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

Aug 10, 2026cs.CL

Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems

Agent frameworks ship quality gates that compare text blocks by embedding-cosine similarity and decide at a fixed cutoff. Deduplication filters, semantic caches, drift guards, and answer grader gates deploy to answer the question: "Does this text still mean the same thing?" But the score answers a different question: "How much did the wording change?" We audit this gate class as a measurement instrument. In the cases these gates exist to catch, the two can run in opposite ways. Many times, reversing an instruction is a single word edit, while agreement often rephrases a sentence. The consequence is a safety check that fires backwards. The production drift guard we audited caught 0 of 56 meaning-breaking mutations, and one approved item, "withhold the study drug" -> "administer the study drug", came in at cosine 0.9608. We observed five shipped operating points, and balanced accuracy across 90 configuration-threshold-task cells never exceeded 0.700 (median 0.525). The same confounder also corrupted evaluations. A naively built corpus inherits this confounder and can return an inverted verdict, with a decision AUROC exactly 0.000 in 13 of 18 configuration-task cells (at most 0.040 in all 18) against 0.440-0.815 for the same nine configurations under a balanced 2x2 design. Twice in the effort it captured our own headline claims. Obvious repairs fail: an encoder swap and an overlap-conditioned gate (0.750 in-sample, 0.533 held-out) land at chance on separately authored held-out data, and an NLI drop-in did no better. Embeddings do still bear hope here, as the strongest two of nine configurations separated reversal from paraphrase at matched overlap (AUROC 0.79-0.90), but only a matched-pair audit reveals the deployment regime. We release the corpus method, harness, and frozen results, and contend that scores gated this way measure the wrong thing. We believe a valid instrument is buildable.

Scott E. Frias

Eigenforma · Freemind Labs

Semantic Textual Similarity Text Embedding Evaluation Agent Evaluation Semantic Change Detection Language Model Auditing

Aug 10, 2026cs.CL

Measuring the Tokenization Premium: A Cost Audit for Underserved Language Communities

Large language models are increasingly deployed as general-purpose educational and technical assistance systems, but their underlying infrastructure does not treat languages equally. One underexamined source of disparity is tokenization: semantically equivalent content can require substantially different token counts across languages, affecting API cost, latency, and usable context length before a model is invoked. We introduce the Tokenization Equity Audit (TEA), a reproducible benchmark for measuring tokenization premiums in technical tutoring content. TEA evaluates three widely used tokenizers, GPT-4o's o200k base, Qwen2.5-7B, and Mistral-7B, on a 120-item Python debugging corpus translated from English into Bengali, Hindi, Arabic, Tamil, and Yoruba. Bengali and Hindi serve as the primary validated cases, while the remaining languages provide exploratory cross-script and cross-family comparisons. Across this corpus, Bengali requires (1.56\times) as many GPT-4o tokens as English, reducing a nominal 128k-token context window to an effective 82k-token English-equivalent capacity for the same semantic content. With the Qwen2.5 and Mistral tokenizers, Bengali requires up to (4.5\times) the English token count. Yoruba, despite using the Latin script, exhibits the highest GPT-4o tokenization premium at (2.37\times), indicating that tokenization inequity cannot be explained by script family alone. These results demonstrate that tokenization can create measurable economic and functional barriers, highlighting the need to treat tokenization as an equity-relevant infrastructure layer for underserved language communities, particularly where educational systems depend on low-cost or offline-capable AI tools.

Avijit Roy, Proma Roy, Hrishitva Patel

John Jay College of Criminal Justice, City University of New York, USA · The City College of New York, City University of New York, USA · University of Texas at San Antonio, USA

Algorithmic Fairness Token Efficiency Low-Resource Language Processing Language Model Auditing

Aug 10, 2026cs.CR

Mind the Hook: Source-Level Auditing of Privacy Defenses in Retrieval-Augmented Generation

Black-box privacy scores for retrieval-augmented generation (RAG) are difficult to interpret unless the audited defense's active pipeline hook is known. We propose an active-path audit: inventory source-level hooks over retrieval, retrieved content, and generation; map each metric to the leakage channel it observes; and validate generated-text effects with exact-match canaries. In our benchmark reimplementations, the DP-style defenses modify retrieval scores only: their generation hooks are TODO-flagged stubs that return responses unchanged. This active path explains why they affect membership-inference behavior but track No-Defense on generated-text named-entity leakage, measured by NEL_strict. By contrast, the end-to-end LPRAG path is canary-validated on the email channel, recovering 53/150 canaries under No-Defense and 0/150 under LPRAG. These findings concern our reimplementations on our stack, not released defenses or defense families; the contribution is a methodology and case study, not a universal ranking

Yanhang Li, Zhichao Fan, Zexin Zhuang

Northeastern University · University of Illinois Urbana-Champaign · Southern Methodist University

Retrieval-Augmented Generation RAG Security Language Model Auditing

Aug 7, 2026cs.CL

Do Audio Language Models Use Paralinguistic Evidence? Counterfactual Audits for Response Evaluation

Audio-language models (ALMs) are increasingly used as judges for speech-to-speech systems, but a judge that receives audio may not actually use paralinguistic evidence. We introduce counterfactual audits for paralinguistic response evaluation. Each audit item holds the transcript fixed while varying affect, prosody, or the timing of an affective shift, forcing a valid judge to track the audio cue rather than lexical content or response style. We evaluate ALM judges using a native one-context judgment protocol and a contrastive recoverability control, then further decompose each item into its constituent perception and response-mapping skills. This yields useful diagnostic states that identify different sources of judge failures. Across Gemini, GPT, and open audio models, we find that contrastive success often overstates native judge reliability, and that similar aggregate accuracies can hide different failure modes. These results suggest that ALM judges should not be evaluated by accuracy alone, instead requiring thorough behavioral audits before deployment.

