Language models can find thousands of severe software vulnerabilities, and agents are increasingly being misused for cyberattacks. To avoid detection, attackers frequently distribute their misuse, splitting a harmful task across many user accounts so each individual transcript looks benign. Because safety monitors score only one agent context at a time, they are structurally blind to misuse that is only visible in aggregate, across many accounts. We show this gap is real by building, to our knowledge, the first distributed agent attack, a multi-agent scaffold that completes hard cybersecurity tasks while hiding the harmful objective across subagents with limited contexts, evading a standard monitor that catches it only a fifth as often as prior agent attacks. Towards a defense, we develop an online stateful monitor that uses real-time clustering to collect weak suspiciousness signals across many agent transcripts, and escalates only rarely to a language model that flags misuse across user accounts. In evaluations with large-scale simulated datacenter traffic, our monitor Pareto dominates standard monitors, catching distributed attacks 30

Popular explanation methods often produce unreliable feature importance scores due to missingness bias, a systematic distortion that arises when models are probed with ablated, out-of-distribution inputs. Existing solutions treat this as a deep representational flaw that requires expensive retraining or architectural modifications. In this work, we challenge this assumption and show that missingness bias can be effectively treated as a superficial artifact of the model's output space. We introduce MCal, a lightweight post-hoc method that corrects this bias by fine-tuning a simple linear head on the outputs of a frozen base model. Surprisingly, we find this simple correction consistently reduces missingness bias and is competitive with, or even outperforms, prior heavyweight approaches across diverse medical benchmarks spanning vision, language, and tabular domains.

Purpose Accurate assessment of the Critical View of Safety (CVS) during laparoscopic cholecystectomy is essential to prevent bile duct injury, a complication associated with significant morbidity and mortality. While large vision-language models (LVLMs) offer flexible reasoning, their predictions remain difficult to audit and unreliable on safety-critical surgical tasks. Methods We introduce Sum-of-Checks, a framework that decomposes each CVS criterion into expert-defined reasoning checks reflecting clinically relevant visual evidence. Given a laparoscopic frame, an LVLM evaluates each check, producing a binary judgment and justification. Criterion-level scores are computed via fixed, weighted aggregation of check outcomes. We evaluate on the Endoscapes2023 benchmark using three frontier LVLMs, comparing against direct prompting, chain-of-thought, and sub-question decomposition, each with and without few-shot examples. Results Sum-of-Checks improves average frame-level mean average precision by 12-14% relative to the best baseline across all three models and criteria. Analysis of individual checks reveals that LVLMs are reliable on observational checks (e.g., visibility, tool obstruction) but show substantial variability on decision-critical anatomical evidence. Conclusion Structuring surgical reasoning into expert-aligned verification checks improves both accuracy and transparency of LVLM-based CVS assessment, demonstrating that explicitly separating evidence elicitation from decision-making is critical for reliable and auditable surgical AI systems.

Interpretability aims to explain the behavior of deep neural networks. Despite rapid growth, there is mounting concern that much of this work has not translated into practical impact, raising questions about its relevance and utility. This position paper argues that the central missing ingredient is not new methods, but evaluation criteria: interpretability should be evaluated by actionability—the extent to which insights enable concrete decisions and interventions beyond interpretability research itself. We define actionable interpretability along two dimensions—concreteness and validation—and analyze the barriers currently preventing real-world impact. To address these barriers, we identify five domains where interpretability offers unique leverage and present a framework for actionable interpretability with evaluation criteria aligned with practical outcomes. Our goal is not to downplay exploratory research, but to establish actionability as a core objective of interpretability research.

LLM-based vulnerability detectors are increasingly deployed in CI/CD security gating, yet their resilience to evasion under syntax- and compilation-preserving edits remains poorly understood. We evaluate five attack variants spanning four carrier families of behavior-preserving code transformations on a unified C/C++ benchmark (N=5000) and introduce Complete Resistance (CR), measuring the fraction of correctly detected vulnerabilities that withstand all attack variants. Our findings reveal a significant robustness gap: models achieving 70%+ clean recall exhibit CR as low as 0.12%, meaning over 87% of detected vulnerabilities can be evaded by at least one syntax-preserving edit. Universal adversarial strings optimized on a 14B surrogate transfer effectively to black-box APIs including GPT-4o, while on-target optimization further amplifies evasion (up to 92.5% ASR). These results indicate that clean benchmark accuracy alone is insufficient as a security guarantee for deployed vulnerability detectors.

