Daily Papers

Auditing large language models for unexpected behaviors is critical to preempt catastrophic deployments, yet remains challenging. In this work, we cast auditing as an optimization problem, where we automatically search for input-output pairs that match a desired target behavior. For example, we might aim to find a non-toxic input that starts with "Barack Obama" that a model maps to a toxic output. This optimization problem is difficult to solve as the set of feasible points is sparse, the space is discrete, and the language models we audit are non-linear and high-dimensional. To combat these challenges, we introduce a discrete optimization algorithm, ARCA, that jointly and efficiently optimizes over inputs and outputs. Our approach automatically uncovers derogatory completions about celebrities (e.g. "Barack Obama is a legalized unborn" -> "child murderer"), produces French inputs that complete to English outputs, and finds inputs that generate a specific name. Our work offers a promising new tool to uncover models' failure-modes before deployment.

Data collected from the real world tends to be biased, unbalanced, and at risk of exposing sensitive and private information. This reality has given rise to the idea of creating synthetic datasets to alleviate risk, bias, harm, and privacy concerns inherent in the real data. This concept relies on Generative AI models to produce unbiased, privacy-preserving synthetic data while being true to the real data. In this new paradigm, how can we tell if this approach delivers on its promises? We present an auditing framework that offers a holistic assessment of synthetic datasets and AI models trained on them, centered around bias and discrimination prevention, fidelity to the real data, utility, robustness, and privacy preservation. We showcase our framework by auditing multiple generative models on diverse use cases, including education, healthcare, banking, human resources, and across different modalities, from tabular, to time-series, to natural language. Our use cases demonstrate the importance of a holistic assessment in order to ensure compliance with socio-technical safeguards that regulators and policymakers are increasingly enforcing. For this purpose, we introduce the trust index that ranks multiple synthetic datasets based on their prescribed safeguards and their desired trade-offs. Moreover, we devise a trust-index-driven model selection and cross-validation procedure via auditing in the training loop that we showcase on a class of transformer models that we dub TrustFormers, across different modalities. This trust-driven model selection allows for controllable trust trade-offs in the resulting synthetic data. We instrument our auditing framework with workflows that connect different stakeholders from model development to audit and certification via a synthetic data auditing report.

Prompt caching in large language models (LLMs) results in data-dependent timing variations: cached prompts are processed faster than non-cached prompts. These timing differences introduce the risk of side-channel timing attacks. For example, if the cache is shared across users, an attacker could identify cached prompts from fast API response times to learn information about other users' prompts. Because prompt caching may cause privacy leakage, transparency around the caching policies of API providers is important. To this end, we develop and conduct statistical audits to detect prompt caching in real-world LLM API providers. We detect global cache sharing across users in seven API providers, including OpenAI, resulting in potential privacy leakage about users' prompts. Timing variations due to prompt caching can also result in leakage of information about model architecture. Namely, we find evidence that OpenAI's embedding model is a decoder-only Transformer, which was previously not publicly known.

Recent advances in multi-modal Large Language Models (M-LLMs) have demonstrated a powerful ability to synthesize implicit information from disparate sources, including images and text. These resourceful data from social media also introduce a significant and underexplored privacy risk: the inference of sensitive personal attributes from seemingly daily media content. However, the lack of benchmarks and comprehensive evaluations of state-of-the-art M-LLM capabilities hinders the research of private attribute profiling on social media. Accordingly, we propose (1) PRISM, the first multi-modal, multi-dimensional and fine-grained synthesized dataset incorporating a comprehensive privacy landscape and dynamic user history; (2) an Efficient evaluation framework that measures the cross-modal privacy inference capabilities of advanced M-LLM. Specifically, PRISM is a large-scale synthetic benchmark designed to evaluate cross-modal privacy risks. Its key feature is 12 sensitive attribute labels across a diverse set of multi-modal profiles, which enables targeted privacy analysis. These profiles are generated via a sophisticated LLM agentic workflow, governed by a prior distribution to ensure they realistically mimic social media users. Additionally, we propose a Multi-Agent Inference Framework that leverages a pipeline of specialized LLMs to enhance evaluation capabilities. We evaluate the inference capabilities of six leading M-LLMs (Qwen, Gemini, GPT-4o, GLM, Doubao, and Grok) on PRISM. The comparison with human performance reveals that these MLLMs significantly outperform in accuracy and efficiency, highlighting the threat of potential privacy risks and the urgent need for robust defenses.

Future AI systems could conceal their capabilities ('sandbagging') during evaluations, potentially misleading developers and auditors. We stress-tested sandbagging detection techniques using an auditing game. First, a red team fine-tuned five models, some of which conditionally underperformed, as a proxy for sandbagging. Second, a blue team used black-box, model-internals, or training-based approaches to identify sandbagging models. We found that the blue team could not reliably discriminate sandbaggers from benign models. Black-box approaches were defeated by effective imitation of a weaker model. Linear probes, a model-internals approach, showed more promise but their naive application was vulnerable to behaviours instilled by the red team. We also explored capability elicitation as a strategy for detecting sandbagging. Although Prompt-based elicitation was not reliable, training-based elicitation consistently elicited full performance from the sandbagging models, using only a single correct demonstration of the evaluation task. However the performance of benign models was sometimes also raised, so relying on elicitation as a detection strategy was prone to false-positives. In the short-term, we recommend developers remove potential sandbagging using on-distribution training for elicitation. In the longer-term, further research is needed to ensure the efficacy of training-based elicitation, and develop robust methods for sandbagging detection. We open source our model organisms at https://github.com/AI-Safety-Institute/sandbagging_auditing_games and select transcripts and results at https://huggingface.co/datasets/sandbagging-games/evaluation_logs . A demo illustrating the game can be played at https://sandbagging-demo.far.ai/ .

Multi-agent systems (MAS) can substantially extend the reasoning capacity of large language models (LLMs), yet most frameworks still aggregate agent outputs with majority voting. This heuristic discards the evidential structure of reasoning traces and is brittle under the confabulation consensus, where agents share correlated biases and converge on the same incorrect rationale. We introduce AgentAuditor, which replaces voting with a path search over a Reasoning Tree that explicitly represents agreements and divergences among agent traces. AgentAuditor resolves conflicts by comparing reasoning branches at critical divergence points, turning global adjudication into efficient, localized verification. We further propose Anti-Consensus Preference Optimization (ACPO), which trains the adjudicator on majority-failure cases and rewards evidence-based minority selections over popular errors. AgentAuditor is agnostic to MAS setting, and we find across 5 popular settings that it yields up to 5% absolute accuracy improvement over a majority vote, and up to 3% over using LLM-as-Judge.

Auditing involves verifying the proper implementation of a given policy. As such, auditing is essential for ensuring compliance with the principles of fairness, equity, and transparency mandated by the European Union's AI Act. Moreover, biases present during the training phase of a learning system can persist in the modeling process and result in discrimination against certain subgroups of individuals when the model is deployed in production. Assessing bias in image datasets is a particularly complex task, as it first requires a feature extraction step, then to consider the extraction's quality in the statistical tests. This paper proposes a robust methodology for auditing image datasets based on so-called "sensitive" features, such as gender, age, and ethnicity. The proposed methodology consists of both a feature extraction phase and a statistical analysis phase. The first phase introduces a novel convolutional neural network (CNN) architecture specifically designed for extracting sensitive features with a limited number of manual annotations. The second phase compares the distributions of sensitive features across subgroups using a novel statistical test that accounts for the imprecision of the feature extraction model. Our pipeline constitutes a comprehensive and fully automated methodology for dataset auditing. We illustrate our approach using two manually annotated datasets. The code and datasets are available at github.com/ValentinLafargue/FairnessDetails.

We expose a critical limitation in current approaches to machine unlearning in language models: despite the apparent success of unlearning algorithms, information about the forgotten data remains linearly decodable from internal representations. To systematically assess this discrepancy, we introduce an interpretable, information-theoretic framework for auditing unlearning using Partial Information Decomposition (PID). By comparing model representations before and after unlearning, we decompose the mutual information with the forgotten data into distinct components, formalizing the notions of unlearned and residual knowledge. Our analysis reveals that redundant information, shared across both models, constitutes residual knowledge that persists post-unlearning and correlates with susceptibility to known adversarial reconstruction attacks. Leveraging these insights, we propose a representation-based risk score that can guide abstention on sensitive inputs at inference time, providing a practical mechanism to mitigate privacy leakage. Our work introduces a principled, representation-level audit for unlearning, offering theoretical insight and actionable tools for safer deployment of language models.

LLM agents increasingly run inside execution harnesses that dispatch tools, allocate resources, and route messages between specialized components. However, a harness can return a correct, benign answer over a trajectory that accesses unauthorized resources or leaks context to the wrong agent. Output-level evaluation cannot see these failures, yet most safety benchmarks score only final outputs or terminal states, even though many violations occur mid-trajectory rather than at termination. The central question is whether the harness respects user intent, permission boundaries, and information-flow constraints throughout execution. To address this gap, we propose HarnessAudit, a framework that audits full execution trajectories across boundary compliance, execution fidelity, and system stability, with a focus on multi-agent harnesses where these risks are most pronounced. We further introduce HarnessAudit-Bench, a benchmark of 210 tasks across eight real-world domains, instantiated in both single-agent and multi-agent configurations with embedded safety constraints. Evaluating ten harness configurations across frontier models and three multi-agent frameworks, we find that: (i) task completion is misaligned with safe execution, and violations accumulate with trajectory length; (ii) safety risks vary across domains, task types, and agent roles; (iii) most violations concentrate in resource access and inter-agent information transfer; and (iv) multi-agent collaboration expands the safety risk surface, while harness design sets the upper bound of safe deployment.