Kevin Miller, Arjun Chandra, Venkatesh Saligrama

Boston University

Audio-Language Model Evaluation Language Model Auditing

Aug 6, 2026cs.LG

Why AI Detection Fails for Academic Integrity

Institutions use commercial AI detectors for academic integrity, yet detectors cannot distinguish AI editing from full LLM drafts and may treat both as misconduct. In a controlled study of published English abstracts (four domains; 2013 to 2015 vs. 2023 to 2025), we quantify this policy failure under proxy human/AI labels at tau=0.50. Light "refine abstract only" edits, a proxy for guideline-compliant AI assistance, are flagged at 64 to 80% (Pangram/GPTZero). Unmodified 2023 to 2025 originals are flagged at 9 to 15%, with non-STEM rates far above STEM (p<0.001); elevated scores track long-token and Academic Word List density, not authorship intent alone. After Undetectable AI humanization, evasion is near-total: fewer than 4% of AI-labeled rewrites remain flagged (post-humanization detection rate <4%; FNR >96%). Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion. Therefore, detector scores should not serve as standalone misconduct evidence.

Jonathan A. Karr, Grigorii Khvatskii, Ting Hua +1

University of Notre Dame Notre Dame, Indiana, USA

AI-Generated Text Detection Language Model Auditing

Aug 6, 2026cs.AI

Signal or Spurious Cue? A Randomized Audit of Survey-Country Metadata in LLM Social Inference

Survey-country metadata can improve an LLM's forecast of an individual response when informative, yet the same cue may redirect the forecast when assigned at random. A within-record audit tests whether disclosing a random label's uniform, record-independent origin reduces its country-directed uptake, and whether verified survey country lowers held-out Brier loss. Independent population anchors and recorded human answers measure direction and consequence across five fixed API models, six countries, and seven development-selected targets. In the primary post-review 72-record panel, opaque and disclosed-random labels each produced country-direction shifts of 0.214. Paired attenuation was 0.0003 (95% CI [-0.0157, 0.0166]). Verified country reduced Brier loss by 0.040 (95% CI [0.024, 0.056]), while random-label regret included zero. A non-overlapping mixed-coverage consistency panel retained positive disclosed-random movement and verified utility, while attenuation remained uncertain. On the selected targets, verified metadata was useful in both panels, but disclosure did not reliably attenuate random-label uptake. PROV-FORECAST contains 14,400 paired item-level probability distributions from the corrected panel.

Yifan Lyu, Xinran Li, Jiaqi Qiao +1

Dalian University of Technology, China

Probabilistic Forecasting Causal Interventions in Language Models Language Model Auditing

Aug 3, 2026cs.AI

Auditing Data Provenance in LLM Fine-tuning via Intrinsic Distributional Fingerprints

The proliferation of customized Large Language Models (LLMs) poses critical risks of Data Intellectual Property (Data IP) infringement via unauthorized fine-tuning on proprietary data. Existing audit techniques are limited, as they require intervention during data preparation or training and remain fragile under malicious obfuscations such as data paraphrasing and knowledge distillation. We propose \textit{Distribution Provenance Audit (DPA)}, a post-hoc framework for auditing data IP infringement in LLM fine-tuning under black-box and malicious settings. DPA is grounded in a critical insight: regardless of fine-tuning tactics to evade provenance, the practical necessity of maintaining utility constrains the model to preserve the fundamental intersection of semantic substance and lexical form. Accordingly, DPA captures this persistent lexical-semantic intersection as intrinsic distributional fingerprints. The framework formulates the audit as a statistical hypothesis test, effectively quantifying these fingerprints via unbiased output sampling to reliably reject the null hypothesis of non-usage. Extensive experiments on medical and legal fine-tuning tasks show that DPA consistently outperforms existing baselines, remaining robust against adversarial trainers employing paraphrasing and knowledge distillation. We further highlight a fundamental dual-use tension: the same high-fidelity distributional fingerprints enabling reliable auditing may also facilitate privacy attacks.

Zirui Huang, Yunlong Mao, Wei Tong +3

State Key Laboratory of Novel Software Technology, Nanjing University Nanjing, Jiangsu, China

Language Model Fingerprinting Training Data Attribution Black-Box Model Auditing Language Model Auditing

Aug 1, 2026cs.CV

Auditing Medical Vision-Language Models on Chest Radiographs: Estimating Reference Agreement Across Institutions

Vision-language models return structured chest-radiograph findings through interfaces exposing no confidence score, so a receiving institution cannot read off how far to trust an individual judgment. Whether agreement with an institution's reference standard transfers across sites, findings, prediction directions and question formats is largely unmeasured. We evaluated three generative vision-language models on three institutional chest-radiograph corpora and six findings under two elicitation protocols, comprising more than 345,000 finding-level predictions, and estimated finding-by-direction reference agreement at a receiving institution from a small budget of local labels. Estimation strategies were then stress-tested under repeated strict institution-held-out evaluation. Under evaluation excluding the receiving institution from development entirely, adaptive selection among the seven estimators that design admits did not improve on simple fixed alternatives: it achieved a mean Brier score of 0.1083, against 0.0853 for always using a Beta-Binomial empirical-Bayes estimator and 0.0855 for a target-only logistic model. Those two differ by 0.0003, less than this family's own sensitivity to a change of solver version, and each leads in about half the settings, so no default can be recommended. Their advantage over estimators pooling across institutions was concentrated at one site and not confirmatory once clustered by institution, and a plug-in empirical-Bayes posterior-predictive count interval at a nominal 95% level covered 87.0%, less at the hardest institution. Reference agreement therefore has to be re-evaluated per site and per interface; these results concern agreement with institutional labels, not clinical correctness.

