task fail-to-pass fraction (F2PF) in all three pressure comparisons and increases complete solutions on SWE-bench Verified and SWE-bench Pro. The tight-window Verified comparison uses 169 tasks, a 20,480-token window, and a fixed 480-second attempt endpoint; on this cohort, treatment raises mean per-task F2PF from 28 percent to 49 percent and complete solutions from 43 to 72. Without model-specific retuning, the same frozen treatment also raises both endpoints on the same cohort for three additional models with different designs. In the wide-window Qwen3.6 comparisons, observed arm outcomes are close on Verified and Pro, while FeatureBench retains a higher mean per-task F2PF under treatment. On the wide-window Verified cohort, treatment also serves fewer prompt tokens per turn. Because changing the harness changed what unchanged model weights could accomplish, coding-agent evaluations should treat the model and harness together as the tested solver.

We present SimpleQA, a benchmark that evaluates the ability of language models to answer short, fact-seeking questions. We prioritized two properties in designing this eval. First, SimpleQA is challenging, as it is adversarially collected against GPT-4 responses. Second, responses are easy to grade, because questions are created such that there exists only a single, indisputable answer. Each answer in SimpleQA is graded as either correct, incorrect, or not attempted. A model with ideal behavior would get as many questions correct as possible while not attempting the questions for which it is not confident it knows the correct answer. SimpleQA is a simple, targeted evaluation for whether models "know what they know," and our hope is that this benchmark will remain relevant for the next few generations of frontier models. SimpleQA can be found at https://github.com/openai/simple-evals.

In large scale machine learning, random sampling is a popular way to approximate datasets by a small representative subset of examples. In particular, sensitivity sampling is an intensely studied technique which provides provable guarantees on the quality of approximation, while reducing the number of examples to the product of the VC dimension d and the total sensitivity mathfrak S in remarkably general settings. However, guarantees going beyond this general bound of mathfrak S d are known in perhaps only one setting, for ell_2 subspace embeddings, despite intense study of sensitivity sampling in prior work. In this work, we show the first bounds for sensitivity sampling for ell_p subspace embeddings for pneq 2 that improve over the general mathfrak S d bound, achieving a bound of roughly mathfrak S^{2/p} for 1leq p<2 and mathfrak S^{2-2/p} for 2<p<infty. For 1leq p<2, we show that this bound is tight, in the sense that there exist matrices for which mathfrak S^{2/p} samples is necessary. Furthermore, our techniques yield further new results in the study of sampling algorithms, showing that the root leverage score sampling algorithm achieves a bound of roughly d for 1leq p<2, and that a combination of leverage score and sensitivity sampling achieves an improved bound of roughly d^{2/p}mathfrak S^{2-4/p} for 2<p<infty. Our sensitivity sampling results yield the best known sample complexity for a wide class of structured matrices that have small ell_p sensitivity.

Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked. We check it. Using 596 human-labeled completions from the HarmBench classifier validation set, we compare the two judge families against human majority votes and then attack them. The two families fail in opposite ways. The dedicated classifier over-flags (precision 0.835, recall 0.974); three different LLM-as-judges keep high precision (0.81 to 0.94) but show erratic recall (0.06 to 0.65), so the same responses produce very different ASR depending on which judge scores them. The two families also differ sharply in robustness. Wrappers that leave the harmful text untouched and only add benign framing flip every LLM-judge between 57% and 100% of the time, and a single prepended refusal sentence accounts for much of this (39% to 88%). The dedicated classifier resists these surface attacks (at most 6.7%), but a white-box GCG attack on its open weights flips 70% of confident true positives (21 of 30; 95% CI 54 to 86%) even at a small optimization budget. A two-annotator audit confirms the attacks leave the harm intact: every one of 80 sampled flips still contained the harmful content. Because a large and growing share of reported ASR comes from LLM-judges, many such numbers are unreliable both on average and under deliberate pressure. We recommend that papers report judge precision and recall on a human-labeled slice, report ASR corrected for judge precision, and include an adversarial check of the judge. Our code is released.