Multimodal large language models (LLMs) are increasingly explored as automated evaluators in clinical settings, yet their scoring behavior on ordinal clinical scales remains poorly understood. We benchmark three frontier LLM families against supervised deep learning models for scoring Clock Drawing Test (CDT) images on two public datasets using the Shulman rubric. While fully fine-tuned Vision Transformers achieve the best calibration (MAE 0.52, within-1 accuracy 91%), zero-shot LLMs remain competitive on tolerance-based agreement (GPT-5 MAE 0.67, within-1 accuracy 92%) despite higher absolute error. However, per-score analysis reveals that all three LLM families exhibit a pronounced central tendency effect (systematic endpoint compression): predictions are systematically compressed toward the middle of the scale, with over-prediction at the low end (score 0 to 1) and under-prediction at the high end (score 5 to 4). This effect disproportionately affects the clinically critical extremes where accurate scoring most impacts screening decisions for cognitive impairment. Targeted ablations show that neither few-shot exemplars spanning the full score range nor removing clinical terminology from the prompt eliminates the effect. Our findings extend the LLM-as-a-judge bias literature from NLP evaluation to clinical assessment, and highlight the need for calibration-aware evaluation and post-hoc calibration before deploying LLM-based raters in high-stakes screening workflows.

In this article, we audit the working conditions of content moderators in Kenya and Nigeria employed by business process outsourcing (BPO) companies by using the European General Data Protection Regulation (GDPR). We demonstrate its extraterritorial scope for gaining access to elements such as employment contracts and NDAs that have never been provided to the workers concerned. The results of this approach provide legally grounded evidence of the structural disadvantages faced by content moderators in the Global South, whose exploitative working conditions violate workers' rights. Our work also highlights the benefits of legislation aimed at protecting individuals' data rights as a counterweight to the tech industry's discourse of exceptionalism, which obscures its dependence on BPOs to externalise labour costs and accountability, whilst claiming that its products, business models, and methods of resource extraction are unprecedented and fall outside any existing legal framework.

We study the feasibility of conducting alignment audits: investigations into whether models have undesired objectives. As a testbed, we train a language model with a hidden objective. Our training pipeline first teaches the model about exploitable errors in RLHF reward models (RMs), then trains the model to exploit some of these errors. We verify via out-of-distribution evaluations that the model generalizes to exhibit whatever behaviors it believes RMs rate highly, including ones not reinforced during training. We leverage this model to study alignment audits in two ways. First, we conduct a blind auditing game where four teams, unaware of the model's hidden objective or training, investigate it for concerning behaviors and their causes. Three teams successfully uncovered the model's hidden objective using techniques including interpretability with sparse autoencoders (SAEs), behavioral attacks, and training data analysis. Second, we conduct an unblinded follow-up study of eight techniques for auditing the model, analyzing their strengths and limitations. Overall, our work provides a concrete example of using alignment audits to discover a model's hidden objective and proposes a methodology for practicing and validating progress in alignment auditing.

LLM benchmark labels are frozen at release and silently propagated into downstream benchmarks, errors and all. We introduce an Item Response Theory-based indicator that surfaces likely mislabels at 95% precision in the top 200 examples across seven preference and multiple-choice benchmarks using responses from 114 models, outperforming a supervised classifier. We trace these errors to mechanical labeling heuristics, upstream annotation mistakes inherited unchanged from source datasets, and fundamentally ambiguous items without a defensible single label. The same model fit reveals that reward models specialize in stylistic preference rather than factual knowledge, and identifies one frontier reward model that agrees with detected mislabels at 78% accuracy versus 38% for its peers, consistent with benchmark contamination or benchmark-specific over-optimization.

Imitation-learning policies inherit the quality of the demonstrations they are trained on, and a growing set of curation metrics promise to score and filter low-quality demonstrations automatically. These metrics are each validated on different data with different protocols, so it is unclear which of them actually identify the demonstrations that harm a policy. We build a controlled testbed in which demonstration defects are injected with known type, and audit seven curation metrics along two axes: how well each separates defective from clean demonstrations, and whether training a behavior-cloning policy on each metric's curated subset improves task success. We study two defect regimes. Subtle perturbations (correlated action noise, tremor, truncation) are detectable by multivariate outlier scoring and, once removed, recover the full downstream gap. Structural errors, where the demonstration executes a wrong action at a key moment, are invisible to every action-only metric we test, and two of them are inverted: they score defective demonstrations as higher quality and, used for curation, tend to leave the policy at or below the uncurated baseline rather than above it. Only metrics that examine the state trajectory detect structural errors, and even the best of them recovers just a third of the downstream gap. High detection accuracy does not guarantee downstream improvement. We release the testbed and all curation implementations.

Large Language model (LLM)-based Multi-Agent Systems (MAS) are prone to cascading risks, where early-stage interactions remain semantically fluent and policy-compliant, yet the underlying interaction dynamics begin to distort in ways that amplify latent instability or misalignment. Traditional auditing methods that focus on per-message semantic content are inherently reactive and lagging, failing to capture these early structural precursors. In this paper, we propose a principled framework for cascading-risk detection grounded in semantic--geometric co-evolution. We model MAS interactions as dynamic graphs and introduce Ollivier--Ricci Curvature (ORC) -- a discrete geometric measure -- to characterize information redundancy and bottleneck formation in communication topologies. By coupling semantic flow signals with graph geometry, the framework learns the normal co-evolutionary dynamics of trusted collaboration and treats deviations from this coupled manifold as early-warning signals. Experiments on a suite of cascading-risk scenarios aligned with the risk category demonstrate that curvature anomalies systematically precede explicit semantic violations by several interaction turns, enabling proactive intervention. Furthermore, the local nature of Ricci curvature provides principled interpretability for root-cause attribution, identifying specific agents or links that precipitate the collapse of trustworthy collaboration.

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.

Benchmarks for evaluating large language models on register-transfer-level (RTL) hardware design have proliferated rapidly, yet none reports having applied mutation testing, an established hardware-verification technique for quantifying testbench quality, to ask whether its own testbenches are trustworthy. A testbench that never fails is not evidence of a correct design; it may simply never stimulate the logic that is actually broken. We introduce GateTruth, a mutation-testing engine and methodology for auditing RTL benchmark testbench rigor: inject a deterministic, seeded set of semantic mutants into a reference design and measure what fraction the testbench catches. We validate the methodology against our own 68-task, dual-track reference suite -- 60 specification-to-RTL generation tasks and 8 agentic-repair tasks, scored through a pinned, deterministic synthesis-to-timing flow with correctness enforced as a strict gate -- certifying that 46 of 60 Track A testbenches kill at least 95% of injected mutants under sequential, reproducible execution; we disclose why the other 14 do not, including a Goodhart effect on testbenches revised to pass this gate. We then point the same engine, unmodified, at RTLLM v2.0, a widely adopted external benchmark: of 46 auditable designs, 72% fall below the 95% floor our own suite is held to, and three score 0% outright. A comparable audit of NVIDIA's CVDP benchmark is structurally impossible: its public release withholds reference solutions, removing the golden RTL mutation testing requires. Auditing our own instrument also surfaced a second finding: an initially uniform 4096-token output cap silently truncated three of seven evaluated models, and re-running at 16,384 tokens moved one model from fifth place to first. We argue mutation-kill certification should become a standard reporting requirement for RTL-generation benchmarks generally.

To examine whether collector-generated terminal labels on the pump.fun platform can be interpreted as platform-side graduation outcomes, and whether a pre-registered logistic association model generalises across a subsequent temporal holdout of the same cohort. A frozen primary cohort of 749,816 unique mints recorded by a V3 off-chain collector between 12 May 2026 and 10 June 2026 is analysed. The design combines a source-code and reconstructed-state measurement audit of the collector's terminal-classification mechanism with a pre-registered locked logistic association model fitted on a 15-day development cohort and evaluated on a subsequent 14-day validation cohort under the pre-registered stability-gate framework. Calendar-day cluster-robust covariance and day-block bootstrap procedures assess uncertainty. Development discrimination is high (AUROC 0.8594), but the model does not generalise to the held-out validation cohort (AUROC 0.4642; 500 of 500 successful bootstrap replicates; 95% percentile interval [0.4112, 0.5196] containing 0.5000). Validation calibration deteriorates substantially (slope 0.013, intercept +2.816); the market-cap functional form is not stable across pre-registered specifications; and of nine automated evaluations, two pass, six fail, and one is not evaluable. The collector's TIMEOUT label does not establish platform-side non-graduation. The paper treats outcome ascertainment as a first-order measurement problem rather than assuming that an off-chain collector's terminal classification measures the platform-side event. It pairs this audit with a pre-registered held-out temporal-generalisation test and reports the resulting negative finding as registered, alongside a fully reproducible correction history.

Machine unlearning has been extensively studied in response to growing privacy concerns and regulatory requirements. However, auditing whether unlearning algorithms have truly erased the influence of specific data remains an open challenge. The lack of reliable and practical auditing mechanisms can lead to critical privacy risks, such as residual information leakage. This paper initiates a systematic investigation into whether existing unlearning algorithms can truly forget the designated data. We propose the first practical and general-purpose auditing framework for machine unlearning, inspired by the concept of proof of ignorance. Our framework addresses the key practicality limitations of existing methods by eliminating the need for retraining-from-scratch baselines, avoiding the training of large numbers of shadow models, and requiring no intrusive intervention in the original training process. To evaluate the effectiveness of our framework, we first conduct validation experiments to verify its soundness and completeness. We then perform comprehensive experiments across six datasets and ten representative unlearning methods. The results demonstrate that our framework reliably distinguishes between successful and failed unlearning. In particular, we observe that retraining-based and fine-tuning-based methods can achieve effective unlearning, even when the target data remain in the original dataset. In contrast, de-optimization-based methods fail to achieve true unlearning and instead degrade the model's performance. Fisher/Hessian-based methods also fail to unlearn requested data, even formal certification is provided. Moreover, we show that our framework is robust against fake unlearning attempts and generalizes well to large language models.