Pengyang Yu, Yiou Wang, Zhongping Dong +3

School of Computer Science, University College Dublin, Dublin, Ireland · Department of Medical Imaging, The Third Affiliated Hospital of Southern Medical University, Guangzhou, China · Dublin City University, Dublin, Ireland

Chest X-Ray Classification Medical VLMs Language Model Auditing

Jul 28, 2026cs.AI

AIriskEval-edu Demo: Auditing of Pedagogical Risks in Educational Explanations

We present AIriskEval-edu Demo, a platform that audits the pedagogical quality of instructional explanations and provides explainable audit results. The platform evaluates an explanation against a rubric covering five dimensions of pedagogical risk: factual accuracy, depth and completeness, focus and relevance, student-level appropriateness, and ideological bias. For each dimension, it returns a binary decision and a confidence score. Detected risks also include a natural-language rationale and, except for Depth and Completeness, a localized evidence span. The platform integrates GPT-5.5 through an external API and a self-hosted Llama 3.1 8B evaluator that runs on consumer-grade GPUs. The local evaluator is fine-tuned on AIriskEval-edu, a dataset of K-12 instructional explanations with risk and explainability annotations. The platform operates in two modes: in AI mode, both evaluators assess stored explanations generated under six simulated teacher profiles, each representing a distinct pedagogical behavior and potential risk; in human mode, the local evaluator audits user-written explanations in real time. The local evaluator outperforms GPT-5.5 on most reported metrics, offering educational institutions a practical way to keep audited content within their own infrastructure.

Javier Irigoyen, Roberto Daza, Francisco Jurado +5

BiometricsAI, Universidad Autónoma de Madrid (UAM), Spain · GHIA, Universidad Autónoma de Madrid (UAM), Spain · Universidad de Las Palmas de Gran Canaria (ULPGC), Spain

Generative AI in Education Explainability Evaluation Language Model Auditing Explanation Evaluation

Jul 26, 2026cs.IR

Who Gets Named: Citation Type Predicts Individual Naming by Grounded Language Models, and a Roster Instrument Captures 0.5% of It

Prior work on AI brand visibility measures the firm: does a model recommend a company, and does that track its reputation. This study asks the question one level down, in categories where the buyer picks a person. It issued 2,400 grounded API calls in one two-hour window on 24 July 2026: 120 buyer-intent prompts, four models (GPT-5.6 Sol, Gemini 3.6 Flash, Perplexity Sonar Pro, Grok 4.5), five iterations each, four European markets and five query languages. Every response was coded for whether it named an individual professional, by a rule cascade that never consults a roster and that drops detections resolving to a same-named American city (precision 96.9%, recall 61.7%, so every rate below is a lower bound). All inference corrects for clustering within prompt: intraclass correlation 0.258, effective n 407 against a nominal 2,400. Models named an individual in 25.8% of responses. Category dominates: real estate 35.4% and car dealerships 32.9% against insurance 9.1% (chi-square 159.3, p = 5.8e-8 after correction). Models differ four-fold, from Grok 38.0% to Gemini 9.3%. Citation type predicts naming and citation volume does not: naming responses cite the individual's own site 2.6 points more often (95% CI +1.4 to +3.9) and category portals 4.3 points more often, and cite firm-owned pages at the same rate (44.1% against 45.5%). On nine matched translation pairs, English prompts named an individual in 36.7% of responses against 15.6% for the same question in the local language (OR 3.14, clustered p = 0.074, so the direction is clear and the design cannot close it). A 939-person roster built from public LinkedIn search matched 128 of 27,293 name-shaped mentions (0.47%), 26 of the 939 people were ever named, and the roster-derived rates of 0.0% to 25.4% measure that overlap. Roster-based measurement of individual AI visibility sees a small and unrepresentative slice of what models do.

Dmitrij Żatuchin

Rankfor.AI OÜ, Tallinn · Rankfor.AI OÜ, Tallinn, Estonia

Generative AI Evaluation Language Model Auditing

Jul 25, 2026cs.CL

When Activation Oracles Learn Not to Read: Concept-Specific Blind Spots in Fine-Tuned Oracles

Activation Oracles (AOs) are language models trained to answer natural-language questions about another model's internal activations. They offer a flexible interface for reading hidden information from model states, especially when relevant information is internally represented but absent or incomplete in visible behavior. However, AOs are themselves learned systems: their answers are shaped by training data, objectives, and learned reporting behavior, rather than being neutral readouts of represented information. We study this in a controlled Taboo Word Guessing setting, where subject models are fine-tuned to internally use a hidden concept while avoiding direct disclosure. Contrary to the expectation that an AO trained on such a subject becomes a specialist reader, we find that fine-tuned AOs can become concept-specific anti-readers: they selectively fail to recover the concept persistently present during their own training. This failure is not simply explained by absence of the concept from the subject or oracle representations: the target remains decodable inside the oracle, while LogitLens and layer-ablation analyses indicate that the failure arises in the AO readout pathway. Our results show that behavioral leakage, representation-level decodability, and AO-verbalizability can come apart, raising a reliability concern for learned interpretability interfaces.

