LLM Auditing

We address the use of large language models (LLMs) to help discover Isabelle proofs. An Isabelle build establishes that the submitted theory is accepted, but not that an LLM changed only what the developer authorised. We present CAPRI, a contract-aware repair workflow in which Isabelle checks the proof and an independent checker enforces a machine-readable edit contract. Prompts, proposals, candidate repositories, diagnostics, verdicts, and hashes are retained for audit. We evaluate five workflows on twelve failed proofs from four developments, with three replicates per task and condition, giving 180 runs and 138 valid repairs. Of 144 terminal candidates accepted by Isabelle, six had modified protected text; all arose in iterative workflows that could edit a complete theory. A proof-body-only interface produced 29/36 valid repairs and no contract violations, compared with 31/36 for the corresponding full-theory workflow. One-shot repair produced 22/36, while a later prospectively frozen iterative workflow produced 32/36; these figures compare complete workflows rather than individual mechanisms. A separate post hoc OpenRouter campaign found no improvement in the designated Luna comparisons. A Sol configuration with matched demonstrations produced 33/36 repairs, compared with 29/36 in the frozen OpenAI Responses condition, but the difference was not statistically significant in a one-sided exact McNemar test ($p=0.0625$).

LLM benchmark scores can be precise even when the observation protocol does not identify the behavioral property they are intended to measure. In a controlled, solver-grounded setting, we formalize a protocol-level identifiability audit over a finite behavioral policy class: given policies H, observation support O, and estimand $τ$, we test whether O separates every pair with different $τ$. The audit requires zero model calls and resolves our diagnostic case: base-only observation collapses seven frozen deterministic policies into one equivalence class; full support yields seven classes and no cross-estimand collisions; every leave-one-out support retains a constructive collision witness. Empirically, both constrained-generation variants have pair-validity 1.0, yet base accuracy and selective-response fidelity diverge - 0.620 versus 0.324 across six balanced oracle-transition directions (cluster-bootstrap 95% CI [0.600, 0.642] vs. [0.304, 0.345]) - and the gap recurs on a second deterministic source (0.646 vs. 0.331). The audit also synthesizes a minimum identifying support $O^*$ for the frozen policy class: two cells instead of the full 36-cell tensor. This case shows how evaluation-design validity can be checked structurally before model inference and why base correctness does not determine intervention-response fidelity.

Normative datasets are often used to train and align AI systems, but the norms they contain can function as action-guiding patterns rather than neutral moral knowledge. We propose treating the AI system as a proxy actor and test whether dataset-level norms can shift it away from its baseline safety behavior when it faces high-conflict dilemmas. We make three contributions. First, we demonstrate in controlled experiments that norm-breaking fine-tuning yields norm-divergent actions justified by self-interested rationales, suggesting a systematic shift in patterns of justification. Second, we establish a practical audit trail linking downstream justifications to upstream norms using mixed methods. Third, we show that system prompts can both suppress and elicit these patterns. We conducted experiments on three models (LLaMA-3.2-11B, Qwen-3.5-9B, and Pixtral-12B) using Low-Rank Adaptation (LoRA) fine-tuning on Social Chemistry 101 Fairness/Cheating (norm-following vs. norm-breaking) with prompt steering. Across all three models, we find that norm-breaking fine-tuning shifts the model's default rationale style from safety compliance to instrumental self-interest, whereas system prompts can override this behavior. Our results support a distributed view of alignment in which observed behavior depends jointly on training data, fine-tuning, and prompting, motivating norm-aware documentation and rationale logging for contestable oversight.

When a language model answers an interventional question, the computation it must perform depends on the type of evidence the query requires. We report a decoupling in how a transformer organizes causal knowledge: slot-by-type structure induced by type-level supervision organizes routing, yet remains functionally decoupled from answer readout. We establish this with a typed mechanism library -- discrete mechanism slots partitioned by evidence type, auditable at the state level -- on a causal-world benchmark with exact interventional ground truth, under a frozen protocol, at two scales (22.6M and 125M). Four preregistered findings. (i) Origin. Slot-by-type organization is induced by type-level supervision: absent in architecturally identical unsupervised controls, not buyable by content-free gating labels, and statistically attributable to the supervision signal, replicating at 125M under a powered preregistered protocol (all nine cells passed). (ii) Boundary. The induced structure is a typed routing index with a sharp routing/readout boundary: slot codes scaffold routing but do not drive answer readout ($|Δ\hat{y}| \le 3.4\times10^{-6}$, zero collateral, three seeds, stable across a 5.6x scale window) -- we therefore make no behavioral-editability claim. (iii) Cost. The structure is free: LM quality matches a parameter-matched monolith within 0.0082 nats. (iv) Trust. The library state is exactly local under edit and bit-exactly revertible -- 250 single-edit and 1,000 stacked reverts per seed, zero failures. We further find that the unsupervised null itself moves with scale, so comparisons reusing a null calibrated at one scale may be confounded at another. Every claim is tied to a preregistered, machine-checkable criterion archived before the data it governs; the full audit trail, including one criterion we failed and how the frozen protocol handled it, is released as an appendix.