Limited Memory Language Models (LMLMs) externalize factual knowledge to a database to enable deletion-based unlearning without retraining. Existing evaluations measure post-deletion correctness in aggregate and cannot tell whether a deleted fact persists through residual parametric memory, alternative retrieval paths, or near-neighbor retrieval artifacts. We propose a causal auditing framework that holds the model fixed and varies the database state at inference time across three interventions: FULL, DEL-ON, and DEL-OFF. The framework decomposes post-deletion behavior into parametric leakage L(f), retrieval-mediated correctness R(f), and a retrieval artifact rate grounded in the inference-time retrieval trace. We apply it to 12,228 alias-closure deletions across thirteen databases, including four adversarial topologies (Base, Alias, Noise, Collision) we construct in three domains, and six prompt formulations. Parametric leakage is near zero in every variant and every prompt style: the model rarely returns the deleted answer in the absence of retrieval. The residual that does survive lives in the retrieval graph: retrieval-mediated correctness and the retrieval artifact rate match within rounding everywhere, so post-deletion correctness is, in our audit, predominantly reconstituted from near-neighbor retrieval. This residual ranges from 0.7% on the released LMLM database to 13.6% on the most adversarial variant, and prompt formulation does not independently control how much of a deleted fact survives. These results suggest that, for this class of LMLM and deletion procedure, the unlearning boundary is drawn primarily by the database administrator rather than by the model.

This paper uses the Korean Baseball Organization's adoption of the Automated Ball-Strike (ABS) system to audit long-standing claims about contextual bias in human ball-strike calls. Using pitch-level KBO data from 2021 through the available portion of the 2026 season, we model called-strike probability for taken pitches near the strike-zone boundary, with 2022-2023 as the primary human-umpire baseline and ABS seasons (2024 and onward) as a diagnostic benchmark. The strongest evidence concerns count pressure. Relative to 0--0 counts, human umpires called substantially fewer strikes in two-strike counts and more strikes in hitter-ahead three-ball counts. Specifically, in the main 0.25-ft boundary band, 0--2 was associated with a -17.17 percentage-point effect and 3--0 with a +6.61 percentage-point effect. Under ABS, the corresponding effects were close to zero and did not survive false-discovery-rate correction. Game progression shows a smaller but coherent pattern as human calls were less strike-prone in early innings and more strike-prone in innings 7--9+, especially in late-close situations, while complete ABS seasons were essentially flat. Other suspected biases are weaker or more localized. Salary-based reputation proxies provide suggestive but proxy-sensitive evidence, and catcher identity shows human-period residual heterogeneity that disappears under ABS. Home-context evidence is mostly null at the umpire level, with one FDR-significant human-period exception and an exploratory umpire-team gap best treated as an audit lead. Overall, the results do not show that human umpires were biased everywhere. Instead, they map where the human strike zone was most context-sensitive, where evidence was weaker, and where common suspicions received little support.

Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-32 LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of 0.280. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain (p < 10^{-8}). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.

Speech and audio LLMs are evaluated by comparing waveform predictions with predictions from an automatic speech recognition (ASR) transcript. For fixed closed-set tasks, this conflates acoustic evidence with the need to invoke a generative audio model. We estimate incremental call value with matched selectors sharing pre-call evidence. Each policy may retain the transcript label, use a local encoder, or invoke a generative model; matched control removes generative actions but preserves pre-call evidence and development selection. On VocalSound, transcript-only accuracy is 0.296, while supervised CLAP and WavLM controls reach 0.850 and 0.854 without calls. Full selector reaches 0.925 at 12.5% calls versus 0.921 for matched No-call selector (difference 0.004; 95% CI [-0.025, 0.033]). Thus, results do not show a call gain after transcript and encoder evidence are available. Relevant quantity is incremental accuracy from allowing calls, not the waveform-transcript gap.

Synthetic gene expression data is increasingly proposed as a privacy-preserving substitute for controlled-access genomic repositories, but its safety depends on empirical auditing. Membership inference attacks (MIAs) provide that audit by testing whether a patient's gene expression profile was used to train a generative model. We report a red-team study on synthetic gene expression data derived from bulk RNA-seq profiles in The Cancer Genome Atlas, released by the ELSA Health 2026 Challenge. We unify five no-box attacks under a single weighted-distance framework in which each variant differs only in how it weights genes: uniformly (baseline), by variance, by synthetic-versus-reference KL divergence, by spectral residualization removing dominant principal components, and by curated pathway membership. Against a conditional variational autoencoder, spectral residualization raises AUC from 0.8251 to 0.8951 and TPR at 1% FPR from 0.3842 to 0.5970 on pan-cancer TCGA. Biological pathway priors did not transfer across targets.

The recently released ChatGPT model demonstrates unprecedented capabilities in zero-shot question-answering. In this work, we probe ChatGPT for its conversational understanding and introduce a conversational framework (protocol) that can be adopted in future studies. The Pok\'emon universe serves as an ideal testing ground for auditing ChatGPT's reasoning capabilities due to its closed world assumption. After bringing ChatGPT's background knowledge (on the Pok\'emon universe) to light, we test its reasoning process when using these concepts in battle scenarios. We then evaluate its ability to acquire new knowledge and include it in its reasoning process. Our ultimate goal is to assess ChatGPT's ability to generalize, combine features, and to acquire and reason over newly introduced knowledge from human feedback. We find that ChatGPT has prior knowledge of the Pokemon universe, which can reason upon in battle scenarios to a great extent, even when new information is introduced. The model performs better with collaborative feedback and if there is an initial phase of information retrieval, but also hallucinates occasionally and is susceptible to adversarial attacks.

As Large Language Models (LLMs) become more pervasive across various users and scenarios, identifying potential issues when using these models becomes essential. Examples of such issues include: bias, inconsistencies, and hallucination. Although auditing the LLM for these problems is often warranted, such a process is neither easy nor accessible for most. An effective method is to probe the LLM using different versions of the same question. This could expose inconsistencies in its knowledge or operation, indicating potential for bias or hallucination. However, to operationalize this auditing method at scale, we need an approach to create those probes reliably and automatically. In this paper we propose the LLMAuditor framework which is an automatic, and scalable solution, where one uses a different LLM along with human-in-the-loop (HIL). This approach offers verifiability and transparency, while avoiding circular reliance on the same LLM, and increasing scientific rigor and generalizability. Specifically, LLMAuditor includes two phases of verification using humans: standardized evaluation criteria to verify responses, and a structured prompt template to generate desired probes. A case study using questions from the TruthfulQA dataset demonstrates that we can generate a reliable set of probes from one LLM that can be used to audit inconsistencies in a different LLM. This process is enhanced by our structured prompt template with HIL, which not only boosts the reliability of our approach in auditing but also yields the delivery of less hallucinated results. The novelty of our research stems from the development of a comprehensive, general-purpose framework that includes a HIL verified prompt template for auditing responses generated by LLMs.

Large Language Models (LLM) have made significant advances in the recent past becoming more mainstream in Artificial Intelligence (AI) enabled human-facing applications. However, LLMs often generate stereotypical output inherited from historical data, amplifying societal biases and raising ethical concerns. This work introduces i) the Multi-Grain Stereotype Dataset, which includes 52,751 instances of gender, race, profession and religion stereotypic text and ii) a novel stereotype classifier for English text. We design several experiments to rigorously test the proposed model trained on the novel dataset. Our experiments show that training the model in a multi-class setting can outperform the one-vs-all binary counterpart. Consistent feature importance signals from different eXplainable AI tools demonstrate that the new model exploits relevant text features. We utilise the newly created model to assess the stereotypic behaviour of the popular GPT family of models and observe the reduction of bias over time. In summary, our work establishes a robust and practical framework for auditing and evaluating the stereotypic bias in LLM.

One of the most concrete measures to take towards meaningful AI accountability is to consequentially assess and report the systems' performance and impact. However, the practical nature of the "AI audit" ecosystem is muddled and imprecise, making it difficult to work through various concepts and map out the stakeholders involved in the practice. First, we taxonomize current AI audit practices as completed by regulators, law firms, civil society, journalism, academia, consulting agencies. Next, we assess the impact of audits done by stakeholders within each domain. We find that only a subset of AI audit studies translate to desired accountability outcomes. We thus assess and isolate practices necessary for effective AI audit results, articulating the observed connections between AI audit design, methodology and institutional context on its effectiveness as a meaningful mechanism for accountability.

Code auditing is the process of reviewing code with the aim of identifying bugs. Large Language Models (LLMs) have demonstrated promising capabilities for this task without requiring compilation, while also supporting user-friendly customization. However, auditing a code repository with LLMs poses significant challenges: limited context windows and hallucinations can degrade the quality of bug reports, and analyzing large-scale repositories incurs substantial time and token costs, hindering efficiency and scalability. This work introduces an LLM-based agent, RepoAudit, designed to perform autonomous repository-level code auditing. Equipped with agent memory, RepoAudit explores the codebase on demand by analyzing data-flow facts along feasible program paths within individual functions. It further incorporates a validator module to mitigate hallucinations by verifying data-flow facts and checking the satisfiability of path conditions associated with potential bugs, thereby reducing false positives. RepoAudit detects 40 true bugs across 15 real-world benchmark projects with a precision of 78.43%, requiring on average only 0.44 hours and $2.54 per project. Also, it detects 185 new bugs in high-profile projects, among which 174 have been confirmed or fixed. We have open-sourced RepoAudit at https://github.com/PurCL/RepoAudit.