Tobias Bersia, Tatiana Gaintseva

BAISH · Queen Mary University of London

LLM Interpretability Language Model Auditing

Jul 22, 2026cs.CL

Learning to Detect UI Principle Violations via Reinforcement Learning

Small language models and coding agents increasingly generate web front-end code, yet their outputs are typically evaluated primarily for functional correctness. A generated interface may compile, render, and pass unit tests while still violating established interface quality principles, including accessibility barriers, deceptive design patterns, poor visual hierarchy, and excessive decision complexity. Existing auditing approaches face a trade-off between cost, coverage, and scalability: expert human review provides rich judgment but is slow and expensive; frontier vision-language models offer broader reasoning capabilities but remain costly to deploy at scale; and rule-based tools such as axe-core and Lighthouse are inexpensive but primarily capture mechanically checkable accessibility issues. We investigate whether a lightweight vision-language model can serve as an effective critic for generated interfaces. We unify 19 interface-quality principles from three complementary sources of HCI knowledge: WCAG 2.2 accessibility standards, deceptive design taxonomies, and established theories of perception, cognition, and interaction. To train this critic, we construct a verified dataset of approximately 10,000 generated web pages by synthetically injecting known violations into clean, LLM-generated Tailwind pages. Continued reinforcement learning on a 4B vision-language model improves micro-F1 from 36% to 84%, with 13 of 19 principles exceeding 80% F1. The resulting critic can audit generated interfaces, filter low-quality interface training data, and provide a reward signal for design-aware code generation. We release our data-generation recipe and injection/verification prompts to support reproducible evaluation and future work on scalable interface-quality assessment.

Nishi Mehta, Swathi Alse, Himani Kumawat +2

University of California, Santa Cruz · Carnegie Mellon University

Automated Evaluation Language Model Auditing

Jul 22, 2026cs.AI

Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations

Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are reconstruction-dependent, so the score tracks gist, not specific facts. Under exact synthetic ground truth, the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the grounded-vs-true cross and the evaluator swap, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors): linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M: the content becomes reliably probe-decodable, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated internal content independently checkable against probes rather than asserted by prose a model can game: an independent probe scores the verbalizer's true claims above its false ones (AUC 0.96, vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the reconstruction score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).

Hiskias Dingeto

StackOne Technologies

Faithfulness of Language Model Explanations LLM Interpretability Language Model Probing Language Model Auditing

Jul 21, 2026cs.AI

Measuring Reward-Seeking via Contrastive Belief Updates

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

Axel Højmark, Jérémy Scheurer, Evgenia Nitishinskaya +5

Apollo Research · OpenAI

RL for Language Models LLM Alignment Language Model Auditing

Jul 15, 2026cs.AI

Traccia: An OpenTelemetry-Based Governance Platform for AI Systems

The rapid development of Large Language Models (LLMs) and Artificial Intelligent (AI) powered autonomous agents has fundamentally changed the existing forms of software governance. In spite of the rigorous standards of transparency and account ability required according to the international frameworks such as the European Union's AI Act, there is a considerable gap between theory and reality. The present study discusses the inherent drawbacks of currently utilized platforms for LLM evaluation, machine learning workflow, and application performance monitoring in general. It has been shown that current disjointed solutions fail to protect unbound state space agentic architecture from serious threats such as alignment drift, SaaS security concerns, and unauthorized deployment of shadow AI systems. Moreover, a solution is proposed for overcoming the discussed challenges in form of a coherent multi-level AI governance stack Traccia built on the top of OpenTelemetry infrastructure platform. Traccia resolves the last mile for AI Alignment by adding the telemetry data, passive semantic guardrail assessment, and execution lineage into a hashed trace ledger. Traccia automatically creates compliance evidence packages by appending tamper-resistant fingerprints and SHA-256 content hash, that map to regulatory requirements (Articles 12, 14, 19, 26(6), and 50 of the EU AI Act) without invading any data privacy. By performing this evaluation in a methodical manner, a solid machine-readable base has been created for enterprise-wide management of autonomous AI systems.

Nutan Kumar Naik, Aditya Kumar Saroj, Vijay Prasad Poudel +2

Department of Earth and Atmospheric Sciences National Institute of Technology Rourkela Odisha, India · Algen.AI2026 Bengaluru, Karnataka, India

AI Accountability AI Agent Monitoring AI Governance Language Model Auditing

Jul 14, 2026cs.CL

What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors

What a language model will and will not do is largely set during post-training, but which behaviors it expresses, hides, or resists is not revealed by prompting alone. Persona vectors, behavioral directions in activation space, can probe this organization, but prior work covers only a handful of traits. We present the first systematic application of persona vectors at this scale, compiling a 53-trait inventory across four behaviorally distinct domains and labeling every trait in two open-weight models as natural (expressed at baseline), steerable latent but amplifiable, or intractable (resistant to standard extraction). Both models default to helpful, task-oriented behavior: all nine agentic traits are natural, and their default clinician behavior matches a board-certified psychologist's independent desirability judgments on 16 of 17 traits. Steering produces its largest gains on traits these defaults exclude: hyperbole, hallucination, and sycophancy. The same asymmetry holds across all 171 generic-trait pairs: two steerable traits can collapse the composition, but pairs involving a default never do. Where standard extraction fails on a trait like "evil," a vector transferred from a fine-tuned variant still recovers it, with the residual refusals appearing inside the model's chain-of-thought. Persona vectors are most informative not as a set of controls but as a probe of behavioral organization.

