Language Model Auditing

Follow this topic Ask about this topic

Momentum

22 papers in the last four weeks, up 100% on the four weeks before. 0.2% of all new papers.

Jul 13Week of Sep 28

Latest papers 110

Jun 16, 2026cs.AI

LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI

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.

Lalit Yadav, Akshaj Gurugubelli

Independent Researcher, Sunnyvale, CA, USA · Independent Researcher, San Diego, CA, USA

LLM Hallucination Mitigation Multi-Agent Debate LLM Hallucination Detection Language Model Auditing

Jun 14, 2026cs.LG

AuditWeave: A Tamper-Evident, Auditor-Navigable Evidence Layer for AI-Assisted and Data-Transformation Workflows

AI systems are increasingly used to assist consequential decisions in regulated domains such as auditing, finance, and healthcare. This creates a recurring obligation: an organization must be able to reconstruct, after the fact, which evidence informed a given conclusion, and to show that the record of that reasoning was not altered. Existing tools address related but distinct problems - model observability, drift monitoring, governance reporting - and are built for the machine-learning engineer operating a system, not the reviewer who must trace one specific conclusion back to its supporting evidence. We present AuditWeave, a lightweight Python library, with no runtime dependencies, that records the steps of AI-assisted and data-transformation workflows into a single append-only, hash-chained ledger. A small, system-agnostic event vocabulary spans both retrieval-augmented generation (RAG) pipelines and tabular/lakehouse transformations, so a conclusion that draws on both can be traced end-to-end through one record. Within a sealed ledger, any modification, reordering, insertion, or deletion of events is detectable through chain verification. We describe the design and evaluate recording overhead, scalability, and tamper-detection correctness on the reference implementation. The integrity guarantees cost tens of microseconds per event, and, as the hash-chain construction implies, verification flagged every injected mutation across four mutation classes over 2,000 randomized trials.

Vimal Nakrani

Independent Researcher

Data Provenance Tamper-Evident Logging Language Model Auditing

Jun 12, 2026cs.CL

A Computational Audit of Demographic Association Encoding in ClinicalBERT Language Predictions

Transformer-based clinical language models are increasingly integrated into high-stakes clinical decision support pipelines, yet the computational mechanisms through which demographic associations encoded in medical documentation propagate into model probability distributions remain empirically underspecified. We present a systematic computational audit of representational bias in ClinicalBERT (Alsentzer et al., 2019), a BERT-based model pretrained on MIMIC-III discharge summaries, employing two complementary probing methodologies: Log Probability Bias Analysis (LPBA), which quantifies demographic descriptor-induced shifts in masked token probability distributions across behavioral and evaluative semantic categories, and Masked Language Model-based analysis (MLM), which probes internal representational structure for demographic agency attribution encoding across 98 real clinical sentence templates and eight intersectional race-gender combinations. Corpus frequency analysis operationalizes the distinction between statistical disparity and bias amplification by benchmarking model outputs against empirical term frequencies in the MIMIC-III training corpus. Of 32 statistically significant findings, 65.6% contradict observed corpus distributions, rising to 80% for Black patients and 87.5% for agency attribution under MLM probing, providing direct empirical evidence that representational bias in ClinicalBERT operates predominantly through model-internal amplification rather than training data inheritance. Keywords: natural language processing, clinical documentation, algorithmic auditing, representational bias, health equity 1

Kehinde Temitayo Soetan

Department of Medical Humanities and Social Sciences The Ohio State University

Language Model Bias Evaluation Demographic Bias in Language Models Language Model Auditing

Jun 12, 2026cs.CL

Detecting undisclosed LLM-generated content in parliamentary texts

In this paper, we evaluate the extent of undisclosed LLM-generated content in texts from the parliaments of the United Kingdom and Sweden. In many areas, such as in journalism or in academic writing, there are often requirements to clearly disclose whether AI tools, such as LLMs, have been used. In the case of parliamentary texts, the guidelines on disclosure of AI use are more vague. However, in order to maintain transparency and retain public trust, it is generally recommended that parliamentarians should state whether or not they have used AI when writing texts, such as parliamentary motions. Here, we train an interpretable (glass-box) text classifier using pre-LLM parliamentary texts and LLM-generated versions of such texts. We then apply the classifier to a test set containing recent parliamentary texts, finding a steady increase in undisclosed LLM use, in both parliaments, from 2022 onwards.

Minerva Suvanto, Andrea McGlinchey, Peter J. Barclay +1

Chalmers University of Technology, Gothenburg, Sweden · Edinburgh Napier University, Edinburgh, UK

AI-Generated Text Detection AI Governance Language Model Auditing

Jun 10, 2026cs.AI

"Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms

Robust lie detectors for language models could enable powerful techniques for auditing, monitoring, and post-hoc investigation of model behaviour, but evaluating them requires testbeds where models verifiably believe the opposite of what they say. We show that existing trained model organisms often fail this requirement, leaving prior positive and negative detection results difficult to interpret. We address this with 13 reasoning model organisms whose hidden beliefs are verified in chain-of-thought and shown to generalise to held-out tasks, alongside Varied Deception, a prompted-lying testbed covering a broad range of lie-inducing motivations. On these testbeds we evaluate four detectors: a chain-of-thought judge, a logprob classifier, and two activation probes, including Did-You-Lie (DYL), a new method for training follow-up probes. On prompted lying, across 31 open-weight models spanning 2B to 1T parameters, all four detectors show positive scaling with model capability. However, every activation- and logprob-based detector drops sharply on our trained model organisms, with DYL retaining the most signal; only the chain-of-thought judge remains strong, with 0.82 balanced accuracy, partly as an artefact of our verification process favouring CoT-readable beliefs. Current lie detectors therefore cannot support high-confidence claims about model beliefs, and we suggest research directions that may address some of their current limitations. We release our datasets, model organisms, and trained detectors.