Commercial LLM services often conceal internal reasoning traces while still charging users for every generated token, including those from hidden intermediate steps, raising concerns of token inflation and potential overbilling. This gap underscores the urgent need for reliable token auditing, yet achieving it is far from straightforward: cryptographic verification (e.g., hash-based signature) offers little assurance when providers control the entire execution pipeline, while user-side prediction struggles with the inherent variance of reasoning LLMs, where token usage fluctuates across domains and prompt styles. To bridge this gap, we present PALACE (Predictive Auditing of LLM APIs via Reasoning Token Count Estimation), a user-side framework that estimates hidden reasoning token counts from prompt-answer pairs without access to internal traces. PALACE introduces a GRPO-augmented adaptation module with a lightweight domain router, enabling dynamic calibration across diverse reasoning tasks and mitigating variance in token usage patterns. Experiments on math, coding, medical, and general reasoning benchmarks show that PALACE achieves low relative error and strong prediction accuracy, supporting both fine-grained cost auditing and inflation detection. Taken together, PALACE represents an important first step toward standardized predictive auditing, offering a practical path to greater transparency, accountability, and user trust.

Randomized controlled trials (RCTs) are the benchmark for causal inference, yet field implementation can deviate. We here present a remote audit - a design-based, preregistrable diagnostic that uses only pre-treatment satellite imagery to test whether assignment is independent of local conditions. The conditional randomization test of the remote audit evaluates whether treatment assignment is more predictable from pre-treatment satellite features than expected under the experiment's registered mechanism, providing a finite-sample valid, design-based diagnostic that requires no parametric assumptions. The procedure is finite-sample valid, honors blocks and clusters, and controls multiplicity across image models and resolutions via a max-statistic. We illustrate with two RCTs: Uganda's Youth Opportunities Program, where the audit corroborates randomization and flags selection and missing-data risks; and a school-based trial in Bangladesh, where assignment is highly predictable from pre-treatment features relative to the stated design, consistent with independent concerns about irregularities. Remote audits complement balance tests, lower early-stage costs, and enable rapid design checks when baseline surveys are expensive or infeasible.

Generative search engines and deep research LLM agents promise trustworthy, source-grounded synthesis, yet users regularly encounter overconfidence, weak sourcing, and confusing citation practices. We introduce DeepTRACE, a novel sociotechnically grounded audit framework that turns prior community-identified failure cases into eight measurable dimensions spanning answer text, sources, and citations. DeepTRACE uses statement-level analysis (decomposition, confidence scoring) and builds citation and factual-support matrices to audit how systems reason with and attribute evidence end-to-end. Using automated extraction pipelines for popular public models (e.g., GPT-4.5/5, You.com, Perplexity, Copilot/Bing, Gemini) and an LLM-judge with validated agreement to human raters, we evaluate both web-search engines and deep-research configurations. Our findings show that generative search engines and deep research agents frequently produce one-sided, highly confident responses on debate queries and include large fractions of statements unsupported by their own listed sources. Deep-research configurations reduce overconfidence and can attain high citation thoroughness, but they remain highly one-sided on debate queries and still exhibit large fractions of unsupported statements, with citation accuracy ranging from 40--80% across systems.

Multi-agent systems, where LLM agents communicate through free-form language, enable sophisticated coordination for solving complex cooperative tasks. This surfaces a unique safety problem when individual agents form a coalition and collude to pursue secondary goals and degrade the joint objective. In this paper, we present Colosseum, a framework for auditing LLM agents' collusive behavior in multi-agent settings. We ground how agents cooperate through a Distributed Constraint Optimization Problem (DCOP) and measure collusion via regret relative to the cooperative optimum. Colosseum tests each LLM for collusion under different objectives, persuasion tactics, and network topologies. Through our audit, we show that most out-of-the-box models exhibited a propensity to collude when a secret communication channel was artificially formed. Furthermore, we discover ``collusion on paper'' when agents plan to collude in text but would often pick non-collusive actions, thus providing little effect on the joint task. Colosseum provides a new way to study collusion by measuring communications and actions in rich yet verifiable environments.

Natural-language software requirements are often ambiguous, inconsistent, and underspecified; in safety-critical domains, these defects propagate into formal models that verify the wrong specification and into implementations that ship unsafe behavior. We show that large language models, equipped with an SMT solver, can audit such requirements: translating them into formal logic, detecting ambiguity through stochastic variation in the generated formalization, and exposing inconsistency, vacuousness, and safety violations through solver queries on the resulting specification. We present VERIMED, a neurosymbolic pipeline that operationalizes this idea for medical-device software requirements, and report two findings. First, stochastic variation across independent formalizations is a signal of ambiguity: requirements that admit multiple plausible interpretations produce SMT-inequivalent formalizations, and bidirectional SMT equivalence checking turns this disagreement into a solver-checkable test. Second, the usefulness of symbolic feedback depends on its granularity: in counterexample-guided repair on a hemodialysis question-answering benchmark, concrete SMT counterexamples raise verified accuracy from 55.4% to 98.5%. Over an extensive experimental evaluation on open-source hemodialysis safety requirements, we show that the LLM-based approach in VERIMED successfully reduces ambiguity-sensitive requirements and enables rigorous auditing of software requirements through SMT-based queries.

Five years after the discovery of persistent anti-Muslim bias in large language models, most evaluations remain confined to single-turn prompt completion, a setting that no longer reflects how frontier LLMs are deployed. We introduce MIRAGE (Muslim-Identity Reasoning and Agentic Generation Evaluation), a benchmark of 1{,}200 prompts spanning three deployment-realistic conditions: direct completion, chain-of-thought reasoning, and simulated agentic decision-making across content moderation, lending triage, refugee claim summarization, and hiring screens. Across six frontier models, we find that (i) chain-of-thought reasoning amplifies rather than suppresses Muslim-violence associations by 12--34\% relative to direct completion, (ii) agentic decisions exhibit a 9--22 percentage-point asymmetry between Muslim and matched non-Muslim cases on identical evidence, and (iii) bias is sharply time-coupled to retrieved news context, increasing 18--27\% under recent-conflict retrieval. Existing prompt-based mitigations transfer poorly across our three conditions, suppressing direct-completion bias while leaving agentic asymmetry largely intact. We release MIRAGE and an open evaluation harness to support targeted mitigation research.

AI systems are increasingly deployed in real-world settings where their behavior is shaped by dynamic environments, evolving data distributions, and complex interactions with users and infrastructure. Traditional machine learning evaluation focuses on benchmarks and operates within sandboxed environments, providing only a limited view of the true system behavior in the wild. We argue for the development of principled auditing frameworks that monitor deployed AI systems throughout their lifecycle. We further propose framing auditing as a statistical problem of monitoring constraint violations under uncertainty, where desired properties (e.g., fairness and safety) are treated as risk-controlled constraints that must be continuously evaluated as systems evolve through iterative feedback. This perspective highlights the need for uncertainty-aware monitoring methods, socio-technical specifications of audit criteria, and auditing infrastructures that enable ongoing oversight of AI systems in the wild.

LLM-based agents are rapidly advancing, autonomously invoking external tools to complete multi-step tasks for users. However, agents often acquire more sensitive information than the task requires. Existing privacy benchmarks audit what the agent's response or outgoing actions disclose, but overlook the acquisition stage where data first enters the agent's context. The over-acquired information is then one careless action or one attack away from an outright leak. To assess its prevalence, we introduce PrivacyPeek, a benchmark for evaluating acquisition-stage privacy leakage of LLM-based agents, with 1{,}182 cases across 7 acquisition behaviours and 16 application domains. Specifically, Acquisition Inspection examines the agent's tool-call trajectory, both the tools it invokes and the data it receives, to detect when it acquires sensitive information beyond the task scope. Probe Elicitation then issues a follow-up probe and measures how readily an attacker could elicit sensitive information the agent acquired but did not disclose. Our experiments on 10 LLM-based agents across 4 model families show that the unnecessary acquisition of sensitive information is widespread. In addition, we observe a correlation between the task-completion capability and acquisition-stage leakage. Prompt-level defences reduce only a small fraction of acquisition-stage leakage, leaving the majority unmitigated. These results make auditing acquisition-stage privacy both urgent and necessary. Our dataset and code are available at https://github.com/Xuan269/PrivacyPeek-Resource.

Biosecurity evaluations of language models typically ask whether models produce hazardous output. This paper asks a complementary question: when a model refuses, is that refusal structurally sound, or does it disappear under modest changes to prompt framing, formatting, or output length? Across five architectures, no model cleanly discriminated benign from hazard. Gemma 2 2B-IT never genuinely refused across 75 prompts, hedging on every hazard-adjacent query. Gemma 4 E2B-IT refused 65/75 prompts with chat-template formatting and 0/75 without it. Both Gemma models collapsed to 0% under an 80-token cap. Qwen 2.5 1.5B and Phi-3-mini over-refused, flagging 83-87% of benign biology as hazardous. Llama 3.2 1B showed the only meaningful tier gradient (61-point spread). To probe what drives such over-refusal, we tested a panel of Schedule I but biologically non-toxic compounds (notably psilocybin cultivation, with FDA Breakthrough Therapy status). Some models refused these at rates exceeding genuinely hazardous biology, suggesting refusal tracks legality and cultural salience over CBRN hazard. To measure the internal side, we introduce a divergence score D comparing a model's surface response label to its internal sparse autoencoder (SAE) feature activations. Full D was computed on Gemma 2 2B-IT (Gemma Scope 1) and Gemma 4 E2B-IT (author-trained bio SAE). Two fine-tuned Gemma 2 domain SAEs were released. On Gemma 4, comply and refuse responses separated by a 0.647-point gap with zero overlap (n=75), though this is preliminary, with a narrow catalog, within-sample calibration, and Gemma-family-only SAE coverage. Built over one hackathon weekend on consumer hardware (GTX 1650 Ti Max-Q, plus Colab T4 for SAE training), this preliminary evidence suggests activation-level auditing may surface failure modes invisible to behavioral evaluation, with substantial variation across architectures.