To identify safety violations, auditors often search over large sets of agent traces. This search is difficult because failures are often rare, complex, and sometimes even adversarially hidden and only detectable when multiple traces are analyzed together. These challenges arise in diverse settings such as misuse campaigns, covert sabotage, reward hacking, and prompt injection. Existing approaches struggle here for several reasons. Per-trace judges miss failures that only become visible across traces, naive agentic auditing does not scale to large trace collections, and fixed monitors are brittle to unanticipated behaviors. We introduce Meerkat, which combines clustering with agentic search to uncover violations specified in natural language. Through structured search and adaptive investigation of promising regions, Meerkat finds sparse failures without relying on seed scenarios, fixed workflows, or exhaustive enumeration. Across misuse, misalignment, and task gaming settings, Meerkat significantly improves detection of safety violations over baseline monitors, discovers widespread developer cheating on a top agent benchmark, and finds nearly 4x more examples of reward hacking on CyBench than previous audits.

We introduce NSF-SciFy, a comprehensive dataset of scientific claims and investigation proposals extracted from National Science Foundation award abstracts. While previous scientific claim verification datasets have been limited in size and scope, NSF-SciFy represents a significant advance with 2.8 million claims from 400,000 abstracts spanning all science and mathematics disciplines. We present two focused subsets: NSF-SciFy-MatSci with 114,000 claims from materials science awards, and NSF-SciFy-20K with 135,000 claims across five NSF directorates. Using zero-shot prompting, we develop a scalable approach for joint extraction of scientific claims and investigation proposals. We demonstrate the dataset’s utility through three downstream tasks: non-technical abstract generation, claim extraction, and investigation proposal extraction. Fine-tuning language models on our dataset yields substantial improvements, with relative gains often exceeding 100%, particularly for claim and proposal extraction tasks. Our error analysis reveals that extracted claims exhibit high precision but lower recall, suggesting opportunities for further methodological refinement. NSF-SciFy enables new research directions in large-scale claim verification, scientific discovery tracking, and meta-scientific analysis.

Large language models and deep research agents supply citation URLs to support their claims, yet the reliability of these citations has not been systematically measured. We address six research questions about citation URL validity using 10 models and agents on DRBench (53,090 URLs) and 3 models on ExpertQA (168,021 URLs across 32 academic fields). We find that 3–13% of citation URLs are hallucinated – they have no record in the Wayback Machine and likely never existed – while 5–18% are non-resolving overall. Deep research agents generate substantially more citations per query than search-augmented LLMs but hallucinate URLs at higher rates. Domain effects are pronounced: non-resolving rates range from 5.4% (Business) to 11.4% (Theology), with per-model effects even larger. Decomposing failures reveals that some models fabricate every non-resolving URL, while others show substantial link-rot fractions indicating genuine retrieval. As a solution, we release urlhealth, an open-source tool for URL liveness checking and stale-vs-hallucinated classification using the Wayback Machine. In agentic self-correction experiments, models equipped with urlhealth reduce non-resolving citation URLs by 6–79× to under 1%, though effectiveness depends on the model's tool-use competence. The tool and all data are publicly available. Our characterization findings, failure taxonomy, and open-source tooling establish that citation URL validity is both measurable at scale and correctable in practice.