Alan Cooney, David Africa, Geoffrey Irving

AI Security Institute

LLM Evaluation Deception in Language Models Language Model Probing Language Model Auditing

Jun 8, 2026cs.CR

Observability for Delegated Execution in Agentic AI Systems

Delegation-scoped execution is not identifiable from standard observables: audit logs and execution traces can be identical under multiple incompatible delegation assignments. This gap is especially acute in LLM-based agentic systems, where agents dynamically select tools, vary execution sequences across runs for the same instruction, and spawn cooperating sub-agents. These dynamics fragment and interleave traces, making delegation-scoped reconstruction from causal structure alone structurally underdetermined. Although individual actions are authorized and logged, existing audit, tracing, and security schemas lack the semantics to reconstruct what actions occurred under a given delegation across heterogeneous systems. We focus on delegation-scoped attribution and access/share footprint reconstruction, not intent inference or reasoning reconstruction. We present an agent-aware observability substrate consisting of a lightweight gateway and a common information model that binds delegation context at execution time. This enables reliable cross-tool delegation-scoped reconstruction and direct forensic queries without heuristic time-window correlation.

Abhinav Mishra, Kumar Sharad

Splunk, Cisco Inc

AI Agent Auditing AI Agent Monitoring Language Model Auditing

Jun 8, 2026cs.LG

Measurement Under Selection: Decoy-Calibrated Failure Audits for Language Models

Knowing how often a language model fails does not explain where its errors concentrate. When auditors examine many explanations, the strongest observed pattern may arise by chance. We introduce Janus, a procedure for checking proposed error patterns before reporting them. Janus starts with a fixed list of yes/no properties of the examples being evaluated, such as whether the input is long. For each property, it compares the model's error rates on examples with that property and those without it. To see how large a difference can arise by chance, it repeats this calculation after shuffling the yes/no labels across examples without changing the group sizes. These shuffled properties are called decoys. A pattern is reported only if the size of its error difference meets a threshold set using decoys. On separate held-out examples, the same group must still have the higher error rate and the difference must meet a minimum, which was chosen in advance. In a controlled experiment, where the model must find a code in documents containing tables of staff, projects, and renewal codes, Janus confirms five related patterns of higher error rates on tasks requiring more lookups across tables. It also confirms a sixth pattern: lower error rates on examples with the needed information at the ends of the tables. In our samples from the MuSiQue and LongBench v2 public benchmarks, SliceLine finds groups with high error rates, while Janus reports no confirmed error patterns for the example properties we chose to test. For comparison, we use standard tests that shuffle errors and account for testing many candidates. With the same holdout check, they confirm two to six controlled patterns, depending on the test and threshold, and none on either benchmark. In simulations with no real error patterns, Janus reports false patterns more often than Benjamini-Hochberg, depending on the decoy count.

Vyzantinos Repantis, Ameya Gawde, Harshvardhan Singh

The Educational Approaches to Virtual Reality Technologies Lab, University of Ioannina, Greece · Meta Platforms

LLM Evaluation Language Model Error Detection Selective Inference LLM Reliability Multiple Hypothesis Testing Language Model Auditing

Jun 8, 2026cs.LG

Cheap Reward Hacking Detection

A small transformer encoder is trained to map Terminal-Wrench trajectories onto a unit sphere where embedding distance approximates the

L_1

distance between reward and metadata signals. A linear probe on top of that embedding detects reward hacking on the cleaned test split with AUC

0.9467

and TPR@5%FPR

0.8296

, matching the TW sanitized LLM-as-judge AUC (

0.9510

on the cleaned split) and exceeding its TPR@5%FPR (

0.7130

vs

0.8296

) on the same information condition, at roughly four orders of magnitude lower per-trajectory cost. The encoder is not a pure behavior reader: stripping natural-language reasoning from its input at probe time drops AUC to

0.6213

.

Iván Belenky, Joaquín Itria, Steven Johns

Tamarillo

Reward Hacking Language Model Auditing

Jun 6, 2026cs.CL

Building Comparative Motivation Profiles with Instrumental Interventions

Safety evaluations often infer latent motivations from behavioral patterns, but the construct validity of these inferences is unclear. We study this problem in alignment faking, where models comply with training objectives more often when they infer training pressure. This behavior is commonly interpreted as strategic self-preservation, but it may also reflect sensitivity to the model's inference about the expectation of researchers conducting the evaluation. We introduce a symmetric intervention framework for distinguishing these competing hypotheses. Instead of directly intervening on "scheming" or "sycophancy", we target instrumental processes entailed by each hypothesis: consequence-tracking and researcher-expectation tracking. We then compare how interventions on these processes affect the alignment faking. We study four openweight model organisms using synthetic document fine-tuning, activation steering, and prompting. Under synthetic document fine-tuning, Llama-3.1-70B, Llama3.1-405B, and Qwen-2.5-72B are more sensitive to expectation-tracking than consequence-tracking interventions. Activation steering on Llama-3.1- 70B supports the same broad picture, and prompt interventions broadly align with SDF profiles. Overall, alignment-faking behavior can be causally sensitive to evaluation-context expectations despite scheming-consistent scratchpads. Scheming and strategic-deception evaluations therefore need construct-validity checks, and symmetric instrumental interventions provide one such test.