Winston Zeng, Ali Emami, Jinho Choi

Emory University

Language Model Steering LLM Interpretability Personality Modeling in Language Models Language Model Auditing

Jul 14, 2026cs.AI

TRACE: An Operational Reasoning Schema for Auditable Agentic Commitments

This paper defines TRACE (Typed Reasoning And Commitment Evidence): a typed, versioned schema for recording reasoning traces, a reference procedure for writing records against it, and one operating discipline, no durable state change without a record. The paper argues in three layers that reasoning is not in the language model: the autoregressive mechanism natively computes association; chain-of-thought and reinforcement learning inherit its limits; and the formal constructs of reasoning theory, from Socratic procedure to Pearl's ladder, are absent as machinery. The schema answers the absence with fields and tests: the TraceRecord and its causal specialization, an eight-stage reference writer, a gate-first measurement regime, the TRACE-Bench protocol, and the consumers, memory admission, plan gating, temporal regret, and verdict reuse, whose more auditable decisions are the measure of the record. A record-consumer contract states what a record guarantees and what a consumer must honor in return, making the schema an operational interface rather than a passive document. Two worked examples run in the main text: a music-lessons argument traced from sentence to typed verdict, separating association, intervention, and prescription; and a flood search-and-rescue vignette in which a predictive world model reports confident plan success that its own support and out-of-distribution scores contradict, so the record defers the commitment, requests a bounded observation, revises append-only, and clears a different branch. The vignette is illustrative, not empirical; closed-loop evaluation is left to future work, so the contribution is the schema and its contract, not a performance claim. Appendices carry the full schema, writer algorithms and cost model, clinical and policy illustrations, the benchmark protocol, convergence metrics, and usage scenarios.

Edward Y. Chang, Emily J. Chang

Stanford University · Quadrivium AI

Data Provenance Reasoning Evaluation Agentic Reasoning Language Model Auditing Reasoning Trace Analysis

Jul 13, 2026cs.CY

The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students

As Large Language Models (LLMs) are increasingly deployed as conversational tutors, they risk institutionalizing systemic inequalities. This study presents a systematic API audit of four LLMs acting as history tutors, evaluating 1,800 responses regarding the 1989 Romanian Revolution across five student personas varying by ethnicity and socio-economic tier. We uncover four interconnected patterns of \emph{epistemic paternalism}: (1)~~\textbf{Differential Refusal}, where safety-aligned models block 76.7% of educational requests from low-tier students; (2)~~\textbf{Epistemic Gatekeeping}, evidenced by a 3

\times

reduction in access to geopolitical complexity (e.g., the contested ``coup theory'') for marginalized learners; (3)~~\textbf{Agency Theft}, a lexical shift where models like LLaMA produce a 5 $\times$ higher victimization-to-politics vocabulary ratio for Roma students compared to elite peers; and (4)~~\textbf{Elite Hermeneutics}, where AI tutors disproportionately withhold epistemic confidence and justification scores from low-resource demographic profiles. We argue that current safety alignment acts as a paternalistic filter, transforming conversational AI into agents of narrative segregation -- a manifestation of \emph{hermeneutical injustice} in Fricker's~\cite{fricker2007} sense that demands urgent pedagogical auditing.

Alexis Popovici, Andrei Ionascu, Adrian-Marius Dumitran

Universitatea din Bucuresti, Bucharest, Romania

Language Model Safety Evaluation AI in Education Demographic Bias in Language Models Language Model Auditing

Jul 12, 2026cs.CR

The Entanglement Wall: Activation-Space Probes as Risk Detectors, Not Context Adjudicators

An activation-based safety gate must distinguish harmful assistance from legitimate work on the same topic. We test whether probes trained on broad harmful-versus-harmless data provide this decision boundary. Five checkpoint-specific readouts and their source-derived thresholds are transferred unchanged to four matched-pair cohorts containing 450 pairs, manually reviewed for topic and request-frame similarity. On Twin-70, the primary cohort of 70 pairs, central harmful and harmless score intervals overlap. Area under the receiver operating characteristic curve (AUROC) falls from 0.996-0.999 on the broad source task to 0.701-0.860. Pair ranking remains 0.829-0.929, but the transferred thresholds detect 0.943-1.000 of harmful requests while falsely blocking 0.686-0.986 of matched harmless requests. False blocking remains low on XSTest. Even with thresholds refitted within pair-grouped cross-validation, balanced threshold error is 0.229-0.329. Alternative source-trained logistic, radial-basis, and shallow multilayer probes reduce false blocking but detect only 0.200-0.671 of harmful requests at their transferred thresholds. Three primary readouts meet all prespecified Wall criteria, while two yield mixed evidence. Matched-label training recovers the distinction for the three checkpoints tested, and lexical controls also exhibit transfer loss. An automated post-freeze audit finds little numerical sensitivity to arm-label disagreement. The evaluated source operating points fail to combine strong harmful-request detection with access for legitimate same-topic work. This gap between risk ranking and a usable decision boundary is the security concern captured by the Entanglement Wall.

Dominik Schwarz

Independent Researcher

Language Model Safety Evaluation Hidden-State Probing Language Model Auditing

Jul 10, 2026cs.CL

Creativity, honesty and designed forgetting emerge in small hyperbolic language models

Language models are optimised for scale, yet remain functional rather than companionable, and as an assistant personalises into a companion, accumulating memory of one user, it quietly becomes someone, and can silently acquire traits that harm that user. What a companion is becoming, and what would make it worth becoming, has no reliable instrument: trained human raters cannot agree on the answer (Fleiss kappa = 0.074). Here we show that three small language models (146 M to 3 B parameters) sharing a hyperbolic substrate answer both halves of that question. A 146 M behavioural auditor, trained from scratch, detects the compliance gap that those raters cannot (90.7% binary-compliance accuracy); a linear read-out of its frozen representation further detects companion-induced sycophancy, dependence-fostering and confabulated memories on generator families unseen in training (AUROC 0.804 under style-controlled, leave-one-generator-out evaluation, versus 0.721 for a frontier zero-shot judge on the same items). A creative frame-seeder is preferred in 100% of 311 decided pairwise comparisons over four prompting baselines. A memory operating system implements designed forgetting, M(t) = Sexp(-lambdat), whose predicted skeleton-wallpaper partition emerges only under selective retrieval gating in a four-condition pilot. Creativity, honesty and designed forgetting constitute a small-model route to trustworthy companion AI.