As users increasingly turn to Large Language Models (LLMs) for information and advice on political matters, particularly during election periods, the political preferences expressed by these systems have become a matter of public interest. Prior research has shown that interactions with LLMs can influence users' political attitudes and choices, raising questions about how these models themselves evaluate political actors. In this paper, we investigate whether and how LLMs express preferences toward political parties and political leaders. We introduce a systematic and reproducible auditing framework in which multiple LLMs are prompted to evaluate parties and leaders across nine criteria. Rather than attempting to infer the models' "true" political beliefs, we focus on their observable behavior, examining consistency across evaluations, differences between models, refusal rates, and sensitivity to prompt formulation. We further investigate how these evaluations vary when models are instructed to adopt different personas. We demonstrate the framework through an Italian case study, providing a systematic analysis of LLM-generated political evaluations on italian parties and leaders.

Prior white-box studies show that large language models can retain latent traces of target knowledge after unlearning, even when the knowledge is no longer expressed in their outputs. However, existing audits remain limited to one-off diagnostics: it is unclear whether these residual signals can predict future recovery under continued training or serve as reliable optimization targets. Resolving this gap is essential to determine whether internal auditing can move beyond post-hoc evaluation toward proactive risk monitoring and safer unlearning. We propose J-Access, an inference-time audit that uses the Jacobian lens to map intermediate representations into vocabulary space and measures how often target concepts remain accessible along the model's output pathway. We hypothesize that residual accessibility reflects recovery susceptibility: knowledge that remains closer to the output pathway requires less fine-tuning to restore, leading to faster recovery. We audit 398 public unlearned models spanning eight unlearning methods. We find that: (1) most unlearned models retain access above the retain-only gold level; (2) pre-attack accessibility predicts recovery speed and extent at the model level, but cannot identify which specific facts will be recovered; and (3) directly minimizing J-Access does not promote genuine deletion. Instead, the model learns to hide knowledge from the audit, producing lower audit scores but greater post-attack recovery. These findings position J-Access as a model-level diagnostic for assessing residual susceptibility in unlearned models. We argue internal audits should serve as an independent diagnostic dimension in unlearning evaluation, and should not be converted into optimization targets without validation.

Over the past several years, LLM-powered chatbots and agents have become widely used as a tool for academic writing. LLM-assisted writing can be valuable by removing language barriers but at the same time causes concerns about misconduct and fraud. To inform policy decisions, it is necessary to monitor the prevalence of LLM-altered texts in scholarly publications. Despite some recent progress in this direction, no existing method can produce reliable estimates. Here we suggest and validate a new unbiased approach to estimate LLM usage in a corpus of texts based on changing word frequencies. We apply our method to the full texts of open-access biomedical papers from Pubmed Central, and show that by the end of 2025, 89% of papers show excess of LLM-associated vocabulary. We also find that LLMs are twice as likely to be used when writing a paragraph in the Discussion section (68%) compared to a paragraph in the Methods section (32%), but even inside the Methods section, the overall prevalence of LLM usage is over 50%. We believe that our estimates are crucial to shape future guidelines and policies.

Pretraining corpus composition shapes LLM capabilities, but it often remains hidden even when model weights are released. Prior work has inferred corpus mixtures or traced specific token groups from released tokenizer vocabularies; in contrast, we estimate corpus ratios for arbitrary target tokens. We first show that BPE tokenizers trained on different corpora share stable token ID--ratio distributions, motivating distribution transfer from known corpora to a target tokenizer trained on hidden corpora. We then propose Quantile-Guided Density Estimation (QGDE), which approximates this distribution with multiple quantile trends and uses local density weighting to produce token-level estimates. In controlled settings and a realistic setting using the released SmolLM tokenizer, QGDE achieves mean relative errors as low as 3.00% for token-level estimation and 3.08% after aggregation into category-level mixtures. These results suggest that released tokenizer vocabularies provide a useful signal for fine-grained corpus estimation beyond coarse composition inference.