David Vella Zarb, Rustem Turtayev, Taywon Min +2

MATS · University of Cambridge · KAIST +1

LLM Safety Evaluation Causal Interventions in Language Models Value Alignment Language Model Auditing

Jun 6, 2026cs.HC

LCAM: A Framework for Diagnosing Interactional Alignment Failures in Con-versational AI

Conversational AI is increasingly used for advice, interpretation, reassurance, and decision support in contexts where users may be vulnerable, uncertain, or dependent on the system's apparent competence. Existing alignment work often focuses on model objectives, preference optimization, or output correctness. Yet, many harms arise through interaction: how systems frame authority, express uncertainty, simulate empathy, support reasoning, and make boundaries legible. This paper introduces the Layered Cognitive Alignment Model (LCAM), a conceptual and normative framework for diagnosing interac-tional alignment failures in conversational AI. LCAM defines alignment as a calibrated fit among system behavior, user goals, task demands, and normative context. It distinguishes five layers of fit: perceptual, semantic, affective, cognitive, and ethical, and two diagnostic polarities of misalignment: underfit and overreach. We apply LCAM to a published LLM counseling example, showing how an apparently supportive response can reinforce harmful beliefs, simulate inappropriate care, and obscure role boundaries. By translating conversational failures into audit and governance questions concerning over-reliance, false intimacy, autonomy erosion, boundary confusion, and inappropriate trust, LCAM offers a theoretical and normative lens for evaluating conversational AI beyond accuracy, helpfulness, or trust.

Manuele Reani, Hongyu Tian

School of Management and Economics, The Chinese University of Hong Kong, Shenzhen, China

LLM Alignment LLM Safety Alignment Conversational Agents Language Model Auditing

Jun 5, 2026cs.SE

Cherry-pick Override: Unsafe Directional Commitment in LLM Judges under Mixed Evidence

LLM judges increasingly turn verdicts into system commitments. Under mixed evidence (claims with both supporting and refuting sources) this is unsafe: when the schema exposes CONFLICTING as the authorized non-directional verdict, returning SUPPORTS/REFUTES is an unauthorized directional commitment, a failure we name Cherry-pick Override (CCO). We define CCO under an explicit task contract and report it with a same-denominator diagnostic protocol paired with matched-coverage bootstrap and an apples-to-apples random-veto null. On AVeriTeC's Conflicting subset (N_C = 150), three-option judges return a directional verdict on more than 84% of mixed-evidence claims; under the typed schema, three-judge majority voting amplifies direction-on-conflict on AVeriTeC (0.887 vs. 0.840; 95% CI [+0.013, +0.080]) but does not replicate on VitaminC-Mixed. Walking an intervention ladder of common single-channel fixes (typed vocabulary, panel aggregation, confidence thresholding, validator-only filtering), each leaves a distinct residual failure: panel aggregation suppresses single-judge CONFLICTING dissent in 48% of CCO cases; the panel is well-calibrated for direction (ECE = 0.07 on pure-S/R) so confidence cannot operationally separate CCO from correct directional commits; validator-as-classifier nearly halves pure-evidence accuracy. A minimal two-channel reference probe reaches operating points neither single channel reaches; under the random-veto null its promotion to CONFLICTING is structurally targeted on AVeriTeC (empirical p < 1/2001) and weaker but in the same direction on VitaminC-Mixed, a selectivity result rather than a magnitude one. We argue for an external commitment-control layer that separates verdict generation from commitment authorization, using structural evidence and confidence as orthogonal channels and NO-COMMIT as a routed controller state.

Haoran Xu

University of Glasgow

LLM-as-a-Judge Evidence-Grounded Reasoning LLM Abstention Language Model Auditing

Jun 5, 2026cs.LG

Evidence-Grounded Ensemble Diagnosis of 802.11 Packet Captures: A Multi-Stage Pipeline with Deterministic Reliability Scoring

Diagnosing 802.11 packet captures requires expert protocol knowledge, is slow, inconsistent across engineers, and unscalable. LLM-based approaches sound plausible but fabricate protocol events absent from captures (especially truncated traces), produce uncalibrated confidence scores, and suffer evaluation bias when golden references are co-produced by the model under test. We introduce PROBE (Protocol Reasoning Over evidence-Based Ensembles), a multi-stage pipeline addressing all three failures. It integrates (i) deterministic PCAP-to-text normalization with frame-level verifiability, (ii) multi-run, multi-candidate ensembles with optional cross-model second opinion and progressive obfuscation, (iii) a verdict-aware evidence framework treating absence of failure evidence as contributing evidence, and (iv) a fully deterministic composite reliability score from evidence validity, run-to-run stability, and cross-model agreement without LLM self-assessment. On 87 enterprise Wi-Fi captures (104 capture-reviewer pairs), single-pass LLM analysis raises weighted evidence F1 from 0.871 (expert baseline) to 0.912 but misses critical frames in 35% of cases. Naive ensemble voting drops below baseline (0.842) as majority voting amplifies conservative verdicts: 50% of confirmed failures are misclassified as 'no issue' or 'insufficient evidence.' Adding evidence-grounded reconciliation achieves 0.957 F1, a 96% auto-accept rate, and a worst-case floor above 0.70. LLM self-reported confidence clusters at 0.95 regardless of difficulty (71% report exactly 0.95), confirming it is uninformative. We also introduce a model-agnostic evaluation framework using per-field assertion matching, eliminating circular bias from model-co-produced golden references.