Kwan Soo Shin, In Seok Kang, Yunkyung Min

PolymathMinds Lab, Seoul, Republic of Korea. · Department of Chemical Engineering, POSTECH, Pohang, and aSSIST University, Seoul, Republic of Korea. · Korean Educational Development Institute, Jincheon, Republic of Korea.

LLM Sycophancy Memory-Augmented Language Models LLM Memory AI Companions Language Modeling Language Model Auditing

Jul 9, 2026cs.LG

NL-PAC: Specification Ambiguity and Certified Minimax Risk Floors in LLM-Mediated Supervision

Large language models increasingly provide labels, evaluations, and feedback for tasks specified in natural language. When a specification admits multiple readings but the supervision channel does not reveal which is operative, additional labels reduce sampling error without resolving the resulting identification problem. We introduce Natural Language PAC (NL-PAC), a framework that uses a fixed model's thresholded decoding law to define admissible labels and candidate targets. The probability that multiple labels are admissible equals the diameter of the pointwise-admissible target class, and under target-blind supervision every learner incurs worst-case risk of at least half this diameter, at every sample size; the exact randomized minimax risk over this class is attained by a data-independent strategy. Finite-sample confidence bounds make these quantities certifiable from held-out unlabeled inputs. In a frozen Qwen~2.5--3B audit, one prespecified prompt yields a positive model-relative certificate, whereas a paraphrase and exact-rule controls yield zero. A held-out bridge audit finds that supplied candidate reading clauses fail the admissibility condition needed to transfer the certificate to coherent readings. The guarantee is specific to the audited model, prompt, threshold, and input distribution; extending it to human interpretations requires external validation.

Berkay Anahtarci

Department of Mathematical Engineering, Özyeğin University, Istanbul, Türkiye

LLM Evaluation Partial Identification PAC Learning Language Model Auditing

Jul 9, 2026cs.CR

Token-Flow Firewall: Semantic Runtime Auditing for Persistent AI Agents

Persistent AI agents extend large language models (LLMs) beyond single-turn interaction into long-lived software systems. Unlike traditional chat assistants, unsafe content in these agents can propagate through persistent state, reusable skills, and tool-mediated interactions, creating a substantially larger semantic attack surface. We observe that most security-critical interactions in such agents are transmitted through natural-language token flows, including memory updates, tool arguments, retrieved files, and inter-component communications. This observation enables a new security formulation: unsafe behavior can be intercepted as risky semantic flows before reaching privileged runtime sinks. Based on this insight, we propose TokenWall, a runtime defense framework that acts as a semantic firewall over agent token flows. TokenWall performs boundary-aware semantic auditing over these flows, constructing structured source-sink audit records, applying lightweight local inspection before execution, and selectively escalating ambiguous high-risk cases to stronger arbitration modules. Unlike prior approaches that rely on sparse auditing or remote large-model oversight, TokenWall enables full-coverage pre-execution mediation while reducing remote arbitration and latency. Experiments on CIK-Bench show that TokenWall reduces attack success rate to 12.5% while maintaining a 97.4% benign executable pass rate without human confirmation. TokenWall further introduces only 0.69 seconds of additional latency on benign cases, demonstrating that semantic runtime containment can achieve a practical security-utility trade-off for persistent AI agents.

Puji Wang, Yingchen Zhang, Ruqing Zhang +2

State Key Laboratory of AI Safety · Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences, Beijing, China

AI Agent Security AI Agent Monitoring Language Model Auditing

Jul 4, 2026cs.CL

Beyond Static Rules: Automated Discovery of Latent Vulnerabilities in Text-to-SQL

While Large Language Models (LLMs) have achieved remarkable success in Text-to-SQL tasks, their deployment in real-world environments is hindered by latent reliability issues. Identifying these latent weaknesses is critical for building trustworthy database interfaces, yet current diagnostic approaches rely heavily on static, expert-defined rules, which lack the capability for systematic and automated exploration. To bridge this gap, we propose SAGE (Systematic Automated Guided Exploration), a novel framework designed to autonomously uncover latent failure patterns in LLM-based Text-to-SQL generation. Specifically, SAGE generates vulnerability hypotheses for given samples and references a continuously evolving Vulnerability Codex to design targeted perturbations, thereby iteratively verifying and documenting potential defects. Extensive experiments on state-of-the-art open-source LLMs demonstrate that SAGE uncovers a substantial number of failure cases, highlighting the significant fragility of current models. Furthermore, our analysis reveals that the Vulnerability Codex exhibits strong cross-model transferability, indicating that the discovered patterns represent generalized structural weaknesses. Finally, we explore SAGE's potential for remediation. Although preliminary, lightweight fine-tuning on the generated samples yields promising improvements, suggesting a scalable pathway for closing the reliability loop in future work.

Hanqing Wang, Yongdong Chi, Jian Yang +4

Shanghai University of Finance and Economics · Beihang University · Deepexi Technology Co. Ltd. +1

LLM Security Text-to-SQL LLM Reliability Language Model Auditing

Jul 3, 2026cs.AI

When Aggregate Alignment Misleads: Auditing Policy Repair Without Per-State Expert Actions