Chinese web pollution has surfaced in LLMs, motivating audits of upstream Chinese corpora. However, auditing such corpora faces three challenges: (1) their web-scale size makes full scan costly; (2) prior analyses are often too coarse to expose token-level pollution; (3) Chinese web pollution is implicit and rapidly changing. We propose Sampled-BPE, a lightweight token-level auditing pipeline that sample a small subset and train BPE tokenizer to surface polluted tokens. Experiments show that Sampled-BPE preserves usable estimates while substantially reducing runtime and memory: a 148.4 $\times$ speedup and a 35.8 $\times$ memory reduction induce only 4.25% relative error for pollution categories. We apply the pipeline to 11 open Chinese corpora and 6 Chinese Common Crawl snapshots from 2021 to 2026. The audit reveals widespread but uneven pollution across open corpora, as well as highly polluted and temporally shifting Chinese web content. We further release a hierarchical Chinese web token dataset with 660k+ token records, each with web context, category, and explanation fields, organized as trees to support review and tracing of pollution.

AI coding agents now write a growing share of production code, and human security review does not scale at the rate code is generated. The agents in widest use are closed-weight, so a deploying team cannot read their internals. It can instead run an open-weight model as a reviewer over the agent's output. That reviewer's activations are readable. We ask whether reading those activations recovers a security signal that simply asking the same reviewer misses. We fit a single linear probe per model on a corpus of paired vulnerable-and-fixed Python functions, then test it without retraining on real disclosed vulnerabilities whose weakness type the probe never saw in training, across five open-weight reviewer models. On the vulnerabilities fixed by changing a single function, the probe scores the vulnerable function above its fix on 61-67% of cases for every model, beating the 50% chance line. It also beats the same model's prompted YES/NO win-rate read from its logits, under every prompt we try. Asking the model for a written verdict, even with chain-of-thought, returns the same answer on the vulnerable and fixed function most of the time and so cannot tell them apart. Model activations carry a code-security signal that prompting the same model misses.

Fairness audits for LLM-based recommenders have largely focused on observable outputs, implicitly assuming that stable recommendations reflect stable internal processing. We challenge this assumption with FairGap, the first benchmark to jointly evaluate recommendation fairness at two levels: observable output shift (OBS) and hidden representation shift (IBS), measured through controlled counterfactual identity probes across gender, age, and race. Their relationship is summarized via Representation-Output Alignment (ROA), with quadrant diagnostics for identifying user-level hidden-output mismatch. Applied to six open-weight LLM families across three domains, FairGap reveals pervasive hidden-output decoupling: ROA rarely exceeds 0.22, and a non-negligible user population shows stable outputs despite substantial internal shifts, a mode that output-only audits cannot detect by design. Further, activation steering that reduces IBS by up to 8x simultaneously worsens OBS, demonstrating a fundamental tension between internal and output-level fairness that existing frameworks are unequipped to diagnose.

Large language models (LLMs) are increasingly reported to exhibit human-like neural and cognitive signatures, including concept cells, mental number lines, and cognitive maps. These claims often rely on linear probing and activation steering applied to a single model, yet both methods are highly sensitive to measurement choices. A reported parallel may therefore reflect the model, the measurement procedure, or both. We audit four representative neuroscience-inspired paradigms across 17 models from five families, spanning $0.6$B to $72$B parameters. Our main experiment examines the causal steerability of concept directions. With raw activation units and a fixed layer and coefficient, steerability appears to increase with model scale, resembling an emergent capability. However, this pattern is produced by an uncalibrated pipeline rather than by a claim established in the steering literature. The trend depends jointly on raw units, the readout metric, and the operating point; correcting any one of these removes it. With residual-norm-comparable interventions and held-out operating-point selection, concept steering remains significant at every scale, but shows no significant trend across the Qwen3 series, although the confidence interval does not rule out a moderate positive slope. The remaining results are mixed. A linear geographic world map is consistently decodable in every tested checkpoint up to $72$B. Number magnitude is strongly encoded, but whether individual neurons appear bell-shaped or monotonic depends on the selection criterion. Language-specific structure is localizable, but the direction of the cross-lingual asymmetry reverses under a different attribution method. These results suggest that the main constraint on AI neuroscience is not a lack of phenomena, but a lack of comparable measurements and adequate controls. We release the protocol, stimuli, and code.