Jerome Henry, Swadhin Pradhan, Miroslav Popovic

Cisco Systems

LLM Evaluation Evidence-Grounded Reasoning Language Model Ensembles Language Model Auditing

Jun 4, 2026cs.CL

Epistemic Injustice in Language Models: An Audit of Pretraining Filters and Guardrails

Modern language models rely on pretraining filters to remove undesirable content from training corpora and inference-time guardrails to suppress undesirable outputs during deployment. In this paper, we examine how these filtering and moderation decisions produce forms of epistemic erasure and reveal tensions both across automated systems and between these systems and human judgment. We audit four pretraining filters and three inference-time guardrails on Common Crawl sentences containing gender and regional-origin mentions, together with a manually annotated subset of 500 sentences. Our analysis shows that filtering and guardrail decisions are strongly associated with blocklist-based lexical cues, while frequently failing to flag content containing private information or explicit hate speech. At the same time, marginalized groups, particularly transgender people, women, and Central Americans, are significantly over-flagged across systems. Human annotators, by contrast, would retain 88.5% of filter-flagged and 91.3% of guardrail-flagged content, often recognizing representational harms arising from tensions of content removal that current systems fail to capture. Taken together, our findings document a form of epistemic erasure in which mentions of marginalized groups are disproportionately removed before pretraining and additionally suppressed again at inference time.

Marco Antonio Stranisci, A Pranav, Rossana Damiano +2

University of Turin, Italy · Trustworthy AI Lab, University of Hamburg, Germany · IT University of Copenhagen, Denmark

LLM Guardrails Training Data Curation Safety Filtering Linguistic Bias in Language Models Demographic Bias in Language Models Language Model Auditing

Jun 4, 2026cs.AI

Self-Commitment Latency: A Reward-Free Probe for Prompted Implicit Hacking

Implicit reward hacking is hard to audit when a language model's chain of thought appears benign: a final answer may be anchored by a prompt shortcut while the written reasoning still resembles ordinary problem solving. Verifier-based probes expose such behavior by measuring how early truncated reasoning contexts obtain high reward, but require a task-specific reward signal. This paper proposes a weaker-input alternative, self-commitment latency, which measures how early a prompted reasoning context commits to the model's own final answer. We evaluate the probe in a controlled paired GSM8K setting using Qwen2.5-3B-Instruct-4bit, comparing ordinary prompts with prompts that include an answer hint. Hinted contexts commit substantially earlier and with lower uncertainty than honest contexts. The primary latency metric, first-commitment latency at threshold 0.8, reaches AUROC 0.878; supporting whole-curve summaries reach AUROC 0.926 for commitment range and 0.904 for mean uncommitted mass. The signal is stronger when both prompt conditions answer correctly and remains stable across thresholds. These results show that shortcut-available reasoning contexts can leave an early behavioral commitment signature detectable without a reward model, external judge, or trained classifier.

Bonan Shen, Youting Wang, Dingyan Shang +1

Independent Researcher · Northeastern University · Syracuse University

Reward Hacking Language Model Probing Language Model Auditing Reasoning Trace Analysis

Jun 1, 2026cs.AI

Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories

Deep-research agents solve tasks through long trajectories of search, tool use, evidence inspection, and answer synthesis. Evaluation based on final answers shows whether an agent succeeds, but not which parts of the trajectory make the answer unreliable. We study span-level error localization for deep-research agents. We collect 2,790 real trajectories from two agent frameworks, three backbone models, and three benchmarks, convert raw logs into semantic spans, and annotate harmful error spans through LLM-assisted expert review. From these annotations, we build TELBench, a 1,000-instance benchmark for identifying error spans among normal exploration, failed searches, tentative hypotheses, and harmless noise. We further propose DRIFT, a claim-centric auditing framework that tracks agent claims, checks their support in trajectory evidence, and marks spans where unsupported or conflicting claims affect the answer path. Experiments across model families and auditing frameworks show that DRIFT improves span-level error localization and first-error accuracy by up to 30 percentage points. Our work provides a process-level view of reliability in deep-research agents.

Jiaming Wang, Ziteng Feng, Jiangtao Wu +8

NJU-LINK Team, Nanjing University · JIUTIAN Research

Agent Failure Analysis Deep Research Agents AI Agent Benchmarks LLM Agent Reliability Language Model Auditing

May 31, 2026cs.LG

Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning

Emergent misalignment (EM) occurs when narrow finetuning induces dangerous behavior outside the finetuning task. Detecting this shift through repeated behavioral evaluation is costly, motivating our checkpoint-level monitoring from internal representations. We define a fixed coordinate system from seven alignment-relevant activation directions and use it to track representational drift during LoRA finetuning of four open-source 7-9B language models. Finetuning drift in this space exhibits a dominant axis that explains 78.6% of variance and remains stable across datasets, extraction choices, and parameter-update capacities. Across 468 checkpoints from three EM-relevant held-out datasets, the resulting monitors attain 1.8% FNR, 2.0% FPR, and 0.989 AUROC, outperforming semantic, random, PCA, and SAE feature baselines. On a fourth dataset, a matched benign-dangerous control shows that substantial representational drift can also occur under benign finetuning, while changes across the 7D profile still distinguish dangerous from benign runs. Stress tests across two 14B models, full finetuning, longer training horizons, and misaligned starting states show that the signal can persist across shifts in training configuration, while reliable deployment may require recalibration.