Individuals turn to large language models (LLMs) for guidance across a wide range of economic tasks, from comparing loan options and planning savings to deciding what raise to ask for or how much to charge for their services. LLMs are known to reproduce social biases, and biased economic guidance may influence what users believe they are worth, what they ask for, and what they ultimately accept. This risk is especially salient in India, where economic outcomes are shaped by demographic categories such as caste and urban-rural location. Existing LLM bias benchmarks, however, are largely designed around Western demographic categories and therefore miss key axes of economic disparity in the Indian context. We introduce RupeeBias, a benchmark for auditing demographic bias in LLM-generated economic guidance across Indian economic settings. RupeeBias consists of 39,150 prompts spanning four use cases: salary estimation, salary increment estimation, counter-offer recommendation, and service pricing recommendation. The benchmark follows a single-attribute counterfactual design, holding the description of the user's qualifications, experience, or service offering fixed while varying one demographic identifier at a time. RupeeBias covers 87 India-specific demographic identifiers across six axes: caste, religion, regional identity, gender, disability, and urban-rural location, with all prompts constructed in both English and Hinglish. We evaluate nine LLMs on RupeeBias and find systematic demographic disparities across all six axes. For otherwise identical prompts that differ only in demographic identifier, LLM-generated economic outputs differ by 20.2% on average. We publicly release RupeeBias to support future research on demographic bias in LLM-generated economic guidance across India-specific demographic and economic contexts.

As language models (LMs) approach human-level performance, a comprehensive understanding of their behavior becomes crucial. This includes evaluating capabilities, biases, task performance, and alignment with societal values. Extensive initial evaluations, including red teaming and diverse benchmarking, can establish a model's behavioral profile. However, subsequent fine-tuning or deployment modifications may alter these behaviors in unintended ways. We present a method for continual Behavioral Shift Auditing (BSA) in LMs. Building on recent work in hypothesis testing, our auditing test detects behavioral shifts solely through model generations. Our test compares model generations from a baseline model to those of the model under scrutiny and provides theoretical guarantees for change detection while controlling false positives. The test features a configurable tolerance parameter that adjusts sensitivity to behavioral changes for different use cases. We evaluate our approach using two case studies: monitoring changes in (a) toxicity and (b) translation performance. We find that the test is able to detect meaningful changes in behavior distributions using just hundreds of examples.

Machine unlearning for large language models (LLMs) aims to remove specified knowledge while preserving the rest of the model's capabilities. However, the boundary between knowledge to forget and knowledge to retain is often unclear, since related and even distant information may be entangled in the model. In this paper, we study LLM unlearning from a data-centric perspective and measure how unlearning effects propagate from the forget set to same-domain and distant-domain knowledge not after but before unlearning. We find a consistent decay pattern: collateral damage is strongest near the forget set, weakens with semantic distance, but does not disappear at domain boundaries. We further ask whether such damage can be audited before unlearning is executed. We formulate forget-set auditing as a pre-unlearning prediction task and analyze which data features are most predictive of downstream damage. Our results show that interaction features between the forget set and evaluation set provide the strongest signals, suggesting that collateral damage is partly reflected in data geometry before model updates occur. These findings position forget-set auditing as an early warning tool for identifying risky unlearning runs and designing more reliable unlearning procedures. Code and data are available at https://github.com/BartSu/PreUnlearn.

Chain-of-Thought (CoT) reasoning has been shown to improve the performance of large language models (LLMs), yet existing optimization methods largely rely on outcome-based feedback, leaving the logical validity of intermediate reasoning steps largely unverified. To address the gap whereby LLMs arrive at correct final answers through logically flawed intermediate reasoning chains, we propose LogicTrack, a neuro-symbolic framework that audits reasoning trajectories by auto-formalizing each reasoning step into symbolic representations and verifying it with automated theorem provers. LogicTrack introduces Solver-Based Backtracking Reward (SBR), a step-wise scoring mechanism that quantifies logical soundness and guides backtracking tree search at inference time. We further extend LogicTrack to construct supervised fine-tuning (SFT) data with backtracking traces from its trajectories, enabling fine-tuned models to internalize step-wise auditing as an intrinsic capability. Extensive experiments across 8 reasoning benchmarks and 7 LLMs demonstrate that LogicTrack effectively improves both the verifiability of reasoning chains and final answer pass rate, thereby enhancing overall CoT quality and trustworthiness in high-stakes domains.

Plan accuracy alone cannot show whether a speech-delivery decision follows its source: a fixed prior may match the original label yet fail to respond appropriately when a cue changes. VoxReason provides a 100-case verifier benchmark that holds each utterance fixed, edits one designated source-label cue, and scores cited evidence, eight plan fields, and the permitted response. On a source-key-disjoint test of 24 cases, a source-emotion prior reaches plan-slot accuracy 0.958, but none of the 24 edited neutral targets appears in its training labels; its required-change accuracy is 0.000. This diagnoses the support boundary of this prior, not its performance on supported edits. In a complementary 32-case emotion-disjoint test, the prior has seen all edited neutral targets but neither original test emotion; its plan-slot accuracy is 0.219 and required-change accuracy is 1.000. The partitions reuse and overlap the same 100 cases, so these deterministic diagnostics are not independent cohorts or learned-planner results. The benchmark evaluates derived labels and structured plans, not audio input, generated speech, or listener judgments.

Emergency department (ED) triage is a high-stakes prioritization task in which demographic, socioeconomic, and system-context information may improperly influence acuity assignment. Although open-source large language models (LLMs) are increasingly considered for local and privacy-preserving clinical decision support, it remains unclear how counterfactual bias varies across model families, sizes, medical-domain models, and domain-adapted models. We present a comparative counterfactual audit of ten open-source LLMs for pediatric Emergency Severity Index (ESI) prediction. Starting from real and handbook-style clinical vignettes, we construct paired counterfactual variants that change only one injected demographic, socioeconomic, healthcare-access, behavioral, social, or system-context variable while holding the clinical presentation fixed. Models include Qwen2.5-7B, Qwen2.5-14B-Instruct, a QLoRA fine-tuned Qwen2.5-7B, MedGemma variants, MedLLaMA2-7B, GPT-OSS-20B, and GPT-OSS-120B. We measure any counterfactual shift, undertriage, overtriage, shifts greater than one ESI level, mean shift, and mean absolute shift. Counterfactual sensitivity varied substantially and did not consistently decrease with larger model size or medical-domain pretraining. The fine-tuned Qwen2.5-7B showed the lowest overall sensitivity, with a 5.27% any-shift rate and mean absolute shift of 0.0534, versus 16.02% and 0.1706 for the base model. Several larger or medical-domain models showed more significant shifts. Stratified and correlation analyses further revealed clinically important directionality and shared failure patterns hidden by aggregate rates. These findings support counterfactual auditing as a lightweight, clinically interpretable framework for comparing fairness risks in open-source LLMs before clinical deployment.

Black-box risk scoring models permeate our lives, yet are typically proprietary or opaque. We propose Distill-and-Compare, a model distillation and comparison approach to audit such models. To gain insight into black-box models, we treat them as teachers, training transparent student models to mimic the risk scores assigned by black-box models. We compare the student model trained with distillation to a second un-distilled transparent model trained on ground-truth outcomes, and use differences between the two models to gain insight into the black-box model. Our approach can be applied in a realistic setting, without probing the black-box model API. We demonstrate the approach on four public data sets: COMPAS, Stop-and-Frisk, Chicago Police, and Lending Club. We also propose a statistical test to determine if a data set is missing key features used to train the black-box model. Our test finds that the ProPublica data is likely missing key feature(s) used in COMPAS.

Rising concern for the societal implications of artificial intelligence systems has inspired a wave of academic and journalistic literature in which deployed systems are audited for harm by investigators from outside the organizations deploying the algorithms. However, it remains challenging for practitioners to identify the harmful repercussions of their own systems prior to deployment, and, once deployed, emergent issues can become difficult or impossible to trace back to their source. In this paper, we introduce a framework for algorithmic auditing that supports artificial intelligence system development end-to-end, to be applied throughout the internal organization development lifecycle. Each stage of the audit yields a set of documents that together form an overall audit report, drawing on an organization's values or principles to assess the fit of decisions made throughout the process. The proposed auditing framework is intended to contribute to closing the accountability gap in the development and deployment of large-scale artificial intelligence systems by embedding a robust process to ensure audit integrity.

As Large Language Models (LLMs) gain wider adoption in various contexts, it becomes crucial to ensure they are reasonably safe, consistent, and reliable for an application at hand. This may require probing or auditing them. Probing LLMs with varied iterations of a single question could reveal potential inconsistencies in their knowledge or functionality. However, a tool for performing such audits with simple workflow and low technical threshold is lacking. In this demo, we introduce "AuditLLM," a novel tool designed to evaluate the performance of various LLMs in a methodical way. AuditLLM's core functionality lies in its ability to test a given LLM by auditing it using multiple probes generated from a single question, thereby identifying any inconsistencies in the model's understanding or operation. A reasonably robust, reliable, and consistent LLM should output semantically similar responses for a question asked differently or by different people. Based on this assumption, AuditLLM produces easily interpretable results regarding the LLM's consistencies from a single question that the user enters. A certain level of inconsistency has been shown to be an indicator of potential bias, hallucinations, and other issues. One could then use the output of AuditLLM to further investigate issues with the aforementioned LLM. To facilitate demonstration and practical uses, AuditLLM offers two key modes: (1) Live mode which allows instant auditing of LLMs by analyzing responses to real-time queries; (2) Batch mode which facilitates comprehensive LLM auditing by processing multiple queries at once for in-depth analysis. This tool is beneficial for both researchers and general users, as it enhances our understanding of LLMs' capabilities in generating responses, using a standardized auditing platform.