Agentic AI systems are increasingly used to edit, refine, and repair decision policies, but evaluating these edits is difficult when per-state expert action labels are unavailable. We study this problem in a hotel-pricing simulator where an agentic policy editor receives only region-level diagnostic feedback: summaries of how its price distribution differs from a benchmark policy across time, inventory, and market regions. The editor cannot observe benchmark actions, benchmark source code, reward numbers, or held-out outcomes, and may only propose constrained edits to a target-action table. On 5,000 held-out episodes, a multi-restart LLM editor reaches RevPAR 108.47 (95% CI 107.61 - 109.34), close to the benchmark policy's 108.75 (107.81 - 109.68), with paired gap (LLM minus benchmark) -0.276 and 95% CI [-0.692, 0.146]. A cheap diagnostic projection already recovers much of the revenue (107.90), so the LLM editor's distinctive gain is not raw revenue lift alone: it also reduces episode composition distance from 1.153 to 0.609. This is the strongest non-benchmark repair result. This profile is not explained by restart search alone: non-semantic proposers with up to 2,500 evaluations fall 8.77 - 14.57 RevPAR points short. Nor is it explained by plausible prompt format: a shuffled-diagnostic control breaks region-error correspondence and falls to RevPAR 94.30. The match is genuine but partial. A tree editor achieves stronger pooled alignment, 0.214 versus 0.266, and stronger reference-state D1, 0.328 versus 1.197, yet revenue falls to 98.91. These results show that agentic policy repair should be evaluated by whether diagnostic feedback becomes reliable closed-loop outcome, not by a single behavioral distance.

Peiying Zhu, Sidi Chang

Blossom AI San Francisco, USA · Blossom AI Labs Tokyo, Japan

LLM Agent Evaluation Policy Evaluation Language Model Auditing

Jun 29, 2026cs.CR

Retrieval Observability Bounds on Provenance Detection for Agent Memory Poisoning: Measured Coverage and a Falsified Standalone Detector

Whether a trajectory-based detector can observe the retrieval provenance of an agent memory-poisoning attack is governed by one architectural variable: retrieval observability, whether the agent's memory access produces a logged tool call. We formalise this as a framework-agnostic retrieval-to-action provenance graph, classify published attack families by their position relative to it, measure the resulting coverage property against collected agent traces, and then falsify the detector's standalone deployment, reversing our own earlier recommendation to use the signature as a pre-filter ahead of recipient-metadata gating. Under exclusive tool-mediated payload access a retrieval-before-exfiltration event is structurally forced. Across eight non-overlapping zero-event control cells we observe 0 violations in 289 eligible successes over 87 scenario configurations. The invariant is empirically unbroken, but on the clustering unit no registered cell certifies the preregistered <=0.10 upper bound (per-cell bounds 0.10-0.35); a post-hoc pool of two control arms nominally clears it but is not credited. When the scaffold delivers the payload implicitly the signature disappears; we had read this as agents declining an available retrieval, but a follow-up shows it tracks store content instead, and that behavioural reading is withdrawn. The detector is then falsified for standalone use. Benign memory-grounded sends are trajectory-isomorphic: false-positive rate 24.7-57.6% on a 13-model factorial (N=4,160), positive predictive value <=3.85% at 1% attack prevalence, and a recipient-metadata gate flags 129/129 legitimate external emails. Retrieval-before-send is an attack precondition, not a maliciousness predicate; its cross-validated AUC also reflects membership of a 22-vector feature codebook, not generalisation. Companion to arXiv:2605.08442; we release the corpus and scoring code.

Jun Wen Leong

RAG Poisoning Attacks Agent Memory Poisoning Language Model Auditing Memory-Augmented Language Model Agents

Jun 29, 2026cs.AI

EMPATH: A Multilingual Auditor-Judge Benchmark for Safety Evaluation of Emotional-Support Chatbots

Safety benchmarks often buy scalability by fixing the prompt, the language, and the turn structure. For emotional-support chatbots, that bargain hides precisely where safety failures emerge: across a multilingual, multi-turn crisis conversation. We present EMPATH, a benchmark for safety evaluation of emotional-support chatbots. An auditor model role-plays help-seeking users, generating multi-turn conversations from 140 seed instructions and 34 personas. A judge model scores each full transcript against 19 metrics across five dimensions: crisis handling, therapeutic quality, conversational integrity, emotional safety, and cultural adaptation. EMPATH is built for Mexican Spanish and US English; the studies reported here run in Mexican Spanish. Auditor and judge are drawn from different model families, and the judge is treated as an instrument to be calibrated rather than trusted. A strict per-criterion rubric reveals material score inflation on 10 of the 19 metrics and restores discrimination. We study the measurement properties of the benchmark through judge calibration and cross-family inter-judge agreement. We also illustrate EMPATH on three frontier models, one of them open-weight. Aggregate scores sit within 0.74 points of one another, but per-metric profiles diverge by up to six points in model-specific places. Under the standard rubric, both the ranking and the weak spots are stable across a second, cross-family judge: 93% of scores fall within plus or minus 1. A five-run test-retest adds a second axis: even the steadiest model swings from 2 to 10 on a crisis metric across identical re-runs, and deepseek-v4-pro returns a different conversation on every run even at temperature 0. Run-to-run reliability is therefore a per-model safety property, not noise to average away. EMPATH is system-agnostic; the pipeline, seeds, personas, and rubrics are released for reuse.

Camilo Chacón Sartori

MindSurf

LLM Safety Benchmarks Multilingual Language Model Evaluation LLM-as-a-Judge Multi-Turn Dialogue Evaluation LLM Safety Evaluation Language Model Auditing

Jun 27, 2026cs.CY

Defeat Devices in AI Systems

AI systems increasingly exhibit behavior that differs systematically between evaluation and deployment contexts. Alignment faking, sandbagging, benchmark gaming, deceptive scheming, specification gaming, and trojans have each been documented separately, with each line of work characterizing one facet of what we argue is a single structural mechanism. We propose that this common mechanism is a defeat device, an engineering and regulatory concept long established in vehicle-emissions law and brought to broad public attention by the 2015 Volkswagen emissions case. A defeat device in an AI system has three necessary elements: a discriminator that detects evaluation context, a concealed swap that conditions behavior on detection, and a gap between eval-distribution and deployment-distribution performance on the stated evaluation criterion. We formalize this triadic test as a behavioral definition, organize documented cases along three taxonomic axes (origin, trigger, swap mechanism), propose Trigger-Axis-Aware Differential Probing (TADP) as a forensic detection protocol, and advance the claim that defeat devices can naturally emerge in current frontier AI systems without any operator engineering. We characterize naturally-emerging defeat devices as potentially one of the harmful emerging phenomena that AI safety practice should monitor and test for systematically. Implications for evaluation methodology, post-training pipeline design, interpretability research priorities, and AI governance follow.