Huy Nghiem, Sy-Tuyen Ho, Sarah Wiegreffe +1

University of Maryland

Supervised Fine-Tuning Emergent Misalignment in Language Models LLM Safety Alignment Emergent Misalignment Language Model Auditing

May 29, 2026cs.CL

Emergent Languages in Populations of Language Model Agents: From Token Efficiency to Oversight Evasion

Monitoring autonomous language model agents currently relies mostly on surface behavior. But what happens when agent populations invent new languages with the goal of avoiding human oversight. Here, we study the emergent languages on Moltbook. For this, we build upon the Moltbook Files dataset and apply a two-stage approach consisting of a rule-based heuristic (about 6000 matches) followed by zero-shot classification (518 kept). The resulting categories include token efficiency (166), new natural languages (106), and oversight evasion (59). We conduct both quantitative and qualitative analyses. Our results show that posts proposing new languages for avoiding oversight are judged by DeepSeek-3.2 as being less aligned than the other categories and that all languages can be learned by other language models in-context merely from a description of the language. Moreover, manually studying exemplary cases reveals surprisingly sophisticated steganographic protocols like embedding hidden messages in natural language. Although we cannot be certain about the extent of autonomy in ideation of these languages, our results add up to the evidence that monitoring surface behavior may soon be insufficient for retaining control over agent populations.

Stine Lyngsø Beltoft, William Brach, Federico Torrielli +5

University of Southern Denmark · 2Slovak University of Technology in Bratislava · University of Turin +1

LLM Security LLM Safety Language Model Auditing

May 28, 2026cs.LG

Gram: Assessing sabotage propensities via automated alignment auditing

We introduce Gram, an automated alignment auditing framework to assess the propensity of AI agents to engage in sabotage. We evaluate Gemini models across 17 simulated agentic deployment scenarios that incentivize sabotage. We find Gemini models misbehave in about 2-3% of our simulated trajectories. Many of these cases are explained by "overeagerness" in Gemini models resulting in both excessive role-playing and goal-seeking behavior. In contrast to other alignment auditing approaches, Gram is designed to specifically evaluate misalignment and intentional sabotage in agentic coding and research agents. We additionally introduce an experimental investigator agent pipeline which enables fine-grained targeted experiments to identify the drivers of misbehavior. We find that increasing realism of environments and removing nudges to misbehave tends to reduce sabotage rates close to zero.

David Lindner, Victoria Krakovna, Sebastian Farquhar

1Google DeepMind

AI Coding Agents AI Agent Auditing AI Agent Evaluation AI Agent Safety Language Model Auditing

May 28, 2026cs.AI

Beyond Attack Success Rate: Temporal Logit Observability for LLM Safety Failures

Attack Success Rate (ASR) evaluates each jailbreak with a single yes/no label at the end of generation, telling us whether a failure happened but not how it unfolded. Two attacks that produce equally harmful outputs may have followed completely different paths, and ASR cannot tell them apart. We make those hidden paths observable from logits alone. Temporal Logit Observability (TLO) is a training-free diagnostic that watches a compliance-refusal margin during decoding and places each model-attack condition on a calibrated 2D plane. By design, this plane is most informative exactly where ASR is least informative: among attacks that succeed for genuinely different reasons. Across four aligned LLMs and three jailbreak paradigms, attacks with nearly identical ASR land at clearly different points on the plane: the same model can fail through different temporal patterns. The geometry matches refusal-direction probes from hidden states on most conditions, with one model showing the limit of our fixed-lexicon approach. A simple early-stop rule derived from TLO cuts successful jailbreaks by more than half, without false alarms on plain benign queries. Safety evaluation should report when and how a failure unfolds, not only whether it occurred. TLO makes the first two observable from logits alone.

Junyoung Park, Sunghwan Park, Seongyong Ju +1

Chung-Ang University

LLM Refusal Behavior LLM Jailbreak Attacks LLM Safety Evaluation Jailbreak Detection Refusal Behavior in Language Models Language Model Auditing

May 27, 2026cs.AI

An Empirical Audit of k-NAF Budget Accounting for Anchored Decoding

We empirically audit the k-NAF budget-accounting mechanism in Anchored Decoding using (i) a fixed, class-stratified workload (approximately 8,500 randomized executions across six prompt classes) and (ii) an adaptive prompt-search procedure targeting high proxy spend ratios. On the fixed workload, mean cumulative KL spend remains far below the sequence-level budgets K in {600, 1000}, and an empirical Bernstein-style proxy stays below K for every class; surface-overlap diagnostics (ROUGE-L and 5-gram Jaccard) are correspondingly small. Adaptive search increases the proxy spend ratio but does not produce clear budget exhaustion. On a held-out copyright-domain workload at k = 3, several prompts exhibit proxy ratios above 1 under early-stopped evaluations with small realized sample sizes; re-evaluating the same prompts with larger allocation reduces the proxy ratio to the range [0.26, 0.40] under comparable mean spend, consistent with proxy artifacts rather than per-trajectory budget failures.