Electrocardiograms (ECG) are electrical recordings of the heart that are critical for diagnosing cardiovascular conditions. ECG language models (ELMs) have recently emerged as a promising framework for ECG classification accompanied by report generation. However, current models cannot forecast future cardiac events despite the immense clinical value for planning earlier intervention. To address this gap, we propose CAMEL, the first ELM that is capable of inference over longer signal durations which enables its forecasting capability. Our key insight is a specialized ECG encoder which enables cross-understanding of ECG signals with text. We train CAMEL using established LLM training procedures, combining LoRA adaptation with a curriculum learning pipeline. Our curriculum includes ECG classification, metrics calculations, and multi-turn conversations to elicit reasoning. CAMEL demonstrates strong zero-shot performance across 6 tasks and 9 datasets, including ECGForecastBench, a new benchmark that we introduce for forecasting arrhythmias. CAMEL is on par with or surpasses ELMs and fully supervised baselines both in- and out-of-distribution, achieving SOTA results on ECGBench (+7.0% absolute average gain) as well as ECGForecastBench (+12.4% over fully supervised models and +21.1% over zero-shot ELMs).

Concept vectors aim to enhance model interpretability by linking internal representations with human-understandable semantics, but their practical utility is often limited by noisy and inconsistent activations. In this work, we uncover the SuperActivator Mechanism: a transformer dynamic that amplifies concept activation gaps, concentrating the most reliable concept evidence into a small set of high-activation tokens. To develop a theoretical understanding of this mechanism, we prove that concept-aligned attention heads multiplicatively amplify pairwise activation gaps, with already-extreme activations growing fastest. We find that this amplification is not just theoretical, but also occurs empirically on large-scale models: while in- and out-of-concept activation distributions overlap considerably, the in-concept distribution develops a positive tail clearly separated from the noise. These high-tail tokens, which we call SuperActivators, appear consistently across concept-positive samples, making them reliable indicators of concept presence. Accordingly, SuperActivator-based detection improves F1 by up to 0.14 over standard concept activation aggregators and prompting baselines across image and text modalities, models, layers, and concept extraction techniques, demonstrating the generality and practicality of our insights. Further empirical analysis demonstrates that the most reliable SuperActivators are sparse, with detection typically peaking when using only 5-10

Ensuring code correctness remains a challenging problem even as large language models (LLMs) become increasingly capable at code-related tasks. While LLM-based program repair systems can propose bug fixes using only a user's bug report, their effectiveness is fundamentally limited by their ability to perform fault localization (FL), a challenging problem for both humans and LLMs. Existing FL approaches rely on executable test cases, require training on costly and often noisy line-level annotations, or demand resource-intensive LLMs. In this paper, we present Bug Attention Probe (BAP), a method which learns state-of-the-art fault localization without any direct localization labels, outperforming traditional FL baselines and prompting of large-scale LLMs. We evaluate our approach across a variety of code settings, including real-world Java bugs from the standard Defects4J dataset as well as seven other datasets which span a diverse set of bug types and languages. Averaged across all eight datasets, BAP improves by 34.6 strongest baseline and 93.4 significantly more efficient than prompting, outperforming large open-weight models at a small fraction of the computational cost.

Stability guarantees have emerged as a principled way to evaluate feature attributions, but existing certification methods rely on heavily smoothed classifiers and often produce conservative guarantees. To address these limitations, we introduce soft stability and propose a simple, model-agnostic, sample-efficient stability certification algorithm (SCA) that yields non-trivial and interpretable guarantees for any attribution method. Moreover, we show that mild smoothing achieves a more favorable trade-off between accuracy and stability, avoiding the aggressive compromises made in prior certification methods. To explain this behavior, we use Boolean function analysis to derive a novel characterization of stability under smoothing. We evaluate SCA on vision and language tasks and demonstrate the effectiveness of soft stability in measuring the robustness of explanation methods.

In reasoning chains generated by large language models (LLMs), initial errors often propagate and undermine the reliability of the final conclusion. Current LLM-based error detection methods often fail to detect propagated errors because earlier errors can corrupt judgments of downstream reasoning. To better detect such errors, we introduce Autoregressive Reasoning Entailment Stability (ARES), a probabilistic framework that evaluates each reasoning step based solely on previously-verified premises. This inductive method yields a nuanced score for each step and provides certified statistical guarantees of its soundness, rather than a brittle binary label. ARES achieves state-of-the-art performance across four benchmarks (72.1% Macro-F1, +8.2 points) and demonstrates superior robustness on very long synthetic reasoning chains, where it excels at detecting propagated errors (90.3% F1, +27.6 points).