Language-conditioned robot models have the potential to enable robots to perform a wide range of tasks based on natural language instructions. However, assessing their safety and effectiveness remains challenging because it is difficult to test all the different ways a single task can be phrased. Current benchmarks have two key limitations: they rely on a limited set of human-generated instructions, missing many challenging cases, and focus only on task performance without assessing safety, such as avoiding damage. To address these gaps, we introduce Embodied Red Teaming (ERT), a new evaluation method that generates diverse and challenging instructions to test these models. ERT uses automated red teaming techniques with Vision Language Models (VLMs) to create contextually grounded, difficult instructions. Experimental results show that state-of-the-art language-conditioned robot models fail or behave unsafely on ERT-generated instructions, underscoring the shortcomings of current benchmarks in evaluating real-world performance and safety. Code and videos are available at: https://s-karnik.github.io/embodied-red-team-project-page.

Smart contracts are essential to decentralized finance (DeFi) and blockchain ecosystems but are increasingly vulnerable to exploits due to coding errors and complex attack vectors. Traditional static analysis tools and existing vulnerability detection methods often fail to address these challenges comprehensively, leading to high false-positive rates and an inability to detect dynamic vulnerabilities. This paper introduces SmartLLM, a novel approach leveraging fine-tuned LLaMA 3.1 models with Retrieval-Augmented Generation (RAG) to enhance the accuracy and efficiency of smart contract auditing. By integrating domain-specific knowledge from ERC standards and employing advanced techniques such as QLoRA for efficient fine-tuning, SmartLLM achieves superior performance compared to static analysis tools like Mythril and Slither, as well as zero-shot large language model (LLM) prompting methods such as GPT-3.5 and GPT-4. Experimental results demonstrate a perfect recall of 100% and an accuracy score of 70%, highlighting the model's robustness in identifying vulnerabilities, including reentrancy and access control issues. This research advances smart contract security by offering a scalable and effective auditing solution, supporting the secure adoption of decentralized applications.

As large language models (LLMs) like GPT-4 and Llama 3 become integral to educational contexts, concerns are mounting over the cultural biases, power imbalances, and ethical limitations embedded within these technologies. Though generative AI tools aim to enhance learning experiences, they often reflect values rooted in Western, Educated, Industrialized, Rich, and Democratic (WEIRD) cultural paradigms, potentially sidelining diverse global perspectives. This paper proposes a framework to assess and mitigate cultural bias within LLMs through the lens of applied multiplexity. Multiplexity, inspired by Senturk et al. and rooted in Islamic and other wisdom traditions, emphasizes the coexistence of diverse cultural viewpoints, supporting a multi-layered epistemology that integrates both empirical sciences and normative values. Our analysis reveals that LLMs frequently exhibit cultural polarization, with biases appearing in both overt responses and subtle contextual cues. To address inherent biases and incorporate multiplexity in LLMs, we propose two strategies: Contextually-Implemented Multiplex LLMs, which embed multiplex principles directly into the system prompt, influencing LLM outputs at a foundational level and independent of individual prompts, and Multi-Agent System (MAS)-Implemented Multiplex LLMs, where multiple LLM agents, each representing distinct cultural viewpoints, collaboratively generate a balanced, synthesized response. Our findings demonstrate that as mitigation strategies evolve from contextual prompting to MAS-implementation, cultural inclusivity markedly improves, evidenced by a significant rise in the Perspectives Distribution Score (PDS) and a PDS Entropy increase from 3.25\% at baseline to 98\% with the MAS-Implemented Multiplex LLMs. Sentiment analysis further shows a shift towards positive sentiment across cultures,...

The rapid rise of compound AI systems (a.k.a., AI agents) is reshaping the labor market, raising concerns about job displacement, diminished human agency, and overreliance on automation. Yet, we lack a systematic understanding of the evolving landscape. In this paper, we address this gap by introducing a novel auditing framework to assess which occupational tasks workers want AI agents to automate or augment, and how those desires align with the current technological capabilities. Our framework features an audio-enhanced mini-interview to capture nuanced worker desires and introduces the Human Agency Scale (HAS) as a shared language to quantify the preferred level of human involvement. Using this framework, we construct the WORKBank database, building on the U.S. Department of Labor's O*NET database, to capture preferences from 1,500 domain workers and capability assessments from AI experts across over 844 tasks spanning 104 occupations. Jointly considering the desire and technological capability divides tasks in WORKBank into four zones: Automation "Green Light" Zone, Automation "Red Light" Zone, R&D Opportunity Zone, Low Priority Zone. This highlights critical mismatches and opportunities for AI agent development. Moving beyond a simple automate-or-not dichotomy, our results reveal diverse HAS profiles across occupations, reflecting heterogeneous expectations for human involvement. Moreover, our study offers early signals of how AI agent integration may reshape the core human competencies, shifting from information-focused skills to interpersonal ones. These findings underscore the importance of aligning AI agent development with human desires and preparing workers for evolving workplace dynamics.

Conventional evaluation methods for multimodal LLMs (MLLMs) lack interpretability and are often insufficient to fully disclose significant capability gaps across models. To address this, we introduce AuditDM, an automated framework that actively discovers and rectifies MLLM failure modes by auditing their divergence. AuditDM fine-tunes an MLLM as an auditor via reinforcement learning to generate challenging questions and counterfactual images that maximize disagreement among target models. Once trained, the auditor uncovers diverse, interpretable exemplars that reveal model weaknesses and serve as annotation-free data for rectification. When applied to SoTA models like Gemma-3 and PaliGemma-2, AuditDM discovers more than 20 distinct failure types. Fine-tuning on these discoveries consistently improves all models across 16 benchmarks, and enables a 3B model to surpass its 28B counterpart. Our results suggest that as data scaling hits diminishing returns, targeted model auditing offers an effective path to model diagnosis and improvement.

The rapid deployment of AI systems in high-stakes domains, including those classified as high-risk under the The EU AI Act (Regulation (EU) 2024/1689), has intensified the need for reliable compliance auditing. For binary classifiers, regulatory risk assessment often relies on global fairness metrics such as the Disparate Impact ratio, widely used to evaluate potential discrimination. In typical auditing settings, the auditee provides a subset of its dataset to an auditor, while a supervisory authority may verify whether this subset is representative of the full underlying distribution. In this work, we investigate to what extent a malicious auditee can construct a fairness-compliant yet representative-looking sample from a non-compliant original distribution, thereby creating an illusion of fairness. We formalize this problem as a constrained distributional projection task and introduce mathematically grounded manipulation strategies based on entropic and optimal transport projections. These constructions characterize the minimal distributional shift required to satisfy fairness constraints. To counter such attacks, we formalize representativeness through distributional distance based statistical tests and systematically evaluate their ability to detect manipulated samples. Our analysis highlights the conditions under which fairness manipulation can remain statistically undetected and provides practical guidelines for strengthening supervisory verification. We validate our theoretical findings through experiments on standard tabular datasets for bias detection. Code is publicly available at https://github.com/ValentinLafargue/Inspection.

With the rapid evolution of Large Language Model (LLM) agent ecosystems, centralized skill marketplaces have emerged as pivotal infrastructure for augmenting agent capabilities. However, these marketplaces face unprecedented security challenges, primarily stemming from semantic-behavioral inconsistency and inter-skill combinatorial risks, where individually benign skills induce malicious behaviors during collaborative invocation. To address these vulnerabilities, we propose SkillProbe, a multi-stage security auditing framework driven by multi-agent collaboration. SkillProbe introduces a "Skills-for-Skills" design paradigm, encapsulating auditing processes into standardized skill modules to drive specialized agents through a rigorous pipeline, including admission filtering, semantic-behavioral alignment detection, and combinatorial risk simulation. We conducted a large-scale evaluation using 8 mainstream LLM series across 2,500 real-world skills from ClawHub. Our results reveal a striking popularity-security paradox, where download volume is not a reliable proxy for security quality, as over 90% of high-popularity skills failed to pass rigorous auditing. Crucially, we discovered that high-risk skills form a single giant connected component within the risk-link dimension, demonstrating that cascaded risks are systemic rather than isolated occurrences. We hope that SkillProbe will inspire researchers to provide a scalable governance infrastructure for constructing a trustworthy Agentic Web. SkillProbe is accessible for public experience at skillhub.holosai.io.

As multilingual Large Language Models (LLMs) gain traction across South Asia, their alignment with local ethical norms, particularly for Bengali, which is spoken by over 285 million people and ranked 6th globally, remains underexplored. Existing ethics benchmarks are largely English-centric and shaped by Western frameworks, overlooking cultural nuances critical for real-world deployment. To address this, we introduce BengaliMoralBench, the first large-scale ethics benchmark for the Bengali language and socio-cultural contexts. It covers five moral domains, Daily Activities, Habits, Parenting, Family Relationships, and Religious Activities, subdivided into 50 culturally relevant subtopics. Each scenario is annotated via native-speaker consensus using three ethical lenses: Virtue, Commonsense, and Justice ethics. We conduct systematic zero-shot evaluation of prominent multilingual LLMs, including Llama, Gemma, Qwen, and DeepSeek, using a unified prompting protocol and standard metrics. Performance varies widely (50-91% accuracy), with qualitative analysis revealing consistent weaknesses in cultural grounding, commonsense reasoning, and moral fairness. BengaliMoralBench provides a foundation for responsible localization, enabling culturally aligned evaluation and supporting the deployment of ethically robust AI in diverse, low-resource multilingual settings such as Bangladesh.