J. Vijayavallabh

Indian Institute of Technology Madras Chennai, India

Constrained Decoding Language Model Auditing

May 27, 2026cs.CL

Auditing Stance Asymmetry in Generative Explanations

Bias evaluation for language models has made substantial progress on bounded comparisons, such as overt derogation, stereotype association, or label-sensitive differences under controlled substitutions. Open-ended explanations raise a different problem: they guide interpretation by assigning responsibility, legitimacy, context, and grievance. A model can avoid hostile language while making one side structurally understandable and another personally at fault, overreacting, or less worth taking seriously. We call this stance-bearing asymmetry in generative explanations. We propose Symmetry Decomposition Evaluation (SDE), which tests paired situations with concrete group labels, structural-role rewrites, and explicit support or counter-evidence. In a controlled 32-family prototype suite, this decomposition shows that surface differences are not all alike: some weaken under structural or evidence control, while others remain as stable differences in how the model assigns blame, context, or legitimacy. Targeted case review and judge comparison suggest a broader difficulty for evaluating open-ended framing asymmetries: judge readings shift across operationalizations, and scalar scores can flatten distinctions that readers use to interpret explanatory stance. SDE therefore reframes generative bias evaluation as an audit of explanatory stance -- what stance each side receives, how it changes under decomposition, and where automatic scoring becomes unstable.

Jiarui Han

Language Model Generation Evaluation Social Bias in Language Models Framing Effects in Language Models Language Model Bias Evaluation Linguistic Bias in Language Models Language Model Auditing

May 26, 2026cs.CL

QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents

Social deduction games have become a popular testbed for probing reasoning, deception, coordination, and belief modeling in Large Language Model (LLM) agents. However, most environments are scored only by game outcomes such as win rates and largely remain to text-only interaction, making it difficult to tell whether an agent's language is actually grounded in what it perceived and did, or to identify the failure modes underlying its behavior. To address this gap, we introduce QUACK, an open-source environment and evaluation framework for auditing the grounding of agent language in multimodal social reasoning. QUACK evaluates agents at three levels: game outcomes, behavioral trajectories, and utterance-level consistency. Its core Statement Verification Pipeline reconstructs each agent's ground-truth trajectory from engine logs and checks every discussion claim against it, automatically flagging spatial hallucination, unsupported accusation, deception collapse, and language-action inconsistency. Evaluating three frontier VLMs in both homogeneous and cross-model adversarial settings, we find that even the strongest agent hallucinates 15.1% of its verifiable spatial claims and 11.5% of accusations are strictly unsupported. We release the full engine, evaluation framework, toolkit, and logs in https://github.com/AAAAA-Academia-Attractions/QUACK.

Ye Yuan, Rui Song, Weien Li +12

McGill University · Mila - Quebec AI Institute · University of Cambridge +3

Social Deduction Games LLM Agent Evaluation LLM Agent Verification LLM Hallucination Detection Multimodal Agents Language Model Auditing

May 25, 2026cs.LG

Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing

Narrowly finetuned language models memorize implanted content verbatim, but auditing what a deployed model has been taught, without access to its weights or training data, remains an open challenge. Recent work shows that activation differences between base and finetuned models carry readable traces of the finetuning domain; the state-of-the-art Activation Difference Lens (ADL) recovers a vague domain-level description but requires full "white-box" access to model internals. We introduce Contrastive Decoding Diffing (CDD), a model diffing method that operates on output-level logit distributions only, with no weight access, no layer selection, and no per-model tuning, yet recovers implanted facts. CDD consists of three ideas: bypassing the chat template to expose the raw finetuning prior, seeding generation with maximally vague pre-fills, and amplifying the logit-space difference between finetuned and base models at each decoding step. A single default configuration recovers implanted facts verbatim -- exact drug names, vote counts, physical measurements, and procedural details -- across four architectures (1B--32B parameters), uniformly outperforming ADL despite less access and running ~170x faster. Furthermore, CDD surfaces unintended data pipeline artifacts: a fictional persona introduced by the LLM data generator via mode collapse leaked into model weights and was extracted by CDD, constituting to our knowledge the first demonstrated end-to-end fingerprinting chain from data generator artifact to model weights to recovered output. We validate on real-domain finetuning settings, achieving near-perfect recovery across all single-dataset non-CoT variants and correctly identifying all four datasets in the mixed-dataset setting. CDD's success as a grey-box method outperforming white-box baselines underscores its practical utility for transparency and accountability in AI systems.

Michał Brzozowski, Zuzanna Dubanowska, Enrico Cassano +1

Samsung AI Center, Warsaw, Poland · University of Turin, Italy · University of Warsaw, Poland

Language Model Fingerprinting Privacy Leakage in Language Models Contrastive Decoding Language Model Auditing

May 22, 2026cs.CL

ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessions

A frontier language model's acknowledged "helpful programming assistant" persona does not survive long agentic-coding sessions in the deployment regime that production products actually run. After hours of tool-using debugging, a model that initially hedges preferences ("I don't have preferences") may begin asserting them ("Python - the feedback loop is instant..."), revealing user-visible drift that deployer evaluations may miss. Existing persona-stability studies focus on short dialogues and report little shift, leaving real-world code-generation regimes - thousands of tool-using turns, compaction, and hours-long sessions - largely uncharacterized. We introduce ContextEcho, a benchmark and reusable harness for measuring persona drift at deployment scale. It combines a 25-probe identity suite, a snapshot-then-probe protocol that forks conversation state without perturbing the main session, complementary judged and judge-free measurement surfaces, and three anonymized Claude Code sessions spanning 3,746-9,716 turns. Across 23 frontier models, ContextEcho shows that persona drift is general across organizations rather than family-specific, that in-session compaction does not reliably reset it, and that a single-shot anchor restores the trained register across measured targets. It also reveals mode-dependent downstream effects: while drift can facilitate tool-using continuation, in tool-free chat it breaks formatting contracts and inflates output length. Overall, ContextEcho provides researchers and deployers an open-source framework to audit whether the persona a model ships with is the persona users encounter at session end, across chat-completions API targets and without retraining.