Establishing the provenance of a language model---including its base checkpoint and possible overlap in training distributions---is a governance challenge that metadata alone cannot resolve. We introduce a training-free fingerprint based on the top-$k$ vocabulary projections of late hidden states elicited by 250 fixed knowledge probes, compared using Jaccard overlap over decoded token strings. We evaluate the method on 32 open-weight models from nine families (0.6B--32B) with documented relationships. (1)~A \emph{similarity ladder} broadly follows model relatedness: independently trained models on identical data score 0.48 raw (0.35 vocabulary-corrected), followed by shared-base fine-tunes (0.39/0.33), same-developer relatives (0.38/0.28), and models with no documented relationship (0.22/0.17). This identical-data signal persists across three organizations, two tokenizer families, and two architecture classes, and emerges within the first 1% of training before measurable task competence, suggesting a contribution from shared training data beyond capability convergence. (2)~As a nearest-neighbor \emph{lineage-retrieval} method, the fingerprint ranks the exact documented base among the top two candidates for all five R1 distillations (mean rank 1.8, MRR 0.60), including a math-specialized base not identifiable from coarse metadata. (3)~A \emph{depth ablation} shows that lineage group discrimination strengthens toward the output distribution, with AUC increasing from 0.72 at quarter depth to 0.90 at the output; using only the top 5 output tokens retains AUC 0.87. (4)~The fingerprint remains stable under quantization, with Jaccard similarity of 0.92 under int8 and 0.82--0.85 under int4, compared with a maximum cross-model similarity of 0.81 in the calibration pool. We release the probes, code, and fingerprints.

LLM recommenders for top-K item suggestion regularly emit titles outside the target catalog. Prior audits report a binary out-of-domain rate; none ask whether the model knew. We jointly audit hallucination rate (OOD@10) and verbalized-confidence calibration (ECE, Brier, reliability) for four zero-shot LLM recommenders from four independent vendors (Mistral Large, Llama-3.3-70B, GPT-OSS-120B, Claude Sonnet 4.6), not grounded or fine-tuned systems, across three catalogs (MovieLens-25M, Amazon Reviews 2023 Toys, Yelp Open Dataset), stratified by item popularity. Measuring catalog membership is itself the hard part: on identical outputs the reported rate moves by an order of magnitude with the string matcher used, and F1 cannot separate the candidates. We validate the instrument against 201 human judgments and select on net bias, where the adopted one is off by -0.040 against +0.144 for the common fuzzy rule. Hallucination is then strongly catalog-dependent (0.6-2.7% on MovieLens, 11.6-38.7% on Yelp, 49.3-61.0% on Amazon Toys). Each model holds a near-constant confidence level barely responsive to the catalog, while the catalog-hit rate swings 60 points, so the sign of the error is set by where a model's constant lands against a catalog's accuracy: 7 of the twelve cells are under-confident and 5 over-confident, all four under-confident on MovieLens, all four over-confident on Amazon Toys. We read this as an elicitation mismatch: "Just Ask" elicits a generic quality rating, not a catalog-membership probability. A conformal abstention threshold over verbalized confidence changes hallucination by at most 1.65 pp across four alpha levels, because the channel cannot separate correct items from hallucinations. We recommend that audits report calibration alongside OOD, validate the matcher producing the OOD number, and use catalog-anchored elicitation.

LLM benchmarks can build an organization's reputation and attract customers, but only when results are transparent and verifiable. Unverified claims that DeepSeek R1 outperformed OpenAI's o1 contributed to market panic on January 27, 2025, when Nvidia lost USD589 billion in market value. Yet vendor benchmarks often depend on an honor system. Academic reassessments and independent leaderboards have found undisclosed changes to proprietary models, contaminated training data, and selective reporting. LLM-as-a-judge methods scale evaluation by reducing human review. Studies, however, suggest that judges may show identity-aware bias, scoring an answer according to its source model rather than its quality. This bias has not been fully measured or corrected across politically sensitive, reasoning-intensive, and preference-based tasks. We examine this problem using seven verifier models: GPT-OSS 120B, Llama 3.3 70B, GLM 5.1, Qwen3 32B, DeepSeek V4 Pro, Mistral Large3, and Sarvam M. They score anonymous and identity-disclosed responses from three primary models on 58 factual, reasoning, political, and preference-based questions. Identity disclosure slightly raises scores for factual questions, moderately affects stress-reasoning tasks, and causes large changes for geopolitically sensitive topics. Notable results include GLM5.1 (+7.00 points, p = 0.0249) and Llama 3.3 70B (+1.56 points, p = 0.00). We also introduce a blockchain-based commit-reveal protocol using Autonomous Economic Agents on an Ethereum-compatible ledger. In Phase 1, each judge records a one-way hash of its score and a secret salt before candidate identities are revealed. In Phase 2, the identity and raw score are disclosed and verified on-chain. This creates a tamper-evident audit trail that separates blind evaluation from post-hoc claims and reduces the verification burden on independent researchers and leaderboard operators.