Prompt sensitivity, referring to the phenomenon where paraphrasing (that is, repeating something written or spoken using different words) leads to significant changes in large language model performance, has been widely accepted as a core limitation of large language models. In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of large language models, or is it largely an artifact of evaluation processes? To answer this question, we systematically evaluate seven large language models (for example, the GPT and Gemini families) across six benchmarks, including both multiple-choice and open-ended tasks on twelve 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 large language model 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 large language models 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.

Early event prediction (EEP) systems continuously estimate a patient's imminent risk to support clinical decision-making. For bedside trust, risk trajectories must be accurate and temporally stable, shifting only with new, relevant evidence. However, current benchmarks (a) ignore stability of risk scores and (b) evaluate mainly on tabular inputs, leaving trajectory behavior untested. To address this gap, we introduce CAREBench, an EEP benchmark that evaluates deployability using multi-modal inputs-tabular EHR, ECG waveforms, and clinical text-and assesses temporal stability alongside predictive accuracy. We propose a stability metric that quantifies short-term variability in per-patient risk and penalizes abrupt oscillations based on local-Lipschitz constants. CAREBench spans six prediction tasks such as sepsis onset and compares classical learners, deep sequence models, and zero-shot LLMs. Across tasks, existing methods, especially LLMs, struggle to jointly optimize accuracy and stability, with notably poor recall at high-precision operating points. These results highlight the need for models that produce evidence-aligned, stable trajectories to earn clinician trust in continuous monitoring settings. (Code: https://github.com/SeewonChoi/CAREBench.)

LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about captcha resolution. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.

Existing language model safety evaluations focus on overt attacks and low-stakes tasks. In reality, an attacker can easily subvert existing safeguards by requesting help on small, benign-seeming tasks across many independent queries. Because the individual queries do not appear harmful, the attack is hard to detect. However, when combined, these fragments uplift misuse by helping the attacker complete hard and dangerous tasks. Toward identifying defenses against such strategies, we develop Benchmarks for Stateful Defenses (BSD), a data generation pipeline that automates evaluations of covert attacks and corresponding defenses. Using this pipeline, we curate two new datasets that are consistently refused by frontier models and are too difficult for weaker open-weight models. This enables us to evaluate decomposition attacks, which are found to be effective misuse enablers, and to highlight stateful defenses as a promising countermeasure.

As LLMs are deployed in knowledge-intensive settings (e.g., surgery, astronomy, therapy), users are often domain experts who expect not just answers, but explanations that mirror professional reasoning. However, most automatic evaluations of explanations prioritize plausibility or faithfulness, rather than testing whether an LLM thinks like an expert. Existing approaches to evaluating professional reasoning rely heavily on per-example expert annotation, making such evaluations costly and difficult to scale. To address this gap, we introduce the T-FIX benchmark, spanning seven scientific tasks across three domains, to operationalize expert alignment as a desired attribute of LLM-generation explanations. Our framework enables automatic evaluation of expert alignment, generalizing to unseen explanations and eliminating the need for ongoing expert involvement.

Large language models' behavior is often shaped by instructions such as system prompts, refusal boundaries, privacy constraints, and tool-use rules that must hold at inference time. Yet in practice these constraints can be violated under long contexts or when user-provided context conflicts with them, creating reliability and safety risks. This motivates inference-time interventions that strengthen instruction influence without retraining. One such intervention is attention steering, which biases attention toward instruction tokens. In this work, we present a unifying theory for attention steering methods by formalizing instruction following as rule-based competition between instruction rules and context-derived rules, with attention mediating which rules dominate. We prove that boosting attention to instruction tokens tilts this competition, making it harder for context to override instruction-following. However, excessive boosting can suppress task-relevant context that should be incorporated alongside the instruction. Guided by this theory, we propose Instruction Attention Boosting (InstABoost), a simple intervention that applies a constant additive bias to instruction-key attention logits across all layers and heads. We evaluate InstABoost against prompting, latent steering, and prior attention steering methods across 15 tasks. InstABoost matches or outperforms all baselines while avoiding the fluency collapse of latent methods and the instruction over-focus of prior attention methods, achieving a stronger steering-quality tradeoff.