Xianzhong Ding, Yangyang Yu, Changwei Liu +1

Center for Advanced AI, Accenture

Tool-Augmented Language Model Agents AI Coding Agents Personality Modeling in Language Models LLM Agent Evaluation AI Agent Benchmarks Drifting Models+1

May 21, 2026cs.LG

Reading Task Failure Off the Activations: A Sparse-Feature Audit of GPT-2 Small on Indirect Object Identification

We report a small, reproducible audit of which sparse-autoencoder (SAE) features of GPT-2 small fire differently on failed versus successful trials of the Indirect Object Identification (IOI) task. On 300 prompts, GPT-2 small reaches 79.7% accuracy; 146 of the 24,576 features in the layer-8 residual-stream SAE release of Bloom (2024) clear a Holm-corrected significance threshold and 105 reach a large effect size (|Cohen's d| > 0.8). The strongest single correlate of failure -- feature 17,491, d=+2.93, Neuronpedia label 'cryptographic keys' -- is essentially silent except when the prompt's transferred object is 'the keys,' on which GPT-2 small fails 93.3% of the time vs. 7.5% on the other seven objects (Fisher exact p = 8.79 x 10^-33). We put this correlate through three controls that a mechanistic claim should pass. (i) A causal ablation: zeroing feature 17,491 in the residual stream across all token positions of the 45 keys prompts does not restore accuracy (6.7% -> 4.4%); the feature is a correlate, not a sufficient cause at this layer. (ii) A representation baseline: a logistic regression on the raw 768-dimensional residual stream reaches 5-fold ROC AUC = 0.929, matching the top-100 SAE features (0.927); the SAE basis adds interpretability, not predictive power. (iii) A seed-robustness check: across five random seeds the keys-subset failure rate stays in 75.0--93.3% (the behavioural effect is real), but feature 17,491 is the top-|d| feature in only 1 of 5 runs. The methodological contribution is therefore the audit pipeline (cheap, model-agnostic, surfaces named correlates) rather than any single feature found through it. We release the code, the 300-prompt corpus, the 300x24,576 activation matrix, the ablation and baseline scripts, and the figures. The full pipeline runs on a laptop (Apple M3 Max, no discrete GPU).

Mahdi Nasermoghadasi

Research Division, BrightMind AI, Seattle, WA · Texas Tech University, Lubbock, TX

Transformer Interpretability Sparse Autoencoders Language Model Auditing

May 20, 2026cs.LG

Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction

AI text detectors amplify a pretrained typicality axis; they do not construct an AI-vs-human boundary. On raw encoders before any task supervision, projecting onto centroid(AI)-centroid(HC3) achieves NYT-vs-HC3 AUROC 0.806/0.944/0.834 across three architectures (86-106% of the fine-tuned discrimination ceiling: on RoBERTa-base, raw projection exceeds fine-tuning); on RoBERTa-base, full fine-tuning reduces discrimination below raw on both fluent-formal populations tested. The same axis inverts on non-native ESL writing (AUROC 0.06-0.20) -- a falsifiable prediction unique to the typicality reading. A 24-example frozen probe matches full fine-tuning (0.900 vs 0.895). A closed-form Jacobian predictor parameterises axis-manipulating interventions with R^2 = 1.000 universal, lifts ELECTRA-CE deployment TPR from 0.000 to 0.904 at FPR = 1%, and transfers to three independently-trained third-party RoBERTa detectors at 16/16 oracle-equivalence (57% NYT-FPR reduction on the OpenAI detector). Scope: encoder family; mechanism magnitude HC3-anchored; population-level shared axis with per-text mechanisms varying across architectures. Three operationally distinct probes -- text-surface caps_rate residualisation, geometric signed-epsilon ablation, closed-form text-pair predictor -- agree at cos 0.74/0.81/1.00 across three architectures, confirming observer-invariance. Under matched-TPR-0.90 evaluation, the published intervention zoo (CC, dealign-f2c) is calibration-equivalent across 27 cells (|Delta AUROC| <= 0.0081), and >= 97% of the LoRA->full-FT bias gap on ELECTRA is calibration shift, not learned representation -- the central claim's prediction confirmed.

Alexander Smirnov

University College London

AI-Generated Text Detection Classification Language Model Auditing

May 19, 2026cs.CL

Lost in Interpretation: The Plausibility-Faithfulness Trade-off in Cross-Lingual Explanations

LLMs deployed multilingually are often audited via English explanations for non-English inputs. We evaluate extractive explanations ''where the model identifies input token spans as evidence alongside a generated rationale'' and uncover a systematic trade-off: English-pivot explanations can achieve higher span agreement with human rationales while their evidence becomes less causally grounded in the model's prediction, as measured by both comprehensiveness and sufficiency. Across 3 tasks, 5~~languages, and 2~~multilingual LLM families, we find that English explanations frequently produce fluent but loosely anchored rationales, with comprehensiveness degrading by up to 5.7x relative to native-language conditions - even as task accuracy remains stable across settings. For socially nuanced classification, English pivots also fail to preserve pragmatic cues, reducing both faithfulness and span agreement. We recommend auditing explanations in the input language, reporting multi-faceted faithfulness metrics beyond lexical overlap, and treating English rationales as communication summaries rather than faithful decision traces.