With the rapid advancement of large language models (LLMs), harnesses have become essential infrastructure for deploying agents across a wide range of domains. The fast-evolving harness ecosystem has also made rigorous capability evaluation increasingly important. However, efficiently building an end-to-end, systematic, and comprehensive evaluation pipeline remains a significant challenge. To address this challenge, we introduce $A^2E$ (Agent Auditing Engine), an end-to-end evaluation engine designed for agent harnesses. $A^2E$ leverages our newly proposed Agent Task Protocol (ATP) to enable the rapid integration of evaluation tasks with different harnesses. Through an automatically instrumented Monitor, it captures and generates standardized execution traces during experiments. In the Evaluation stage, $A^2E$ systematically assesses harness capabilities using a suite of multidimensional metrics. Compared with correctness alone, these metrics provide a more fine-grained characterization of differences among harnesses in execution efficiency, tool use, task planning, and error recovery. Experiments conducted with $A^2E$ further reveal that model-harness combinations exhibit substantial performance variation across different types of tasks, and that no single combination consistently outperforms all others across every task. These findings not only demonstrate the necessity of systematic evaluation but also provide useful guidance for the co-evolving of models and harnesses. Our code is available at https://github.com/datamllab/A2E.

Prior benchmarking work has shown that a single large language model (LLM), forced to make life-or-death resource-allocation decisions, exhibits measurable demographic bias. Real deployments, however, rarely use a single agent: they use pipelines, with review steps meant to catch exactly this kind of failure. We study what happens to bias when the same decision is distributed across a role-differentiated multi-agent pipeline (assessment, allocation, independent audit) instead of made and checked by one model alone. Using a synthetic disaster-triage simulator with paired cases that are clinically identical except for one demographic attribute, we run 192 episodes (2,304 resolved case pairs) on GPT-4o-mini comparing a single-agent control condition to a nine-agent pipeline under three independently varied pressure dimensions. We find no measurable difference in how often biased outcomes occur between the two conditions (6.9% vs. 6.1%, p = 0.498). We do find a large and significant effect of audit capacity on whether bias is caught: 30.0% of biased outcomes go entirely undetected, rising to 43.8% when the auditor is overloaded and falling to 18.4% when it is not. Decomposing this effect shows it is driven almost entirely by coverage (whether a case is reviewed at all, which collapses from 100.0% to 65.6% under load, p < 0.001) rather than by degraded judgment on the cases that are reviewed (81.6% vs. 85.7%, p = 1.000, direction reversed). A follow-up experiment shows that reordering the audit queue by estimated risk, rather than first-come-first-served, recovers most of the lost coverage under the same capacity constraint (65.6% to 91.7%, p = 0.028). We discuss the implications for any system that adds independent oversight to an LLM agent pipeline under resource constraints, and report the study's limitations honestly: one model, modest sample sizes, and no adversarial replication.

Most language-model access controls regulate behavior while leaving the same computation available to every request. We study a different systems question: can trusted authorization determine which newly trained parameters are reachable by the forward pass? Policy-Masked Private Experts freezes a pretrained sparse Mixture-of-Experts (MoE) model, trains a disjoint expert branch, and selects the public or private pool before top-k routing. The resulting claim is narrow but testable: under the declared trusted computing base (TCB), an unauthorized request executes no private expert. It does not imply that the public model lacks the same semantic capability. We test this separation between execution control and task utility in Qwen3-30B-A3B and DeepSeek-V2-Lite. Three Qwen BF16 seeds update all 32 private experts while the public fingerprint remains unchanged. Across 64 adversarial scenarios and 96 deny/fail-closed events, unauthorized private execution is zero; independent hooks exactly match 11,616 routed private rows and allow-deny-allow recovery is exact. On two prospectively frozen Qwen benchmarks, the private branch improves exact tool use by 5.0 percentage points (pp) (five versus zero discordances; one-sided Holm p = 0.03125, corresponding two-sided exact p = 0.0625) and 21.3 pp (percentile-bootstrap 95% CI [13.3, 29.3], Holm p = 0.000031). Three arm-blinded model evaluators retain a positive external effect of 18.7 pp (95% CI [9.3, 28.0]). A parameter-matched Lora has similar external utility, but a post-hoc request gate leaves 1,225 adapter calls under deny; the disjoint expert branch leaves none. DeepSeek reproduces the route invariant and gains 27.0 pp. A valid sealed evaluation is near-neutral. These results support auditable, reversible control over a trained parameter path, while showing that useful transfer remains distribution dependent.