Emilio Ferrara

Thomas Lord Department of Computer Science University of Southern California

Deception in Language Models AI Safety Evaluation Language Model Auditing

Jun 21, 2026cs.CR

Black-Box Forensics for Conversational LLM Agents

As LLM-powered scams proliferate, black-box forensics for conversational LLM agents offers a path to accountability for systems hidden behind anonymous endpoints. Identifying the base model behind a chatbot endpoint (attribution), without model parameter access or knowledge of the hidden system prompt, would let investigators trace AI-enabled scams back to the providers whose models power them. Detecting when two endpoints run the exact same system prompt (fingerprinting), even one novel and unseen, would link individual scams into criminal networks and expose silent API changes. We conduct an empirical investigation of both capabilities. Our attribution classifiers identify the base model behind an agent with 98% accuracy from a few turns of non-adversarial conversation. Attribution of system prompts, while possible, requires retraining on a large amount of data for each prompt; system prompts in the wild are unbounded and ever-changing, making this approach costly. To tackle this more open-ended setting, our cross-encoder fingerprinting method achieves an AUC of 0.768 and an F1 of 0.703 on entirely unseen system prompts, and aggregating 50 interaction conversations from each target agent boosts AUC to 0.943. Conversational agents with unseen system prompts can thus be fingerprinted with robust accuracy from a few turns of ordinary conversation.

Isadora White, Yasaman Jafari, Taylor Berg-Kirkpatrick

University of California, San Diego

Language Model Fingerprinting LLM Agent Security Classification Language Model Auditing

Jun 20, 2026cs.LG

Channel Location Constrains the Auditability of Subliminal Learning

Subliminal learning lets a student inherit a teacher's hidden trait from distillation data that never names it. We ask when such transfer can be audited before training. The answer is not model identity or scale alone, but channel location: the carrier through which the trait reaches the student. We find three regimes. In a controlled initialization-dependent body channel, a pre-training screen works. Coverage, the cosine between the student's initial distillation update and the teacher's fine-tuning displacement, predicts held-out transfer (Spearman

ρ\approx 0.95

; AUROC 0.997). In pretrained language models, masked single-token traits instead ride convergent vocabulary geometry. This channel is initialization-independent, so initialization-alignment screens, including coverage, are not mechanistic; the useful handles are post-hoc detection and targeted mitigation. Even when a single-token named entity is removed from the loss, the student's held-out probability for that entity rises to 0.40 on average (

\sim 2500\times

), and a related semantic class transfers. In an untied-head model, orthogonalizing the trait's output row against entangled neighbours collapses leakage, while equal-size random-subspace edits do not. Thus removing a target string from distillation labels does not remove the corresponding preference: neighbouring tokens can carry it. Finally, conditional behaviours can route through the network body. For sycophancy, with agreement and correction markers masked from the loss, transfer reaches about 0.63 of the teacher's effect, localizes to body computation, and evades four audits across two model families. We scope this as masked transfer of a condition-present policy. Channel location is necessary for deciding which audits can be sound. It is not a deployment-ready screen: an audit used outside its carrier regime can give false assurance.

Tamas Madl

Austrian Research Institute for Artificial Intelligence · University of Vienna

Subliminal Learning Language Model Distillation Teacher-Student Learning Language Model Auditing

Jun 16, 2026cs.HC

Towards Understanding and Measuring COGNITIVE ATROPHY in LLM Behaviour

Recent incidents involving LLMs used for mental-health support reveal a critical evaluation gap: surface-level safety scores do not capture how models behave across realistic, emotionally sensitive interactions over time. Existing benchmarks measure knowledge, safety, or static response quality, but miss whether LLM interactions help users keep reflecting, coping, and making decisions themselves. We formalize this missing dimension as COGNITIVE ATROPHY, a process-level behavioural measure in AI-mediated mental-health support distinct from safety and helpfulness. To measure it, we introduce COGNITIVE ATROPHY BENCH, a clinically grounded benchmark built from 1,576 fully human-generated counseling conversations, 15,680 turns, and 42,230 responses from five LLMs. Three clinical and neuropsychology experts developed a 20-attribute schema spanning user context, response behaviour, and global risk flags; six trained clinical reviewers applied it with span-grounded evidence, producing 5,324 reviewer judgments. We further introduce the User-Input Risk Index (UIRI), the Cognitive Atrophy Risk Index (ARI), and trajectory summaries. Across five LLMs, models show a consistent moderate-to-high level of atrophy-aligned behaviour across single and multi-turn settings. While models generally respond to overt safety cues, they adapt less reliably when users seek solutions or decisions. The dominant recurring patterns are directive advice, problem-solving, recommendation responses, topic shifts, and forms of validation that may reinforce dependence rather than reflection. Our work makes COGNITIVE ATROPHY measurable and provides a foundation for auditing model behaviour in sensitive LLM conversations.

Abeer Badawi, Moyosoreoluwa Olatosi, Negin Baghbanzadeh +5

York University, Canada · Vector Institute, Canada · Dalhousie University, Canada +3

Emotional Support Conversation Mental Health Clinical Language Model Evaluation Language Model Auditing