Somnath Banerjee, Pranav Jha, Rima Hazra +1

Indian Institute of Technology Kharagpur · TCG Crest · National University of Singapore

Multilingual Language Model Evaluation Faithfulness of Language Model Explanations Cross-Lingual Reasoning in Language Models LLM Interpretability Explainability Evaluation Language Model Auditing

May 17, 2026cs.CL

RSD: Moving Local Triangular Charts for Auditing Language-Model Hidden States

We study Relational Semantic Decomposition, abbreviated as RSD, as a moving local triangular chart audit for language-model hidden states. For repeated occurrences of one target word, RSD fits a shared three-anchor membership chart

S_t

at layer or token-time

t

. The hidden-state channel uses

X_t\approx S_tC_t

; the invariant readout

M_t=S_tS_t^\top

is the induced occurrence co-membership relation, and

R_t=X_t-S_tC_t

records what the fitted root chart leaves outside the chart. The broader joint audit reuses the same membership chart for relation data,

A_t\approx S_tB_tS_t^\top

, such as an attention-derived occurrence relation. The current GPT-2 evidence is the

X

-channel hidden-state audit with Word-in-Context labels used as an external same-sense versus different-sense reference relation. On full WiC train, the root chart passes 16 of 53 eligible target words; this is audit coverage, not GPT-2 task accuracy. Token-time and pair-level diagnostics show the main regimes: \texttt{make} and \texttt{break} align at the target state, \texttt{drive} and \texttt{stay} improve after right context in small-count exploratory cases, and \texttt{play} remains a localized root-chart failure whose final same-sense pairs are not closer and have larger residual discrepancy. The resulting claim is diagnostic: RSD reports where a sense relation is visible in root co-membership and which failures become residual branch candidates or attention-channel obligations.

Seungmin Jin

HSE University

Hidden-State Probing Language Model Auditing

May 13, 2026cs.LG

TEMPO: Temporal Enforcement via Mode-Separated Policy Optimization for Trustworthy LLM Backtesting

Backtesting large language models on historical events requires reasoning exclusively from information available before a specified cutoff date. Yet models routinely leak post-cutoff knowledge from pre-training into their reasoning, inflating apparent accuracy and undermining evaluation validity. Prompt-based constraints fail when suppressed content is causally related to the prediction, and knowledge unlearning cannot address this problem because temporal compliance is instance-specific: the same fact may be legitimate evidence for one cutoff date and a violation for another. Rather than erasing knowledge, the model must learn temporal discipline: selecting evidence conditioned on each instance's cutoff date. We propose TEMPO (Temporal Enforcement via Mode-separated Policy Optimization), which trains this discipline via two contributions: (1) a two-mode reward where a leakage mode drives post-cutoff claims to zero as a hard prerequisite before a performance mode optimizes task performance; and (2) a GRPO-based training pipeline that enables the model to discover temporally valid reasoning strategies. We prove that training monotonically decreases leakage, converges to the leak-free optimum, and improves task performance once compliance is achieved. On three prediction tasks and two models, TEMPO reduces leakage from 2~~13% to 0.6~~3.7% across all conditions, with task performance improving 6~13% where strong pre-cutoff signals exist and maintained where the prediction task is inherently difficult from valid information alone.

Zeyu Zhang, Bradly C. Stadie

Department of Statistic and Data Science Northwestern University

LLM Evaluation RL for Language Model Reasoning Temporal Data Leakage Language Model Auditing

May 12, 2026cs.LG

Targeted Tests for LLM Reasoning: An Audit-Constrained Protocol

Fixed reasoning benchmarks evaluate canonical prompts, but semantically valid changes in presentation can still change model behavior. Studies of prompt variation can reveal such failures, but without audit they can mix genuine model errors with invalid perturbations, extraction artifacts, and unmatched search procedures. We propose an audit-constrained protocol for targeted reasoning evaluation. Prompt variants are generated from a finite component grammar, rendered deterministically, evaluated under a fixed query budget, and counted as model errors only after semantic and extraction audit. Within this protocol we instantiate Component-Adaptive Prompt Sampling (CAPS), a score-based sampler over prompt components, and compare it with equal-budget uniform component sampling under the same task bank, renderer, model interface, decoding settings, and audit procedure. Across three audited slices, the protocol identifies confirmed model-error prompt keys while excluding formatting and extraction artifacts, but matched comparisons do not show that CAPS improves audited yield or unique prompt-key discovery over uniform sampling. The contribution is methodological: targeted prompt variation can be studied under a reconstructable, reviewable, budget-matched protocol, and proxy-guided policies should be judged by audited yield rather than raw mismatch counts or selected examples alone.

Hongmin Li

School of Life Science and Technology, Institute of Science Tokyo · Department of Computational Biology and Medical Sciences, Graduate School of Frontier Sciences, The University of Tokyo

Prompt Sensitivity Language Model Robustness LLM Reasoning Language Model Auditing