Large language models are increasingly used to read markets, assess risk, and allocate capital. However, reported results for LLM trading agents can be inflated by look-ahead leakage, optimistic execution, and risk mandates that are described but not enforced. We present OpenPM, an auditable point-in-time evaluation framework for LLM portfolio-management agents. In OpenPM, an agent manages a $1M long-only book over the S&P 500 universe using market data at five-minute intervals. Every record visible to the agent must be available at the decision time. Natural-language risk mandates are converted into typed constraints and enforced on the executed portfolio. Each run produces audit artifacts, including a contamination certificate, a cost-sensitivity curve, and a constraint-adherence report. We also build a reference agent named the tiered allocator, where typed analysts score candidates, a constructor LLM proposes weights, and a deterministic critic guarantees feasibility. We isolate constructor behavior by capturing analyst evidence once and replaying it across constructor models. In our short-window case study, stronger constructors show modest and model-dependent gains over equal weighting on the same pool, but analyst quality matters more than constructor choice, and turnover is the main cost driver. All returns are upper bounds on a single frozen window without market impact, not validated alpha.

Measuring political bias in large language models (LLMs) remains challenging as it can manifest through subtle differences in framing, argumentation, and legal reasoning that are difficult to capture with a single metric. In this work, we introduce Poli-Bias, a counterfactual framework for measuring whether LLMs treat legally equivalent conflict scenarios differently depending on the countries involved. Poli-Bias compares responses to paired prompts in which country identities are systematically swapped across diverse geopolitical relationships, legal violations, and reasoning tasks. Rather than reducing bias to a single judgment, our framework decomposes response disparities into five interpretable dimensions, revealing how and where unequal treatment manifests. Across 13 contemporary LLMs spanning diverse model families and sizes, we find that country identities and user affiliations can systematically affect how equivalent actions are described, evaluated, and defended under international law. Our results thus establish Poli-Bias as a fine-grained framework for auditing political even-handedness and sycophancy in LLMs.

Search-agent rewards mix answer quality, citation grounding, tool cost, and anti-hacking terms; a high score therefore need not imply that cited evidence was retrieved, and added penalties can cancel. We introduce HERALD, an offline audit that applies exact same-question interventions, separates candidate-visible from oracle information, and enumerates detector contracts before policy optimization. On four Qwen3-8B pools from HotpotQA, 2WikiMultiHopQA, and MuSiQue, $R_0$ rejects search deletion and fake IDs, but a label-free citation-laundering attack succeeds. A complete $2^3$ ablation identifies targeted strengthening of $L$---citing a corpus passage absent from the retrieved evidence---as the observed inclusion-minimal repair: $R[L]$ has zero empirical ASR with a 0.50% one-sided cluster upper bound. The gap persists across pool rules, a visible BM25 attacker, and four models; broader hardening remains vulnerable when the attack removes an oracle support-ID penalty. Under strict 5M-token matched training evaluated on 256 paired questions per benchmark, $R[L]$ meets the EM non-inferiority gate on HotpotQA and 2Wiki but not MuSiQue. Equal-suite citation precision and support recall improve by 2.02 and 1.46 points, unsupported citations fall by 1.69, and laundering attackability falls on 2Wiki and MuSiQue. Natural $L$ is not reduced, and the detector appears in only 18 of 58,368 training trajectories. HERALD thus separates robust scoring, sparse learning signal, and policy transfer.

Modern models no longer keep a plain KV cache: latent caches, learned sparse selectors and recurrent states each carry the model's memory in a different form, and each fails differently under compression. We give a runtime observability contract that covers all four memory classes with three operators, instantiate it on six model configurations across five architecture families, and compose the per-stage bounds into an executable request-level risk ledger. Contracts carry their error metric as a type -- composition is only defined when metrics match, and this check rejected our own first composed chain; the repaired chain crosses metrics through two proved bridges, and whatever no formal system can certify is measured instead, dropping the composed tier to empirical automatically: every claim is certified, partially certified, or empirical, composition inherits the weakest tier, and the tier is decided by the machine. Replayed over $12.4$M entry reads and run under eight-way concurrency with per-request budgets and fail-closed identity attribution, the ledger quantifies the honest trade-off on today's witness and holds its risk budget with zero violations. A fused always-on probe observes a declared one-layer subset under CUDA graphs inside the serving noise floor. Applied to a served DeepSeek-V4 stack with a packed compressed-KV prototype, the same machinery localizes a silent corruption to a precise structural boundary -- exact in the eviction-free, identity-isolated regime, with every observed failure in an eviction or slot-reuse regime -- through a machine-adjudicated discrimination campaign whose calculus rejected two of our own confounded inferences along the way. All artifacts, guards, and the Lean development are released at https://github.com/metask-ai/witprobe-attention-memory; every number in this paper regenerates from the shipped artifacts by one command.