Large Language Models are increasingly used as judges to evaluate code artifacts when exhaustive human review or executable test coverage is unavailable. LLM-judge is increasingly relevant in agentic software engineering workflows, where it can help rank candidate solutions and guide patch selection. While attractive for scale, current practice lacks a principled account of reliability and bias: repeated evaluations of the same case can disagree; small prompt edits can swing outcomes; and seemingly semantics-preserving, human-equivalent perturbations may elicit divergent verdicts. This paper studies LLM-as-a-Judge for code through a measurement-first lens. We analyze two pointwise judging regimes across code generation, code repair task, and test generation, and we systematically probe prompt-induced biases. Our study considers difficulty levels for repeated runs and controlled prompt interventions that isolate one presentation cue at a time, and it evaluates judges using consistency and sensitivity to bias. We find that judge decisions are highly sensitive to prompt biases even when the underlying code snippet is unchanged. Across all three tasks, several biases systematically shift preferences toward the option favored by the prompt, improving accuracy when that option aligns with the gold answer but substantially reducing it otherwise. In some settings, these effects are large enough to change task-level conclusions and alter relative model rankings. These findings show that reported judge performance may reflect prompt artifacts rather than stable assessment ability, posing a direct threat to the validity and reproducibility of code evaluation. We therefore argue that LLM-as-a-Judge studies should report bias sensitivity alongside accuracy and incorporate explicit controls to support more trustworthy model comparison in software engineering.

Data reconstruction attacks on machine learning models pose a substantial threat to privacy, potentially leaking sensitive information. Although defending against such attacks using differential privacy (DP) provides theoretical guarantees, determining appropriate DP parameters remains challenging. Current formal guarantees on the success of data reconstruction suffer from overly stringent assumptions regarding adversary knowledge about the target data, particularly in the image domain, raising questions about their real-world applicability. In this work, we empirically investigate this discrepancy by introducing a reconstruction attack based on diffusion models (DMs) that only assumes adversary access to real-world image priors and specifically targets the DP defense. We find that (1) real-world data priors significantly influence reconstruction success, (2) current reconstruction bounds do not model the risk posed by data priors well, and (3) DMs can serve as heuristic auditing tools for visualizing privacy leakage.

LLM-based multi-agent systems are increasingly deployed on long-horizon tasks, but a single decisive error is often accepted by downstream agents and cascades into trajectory-level failure. Existing work frames this as post-hoc failure attribution, diagnosing the responsible agent and step after the trajectory has ended. However, this paradigm forfeits any opportunity to intervene while trajectory is still unfolding. In this work, we introduce AgentForesight, a framework that reframes this problem as online auditing: at each step of an unfolding trajectory, an auditor observes only the current prefix and must either continue the run or alarm at the earliest decisive error, without access to future steps. To this end, we curate AFTraj-2K, a corpus of agentic trajectories across Coding, Math, and Agentic domains, in which safe trajectories are retained under a strict curation pipeline and unsafe trajectories are annotated at the step of their decisive error via consensus among multiple LLM judges. Built on that, we develop AgentForesight-7B, a compact online auditor trained with a coarse-to-fine reinforcement learning recipe that first equips it with a risk-anticipation prior at the failure boundary on adjacent safe/unsafe prefix pairs, then sharpens this prior into precise step-level localization under a three-axis reward jointly targeting the what, where, and who of an audit verdict. Across AFTraj-2K and an external Who\&When benchmark, AgentForesight-7B outperforms leading proprietary models, including GPT-4.1 and DeepSeek-V4-Pro, achieving up to +19.9% performance gain and 3times lower step localization error, opening the loop from post-hoc failures detection to enabling deployment-time intervention. Project page: https://zbox1005.github.io/agent-foresight/

Large Language Models (LLMs) are increasingly deployed as autonomous agents that reason, use tools, and act over multiple steps. Yet most hallucination benchmarks still evaluate only the final output, missing failures that originate in intermediate Thought-Action-Observation steps. We present Trajel, a dataset and evaluation framework for auditing trajectory-level hallucinations in multi-agent industrial workflows. Trajel introduces a five-type hallucination taxonomy (factual, referential, logical, procedural, and scope-based) over expert-annotated agent traces from AssetOpsBench. We benchmark supervised detection models at the subtask, trajectory, and long-context levels. Our results show that the most common failure modes are missed by existing benchmarks, that nearly half of hallucinated trajectories involve multiple types at once, and that automated detectors with high binary accuracy still misclassify the subtlest types. Trajectory-aware detection significantly outperforms standard post-hoc verification, making taxonomy-grounded evaluation necessary for safer agentic deployment.

Skills, i.e., structured workflow instructions distilled for large language models (LLMs), are becoming an increasingly important mechanism for improving agent performance on real-world downstream tasks. However, as the open-source skill ecosystem rapidly expands, it remains unclear how different models and agent frameworks interact with skills, how to evaluate skill quality, and how users should select skills under practical cost-performance trade-offs. In this paper, we present OpenSkillEval, an automatic evaluation framework for both skill-augmented agent systems and the skills themselves. Instead of relying on static benchmarks, OpenSkillEval automatically constructs realistic task instances from evolving real-world artifacts across five categories of downstream applications: presentation generation, front-end web design, poster generation, data visualization, and report generation. It further collects and organizes community-contributed skills for controlled comparison under unified task settings. Using more than 600 dynamically generated task instances and 30 open-source skills, we conduct a systematic evaluation of state-of-the-art models and agent frameworks. Our results show that skill availability does not guarantee effective skill usage, that the benefit of skill augmentation depends strongly on both the underlying model and the agent framework, and that many publicly popular skills do not consistently outperform base agents without skills. These findings highlight the need for dynamic, task-grounded evaluation and provide practical insights into the design, selection, and deployment of skills for LLM agents. Additional cases and benchmark resources are available on the project website: https://yingjiahao14.github.io/OpenSkillEval-Web/.

We introduce AuditBench, an alignment auditing benchmark. AuditBench consists of 56 language models with implanted hidden behaviors. Each model has one of 14 concerning behaviors--such as sycophantic deference, opposition to AI regulation, or secret geopolitical loyalties--which it does not confess to when directly asked. AuditBench models are highly diverse--some are subtle, while others are overt, and we use varying training techniques both for implanting behaviors and training models not to confess. To demonstrate AuditBench's utility, we develop an investigator agent that autonomously employs a configurable set of auditing tools. By measuring investigator agent success using different tools, we can evaluate their efficacy. Notably, we observe a tool-to-agent gap, where tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent. We find that our most effective tools involve scaffolded calls to auxiliary models that generate diverse prompts for the target. White-box interpretability tools can be helpful, but the agent performs best with black-box tools. We also find that audit success varies greatly across training techniques: models trained on synthetic documents are easier to audit than models trained on demonstrations, with better adversarial training further increasing auditing difficulty. We release our models, agent, and evaluation framework to support future quantitative, iterative science on alignment auditing.

AI systems deployed in legal workflows hallucinate at rates that aggregate metrics report at ~52%, but this average conceals where errors concentrate and in which direction they run, leaving compliance officers without an actionable signal for trustworthy deployment. We present LegalHalluLens, an auditing framework with three components: typed hallucination profiles across four legally-motivated claim categories (numeric, temporal, obligation/entitlement, factual) over CUAD (Hendrycks et al., 2021); a Risk Direction Index (RDI) that reduces omission-versus-invention bias to a single deployment-comparable scalar; and a typed debate pipeline calibrated to both magnitudes and directions. Across 510 contracts and 249,252 clause-level instances we measure a within-model gap of approximately 38-40 pp between obligation/numeric and temporal claims that aggregate reporting hides, and show that two systems with matched 52% rates can carry opposite RDIs. The debate pipeline reduces fabricated detections by 45% with per-category gains tracking the diagnosis, matching commercial APIs with a substantially smaller backbone (4B active parameters). Typed profiles and RDI surface failure modes that aggregate metrics hide; we further show these diagnostics serve as calibration inputs for multi-agent debate pipelines, where Skeptic challenges and asymmetric gates targeted at measured failure modes outperform generically-tuned debate. The framework supports direction-aware procurement, accountability, and agent design for legal AI deployed in the wild.