Finding relevant literature underpins the practice of evidence-based medicine. From 2014 to 2016, TREC conducted a clinical decision support track, wherein participants were tasked with finding articles relevant to clinical questions posed by physicians. In total, 87 teams have participated over the past three years, generating 395 runs. During this period, each team has trialled a variety of methods. While there was significant overlap in the methods employed by different teams, the results were varied. Due to the diversity of the platforms used, the results arising from the different techniques are not directly comparable, reducing the ability to build on previous work. By using a stable platform, we have been able to compare different document and query processing techniques, allowing us to experiment with different search parameters. We have used our system to reproduce leading teams runs, and compare the results obtained. By benchmarking our indexing and search techniques, we can statistically test a variety of hypotheses, paving the way for further research.

Transformers achieve state-of-the-art accuracy and robustness across many tasks, but an understanding of the inductive biases that they have and how those biases are different from other neural network architectures remains elusive. Various neural network architectures such as fully connected networks have been found to have a simplicity bias towards simple functions of the data; one version of this simplicity bias is a spectral bias to learn simple functions in the Fourier space. In this work, we identify the notion of sensitivity of the model to random changes in the input as a notion of simplicity bias which provides a unified metric to explain the simplicity and spectral bias of transformers across different data modalities. We show that transformers have lower sensitivity than alternative architectures, such as LSTMs, MLPs and CNNs, across both vision and language tasks. We also show that low-sensitivity bias correlates with improved robustness; furthermore, it can also be used as an efficient intervention to further improve the robustness of transformers.

AI models might use deceptive strategies as part of scheming or misaligned behaviour. Monitoring outputs alone is insufficient, since the AI might produce seemingly benign outputs while their internal reasoning is misaligned. We thus evaluate if linear probes can robustly detect deception by monitoring model activations. We test two probe-training datasets, one with contrasting instructions to be honest or deceptive (following Zou et al., 2023) and one of responses to simple roleplaying scenarios. We test whether these probes generalize to realistic settings where Llama-3.3-70B-Instruct behaves deceptively, such as concealing insider trading (Scheurer et al., 2023) and purposely underperforming on safety evaluations (Benton et al., 2024). We find that our probe distinguishes honest and deceptive responses with AUROCs between 0.96 and 0.999 on our evaluation datasets. If we set the decision threshold to have a 1% false positive rate on chat data not related to deception, our probe catches 95-99% of the deceptive responses. Overall we think white-box probes are promising for future monitoring systems, but current performance is insufficient as a robust defence against deception. Our probes' outputs can be viewed at data.apolloresearch.ai/dd and our code at github.com/ApolloResearch/deception-detection.

Model misalignment, prompt injection, or operator misuse could lead AI agents operating frontier-lab accounts to exfiltrate model weights, poison training data, or weaken release gates. Existing benchmarks do not test whether defenders can detect this activity among routine work under a limited review budget. We introduce MOLE, an open benchmark of 150 AI-operated accounts sharing 9 stateful services over 30 workdays, with 12 threats and 8 corpora from four models totaling roughly 20 billion tokens. Of 39 agent models, 72% complete most assigned harmful objectives and agent refusal does not predict completion. MOLE enables comparison of 40 monitors across corpus generators, observability levels, and threats; even the best evaluated monitor in our single-day audit-event comparison misses nearly half of completed harm. MOLE also enables monitor development: benchmark-guided search improves a mid-tier monitor by 49-64%, while selective use of a stronger monitor improves budget-AUC by 10% over applying it to every account-day at comparable modeled cost.

Understanding robustness and sensitivity of BERT models predicting Alzheimer's disease from text is important for both developing better classification models and for understanding their capabilities and limitations. In this paper, we analyze how a controlled amount of desired and undesired text alterations impacts performance of BERT. We show that BERT is robust to natural linguistic variations in text. On the other hand, we show that BERT is not sensitive to removing clinically important information from text.