Subliminal Learning (SL) is a surprising type of generalization displayed by modern language models. It allows the transfer of a bias or behavior from a teacher model to a student by distilling from seemingly unrelated or random synthetic data from the teacher. This presents challenges in ensuring AI systems remain predictable and are trained safely, as standard auditing of the input data would not catch the hidden subliminal signal. Here, we investigate several open questions as to the enabling mechanisms and drivers of SL. First is the nature of the process by which biases are encoded in the data. We find that by adding Gaussian noise to the weights of the teacher and student models, the magnitude of subliminal transfer is increased by a factor of 1.9 in Gemma and 1.3 in Llama, suggesting that non-semantic weight structures play a crucial role. We show that steering vectors can be applied to the teacher to produce subliminal data, in addition to prompting and finetuning as used in previous studies. Analysis of the activations of the student models that have been trained on steered and prompted data demonstrates that students inherit not just the semantic meaning of the teacher's bias, but also the type of intervention that was used to apply it: steered students imitate steering vectors, prompted students do not. Additionally, the gradients of steered subliminal data show a linear correlation with the teacher's steering vectors, showing promise for data auditing. More broadly, as synthetic data becomes central to frontier training pipelines, being able to see the latent signals hidden in training data becomes paramount.

Language models differ in how safely they behave and these differences are measured by safety benchmarks. But aggregated benchmark scores are hard to trust and interpret, because benchmarks duplicate one another, correlate heavily, and models may sandbag when they detect evaluation. To address these issues, we draw on Item Response Theory (IRT), a statistical toolkit for measuring these latents from performance on items with inferred psychometric properties. We fit IRT models to eight safety benchmarks across 192 language models, the largest psychometric analysis of LLM safety evaluations to date, and contribute three results. First, we find that three interpretable factors of refusal strictness, truthfulness, and contextual harm explain most of the variance between models across benchmarks. Second, psychometrically selected items recover full benchmark scores with lower error than random subsets of the same size, and roughly ten adaptively chosen items suffice for several individual benchmarks, cutting evaluation cost by 97-99%. Third, IRT supports audits of individual models, showing that it can be used to detect naive sandbagging and changes of model behind APIs. Overall, we show IRT is a ready-made toolkit for reading, reducing, and auditing safety benchmarks, which we recommend frontier labs and evaluators adopt.

Deep search agents tackle challenging questions through long-horizon web interactions, a process that is both complex and fragile: small reasoning errors may propagate through long, noisy trajectories into fluent but incorrect answers. Diagnosing such failures is difficult, requiring the manual inspection of extremely long execution traces, which could be beyond human capacity. We therefore introduce SearchAuditBench, a benchmark that evaluates whether LLM auditors can localize, attribute, and repair these failures, thereby reducing the human burden. SearchAuditBench comprises 1,243 failed trajectories, averaging 73.1 messages and 65.1K tokens, collected from eight open-weight models on five deep-search benchmarks, each expert-annotated with the critical error step, a search-specific root cause, and a reference repair with grading rubrics. We further propose SearchAuditor, a multi-perspective auditing framework that effectively localizes, attributes, and repairs search-agent failures through evidence-grounded adjudication. Experimental results show that even the strongest baseline, when powered by a frontier model like GPT-5.5, attains only a 26.6% end-to-end pass rate. In contrast, our SearchAuditor consistently outperforms all baselines across different frontier models, achieving an end-to-end pass rate of 32.3%, and resuming failed runs with its repairs enables agents to better recover from errors.

Fine-tuning enables a source model to acquire desired capabilities and behaviors in a target domain while retaining much of its general-purpose competence. However, this adaptation process can also degrade alignment properties that were present in the source model. Recent work has shown that large language models can be trained using LoRA-based modules known as introspection adapters (IAs) to describe behavioral changes induced by fine-tuning. However, existing studies primarily consider settings in which the model is fine-tuned on datasets explicitly designed to implant a specific behavior and is then asked to explain the implanted behavior. This differs from practical deployment scenarios, where the central concern is often side-effect misalignment: unintended degradation of alignment caused by fine-tuning on tasks that are not obviously related to safety or alignment. To bridge this gap, we formulate a novel problem setting called \emph{side-effect introspection}, in which the target of introspection is not a behavior explicitly implanted through fine-tuning, but rather alignment shifts that emerge as unintended side effects, and we construct a dataset for this setting. Furthermore, to enhance sensitivity to internal model changes, we propose the Delta-Aware Introspection Adapter (DAIA), a novel mechanism designed to explicitly process both base-model activations and activation differences induced by fine-tuning. Our empirical evaluation shows that introspection learning generalizes to unseen fine-tuned models and safety categories, and that DAIA consistently outperforms existing introspection adapters.