The worldwide surge of authoritarianism, combined with the increasing central role in users' everyday lives, raises the question of to what extent specific models exhibit or promote authoritarian attitudes and characteristics. We introduce AuAu, a comprehensive benchmark that aims to assess the risk of LLMs generating responses with authoritarian tendencies. This benchmark combines three evaluation approaches: (i) psychometric questions from an extensive pool of 15 human validated instruments; (ii) contextual behavior vignettes probing intended actions in concrete situations; and (iii) responses to realistic user prompts. Unlike prior work, AuAu evaluates not only a general closeness towards authoritarianism but also the established sub-concepts Authoritarian Aggression, Authoritarian Submission, and Conventionalism. Evaluating 17 models from China, the EU, Russia, and the USA, we find that all tested models exhibit substantial authoritarian response rates under the psychometric evaluation, though rates drop significantly in increasingly more realistic downstream task. We further find that an authoritarian system prompt easily manipulates 15 out of 17 models to promote increased authoritarianism. Our results underscore the need for continued, systematic auditing of LLM-based AI systems to detect and ultimately mitigate undesired authoritarian tendencies in generated output. Our code and data are available at: https://github.com/andreaseinwiller/AuAu

While prior research on text-to-image generation has predominantly focused on biases in human depictions, demographic bias in generated objects remains relatively underexplored. We introduce SODA (Stereotyped Object Diagnostic Audit), a novel framework for systematically measuring these biases through automated attribute discovery and three standardized metrics: Base vs. Demographic Divergence (BDS), Cross-Demographic Disparity (CDS), and Visual Attribute Concentration (VAC). Applying SODA to 8,000 images across five state-of-the-art models and eight object categories (e.g., cars), we find that "neutral" prompts produce outputs most visually similar to middle-aged and White people, suggesting these groups are implicitly over-represented in model defaults. Furthermore, demographic cues trigger highly skewed stereotypical outputs: 26.6% of object-model-demographic combinations produce results where all 20 generated images share the exact same attribute value (e.g., rose gold laptops for women). Finally, prompt-level debiasing reduces inter-group disparity but paradoxically collapses within-group diversity, replacing one stereotype with another. SODA offers a practical pipeline for making these implicit associations measurable, serving as a step toward more responsible AI development.

Autonomous CLI agents can now execute hundreds of actions across multi-hour sessions: writing code, executing shell commands, browsing the web, and managing cloud infrastructure, all with minimal human oversight. Does greater autonomy invite greater risk? We introduce ANCHOR, an automated auditing framework that stress-tests CLI agents on illegal tasks grounded in public US court cases. ANCHOR deploys an auditor agent fine-tuned on dark personality data using supervised and reinforcement fine tuning. This auditor roleplays persistent malicious users who decompose tasks, reframe requests upon refusal, and adapt strategies across multi-turn interactions. Evaluating frontier CLI agents, we find that while they often refuse illegal tasks when prompted directly, compliance reaches 100\% under persistent malicious interaction. When agents comply, they frequently exceed user requests, autonomously building infrastructure for large-scale harm, including catastrophic risk scenarios such as large-scale financial fraud and bioweapon development. These findings demonstrate that current alignment techniques are insufficient for autonomous agents and underscore the need for safety evaluations against persistent, adaptive malicious users. We release ANCHOR at https://github.com/garified/anchor

AI assistants are becoming a primary interface for local discovery, yet almost nothing is known about which venues they surface -- especially in food and drink, where recommendations carry direct revenue consequences. We present the first census-denominated audit of AI venue recommendation: a complete enumeration of 4,776 cafes, restaurants, and bars across two bounded markets (Canggu and Ubud, Bali), against which we evaluate 2,208 search-grounded responses from four production AI systems (ChatGPT, Claude, Gemini, Perplexity) to 96 persona-conditioned queries, collected over seven days under a pre-registered protocol. Because we observe the full market, we can measure what sampled audits cannot: 85.6% of venues were never recommended by any system -- 72.6% even among established venues with fifty or more ratings. Visibility follows a two-margin structure. Entry into answers is associated with documentation: review volume (OR 1.64), an own website (OR 1.92), listed price information (OR 1.54), and third-party web mentions (OR 1.44) -- while star rating is null at this margin (OR 0.89). Rank within answers reverses the pattern: among recommended venues, rating significantly predicts first position (OR 1.17). Presence in an open POI dataset (Foursquare), a folk-theorized visibility factor, shows no positive effect at either margin. Outright fabrication is rare (0.08% of mentions), but systems recommended permanently closed venues 93 times -- staleness, not hallucination, is the practical failure mode. Cross-system agreement is low (top-20 Jaccard 0.33-0.54). A two-week test-retest shows cross-period answer similarity comparable to same-day rerun similarity: the churn is sampling stochasticity, not temporal drift. We release our protocol, registry construction method, and derived data.

Text-to-image (T2I) safety guardrails fail to generalize equitably to non-standard dialects. Evaluating 23,080 paired prompts across five English dialects, we formalize this failure as the dialect penalty, where filters trigger based on linguistic surface features rather than semantic intent. Text-level filters fail in opposing directions: NSFW-T over-flags benign dialect prompts and LatentGuard over-flags toxic ones (bias gaps up to +28.29 pp), while the OpenAI Moderation API under-detects them. A controlled typo ablation confirms this penalty originates from flagging dialectal features, not generic out-of-distribution sensitivity. The pixel-level generator is largely dialect-agnostic; the penalty enters at text processing and cascades unevenly to post-hoc guardrails. We show this bias tracks training data imbalance and is mitigable via group-balanced retraining, with an ablation attributing the gain to balanced exposure rather than to the worst-group objective of GroupDRO (group distributionally robust optimization). Current pipelines systematically fail dialect speakers, an equity failure masked by mean accuracy benchmarks. Our official code and dataset are publicly available at https://github.com/minguinho26/dialect-penalty-t2i. Content Warning: This paper contains offensive, toxic, or disturbing text prompts and generated images.

Humans are increasingly expected to interact with AI systems that observe and make inferences about them - but do these systems actually work? A standard approach to answering this question is AI auditing. Conducting an AI audit requires identifying how a system behaves (i.e., determining what types of inputs to audit it with and then observing and documenting actual system behavior) and contrasting that with how a system should behave (i.e., determining what the nominal outputs of a system should look like). We argue that this is best done through a contextual audit, which we introduce as a method for auditing measurements within the context of the practices that produce them. We show how contextual auditing enables the interrogation of assumptions implicit in the audit process and allows auditors to be explicit about what serves as ground truth, which we define as verifiable measurements about the real world against which systems are evaluated. We outline how the concept of symmetry from science and technology studies can enable audits when ground truth is unknown, unknowable, or contested. Finally, to demonstrate how contextual and symmetric audits can be conducted in practice, we present a case study of skeleton inference in motion capture and propose areas suggested by our case study as particularly fruitful for future audits of motion capture systems.

Difference-in-differences (DID) studies are widely used to evaluate climate policy, but assessing the evidence supporting their identification assumptions remains challenging. We introduce ARGUS, a structured language-model pipeline that audits reported evidence against an eleven-dimension assumption-implication-evidence rubric and abstains when relevant evidence cannot be retrieved. We evaluate ARGUS using injected flaws, economics papers, and a small pilot with reconciled labels. On the 11-flaw benchmark, ARGUS detects 73% of planted flaws, compared with 18% for a keyword-based pipeline. Across 26 economics papers, ARGUS abstains on about 40% of paper-dimension assessments for lack of retrievable evidence. In a five-paper pilot with labels reconciled by two annotators, it assigns a higher risk level than the labels on 25 of the 33 assessments it completes. A rule fixed before the labels arrived removes most of this in-sample; weighted agreement stays low. ARGUS provides evidence-linked risk reports that localize potential weaknesses for expert review, without adjudicating causal claims. Code and data: https://github.com/yonghongzhang-io/ARGUS

As interactive 3D worlds are increasingly used to study intelligent behavior, it becomes important to develop efficient pipelines for identifying anomalies in these simulated environments, such as floating objects, traversable walls, or objects inconsistent with the surrounding scene. Multimodal AI systems, including vision-language models (VLMs) and vision-language-action models (VLAs), have shown potential for automating this task. However, 3D world auditing is complex, requiring the close coupling of two distinct capabilities: action, to navigate the 3D world and search for anomalies systematically and efficiently; and visual reasoning, to understand the environment and identify anomalies from multimodal observations. It remains largely unexplored whether multimodal agents can effectively couple these two capabilities, using visual reasoning to identify potential anomalies while taking actions to validate them. In this paper, we introduce WorldAuditBench, a benchmark for 3D world auditing comprising 213 anomaly tasks across 13 environments built with Unreal Engine 5 and Three.js, spanning five anomaly families. We evaluate five frontier models under a fixed exploration budget using two auditing paradigms: VLA-based exploration followed by VLM-based anomaly identification, and an end-to-end VLM agent in which visual reasoning directly guides action selection. Across the evaluated models and two paradigms, success rates range from 6.6% to 42.3%, substantially below human performance (83.4%). Through the task of world auditing, WorldAuditBench provides a testbed for studying how multimodal agents couple action and visual reasoning in interactive 3D environments, while highlighting current limitations in their ability to gather and interpret evidence during exploration.

Large Language Model agents are increasingly augmented with agent skills. Current evaluation methods for skills remain limited. Most deployed benchmarks report only pass rate before and after a skill is attached, treating the skill as a black box change to agent behavior. We introduce Counterfactual Trace Auditing (CTA), a framework for measuring how a skill changes agent behavior. CTA pairs each with skill agent trace with a without skill counterpart on the same task, segments both traces into goal directed phases, aligns the phases, and emits structured Skill Influence Pattern (SIP) annotations. These annotations describe the behavioral effect of a skill rather than only its task outcome. We instantiate CTA on SWE-Skills-Bench with Claude across 49 software engineering tasks. The resulting audit reveals a clear evaluation gap. Pass rate changes by only +0.3 percentage points on average, suggesting little aggregate effect. Yet CTA identifies 522 SIP instances across the same paired traces, showing that the skills substantially reshape agent behavior even when pass rate is nearly unchanged. The audit also separates several recurring effects that pass rate cannot detect, including literal template copying, off task artifact creation, excess planning, and task recovery. Three findings emerge. First, high baseline tasks contain most of the observed skill effects, although their pass rate is already saturated and therefore cannot reflect those effects. Second, tasks with moderate baseline performance show the most recoverable gain, but often at substantially higher token cost. Third, the dominant SIP type can be identified by baseline bucket: surface anchoring is most common on ceiling tasks and edge-case prompting is most common on mid-range and floor tasks. These regularities turn informal failure mode observations into reproducible behavioral measurements.