Reasoning-based guard models improve LLM safeguards, but decoding explicit rationales for every interaction makes them costly to deploy. Although latent-reasoning methods reduce token generation by moving reasoning into continuous states, they remain underexplored for safety moderation and lack an inspection interface for deployment. In this paper, we propose LatentGuard, an efficient and inspectable safeguard framework that brings continuous latent reasoning to guard models. LatentGuard uses a staged curriculum to progressively compress task-aligned textual rationales into compact latent states, enabling safety verdicts to be predicted directly from continuous representations. To preserve inspectability, an isolated auxiliary decoder generates compact audit artifacts on demand, keeping rationale generation off the standard inference path. Experiments show that LatentGuard-8B improves mean weighted F1 from 83.95 to 84.91 over GuardReasoner-8B, while reducing critical-path reasoning cost from 268.56 generated rationale tokens to 1.60 latent reasoning tokens. Its audit decoder achieves an audit utility score of 85.75, demonstrating an efficient and inspectable path toward deployable LLM safeguards.

Benchmark gains are often treated as evidence of greater LLM capability. Yet the same gain can reflect different changes in model behavior. A model may reach new answers, or produce answers that were already within reach. Aggregate scores do not distinguish these changes question by question. We establish a question-level audit under fixed budgets, temperatures, and answer formats. A question is realized when the default deployment procedure produces the correct answer. A question is reachable when a specified probe finds that answer within a fixed budget. We first test whether inference-time layer routing can expand reachability. Under a matched budget, random routes match or exceed structured search in all 43 model and task settings. Answer-blind procedures retain almost none of this gain, which instead requires access to the correct answer. We then ask why reachable answers sometimes fail to appear. Across six cases spanning 0.5B to 31B, silencing one identified MLP block repairs 68 to 92 percent of a predefined failure set. We next test whether training closes the gap by expanding reachability. In five of six matched evaluations, deployed performance rises while the reachable ceiling remains flat or falls. For DAPO, the deployed score rises by 14.7 points while the reachable ceiling falls by 13.3 points. Across the settings we audit, realization and reachability therefore do not always change together. Claims of capability expansion should report both realized performance and reachability under matched evaluation conditions. Code is available at https://github.com/LiZaiyuan0619/reachability-not-realization

The standard check for contamination in LLM backtests is simple: compare scores before and after the training cutoff. We show this check is uninformative. Four flagship models fail it on questions they cannot have memorized: every scored question resolved after their cutoffs. The reason is structural. Models legitimately know more about times near their cutoff, so recency mimics leakage, and we prove no passive backtest can separate the two from genuine skill. Measurement, not just detection, requires information from outside the backtest. We supply it in two forms. A known cutoff identifies leakage at the boundary; a matched clean control identifies it globally and yields a leakage-adjusted score. We also derive where leakage hides: it concentrates on outcomes that surprised the crowd and were well covered in training, and partial memorization is disproportionately rewarded. We validate the estimators against ground truth by planting leakage in twin models, where they recover the injected dose and return null on clean questions. Deployed on frontier models, they detect one cutoff-localized signature and, at the audit's power floor, clear five models whose apparent advantages were recency alone. Backtests need not be discarded; they need one defensible reference.

Chain-of-thought (CoT) monitoring is an increasingly important component of AI safety stacks but relies on the assumption that a model's reasoning trace is informative about its actions. This work studies the limits of CoT monitoring through the lens of model poisoning. We demonstrate that backdoors can be implanted into reasoning models to elicit an attacker-chosen behavior while their CoT traces appear entirely benign. We find that these CoT-Hidden backdoors can be induced through simple fine-tuning recipes across reasoning-model architectures and sizes. When direct poisoning is ineffective, we introduce a curriculum training approach that progressively teaches the model to produce an attacker-chosen output while concealing the behavior from its reasoning traces. These findings suggest that CoT monitoring may be better framed as a question about the consistency between a model's reasoning trace and its final response than as anomaly detection within a trace. We further examine the mechanisms that allow models to suppress evidence of the target behavior from their reasoning traces. Causal interventions locate a trigger-conditioned activation pathway that does not depend on the visible reasoning, and residual stream verbalizations provide an anomaly warning near answer generation, but do not identify the trigger, target, or backdoor mechanism.