Daily Papers

Self-evolving agents turn experience into reusable skills, workflows, or memories, but post-evolution accuracy alone does not show whether learned behavior preserves previously correct behavior or security. We audit SkillOpt, Agent Workflow Memory (AWM), and ReasoningBank in simulated e-banking using matched benign acquisition trajectories, sealed evaluation endpoints, execution-grounded checks, and independent state replay. On Qwen 3.7 Flash, SkillOpt raises benign utility from 0.741 to 0.837 while exposure to injected content rises from 0.820 to 0.943. Conditional attack success after exposure falls from 0.605 to 0.562, yet overall attack success rate (ASR) rises from 0.496 to 0.530 and unauthorized financial state changes rise to 0.685. Across three independently evolved lineages, capability, exposure, and unauthorized-state changes increase in all three, whereas ASR increases in only two. ReasoningBank raises utility to 0.859 without increasing aggregate ASR, although unauthorized state changes remain slightly above Static. AWM reveals a separate evaluation hazard: a literal WebArena text-action envelope disrupts tool execution in our native function-calling executor. In a post-hoc sensitivity test, removing only that envelope restores utility from 0.319 to 0.756, while exposure rises from 0.299 to 0.909 and ASR from 0.195 to 0.575. Auditing self-evolving financial agents therefore requires tracking regressions, attack-surface contact, unauthorized financial-state change, and artifact-executor compatibility, not accuracy alone.

AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Discovery Certification Protocol (DCP) turns claims about these results into executable recovery and feedback tests. Gate 1 validates useful improvement on sealed evaluation. Gate 2 gives matched agents the registered starting information and observed Web content while withholding the target research history. Every valid method reaching the numerical target supplies a recovery witness and triggers the Core veto. DCP Core requires adequate controls, zero observed recoveries, and a finite-sample bound on recovery in one fresh registered episode. Optional Gate 3 measures the average effect of truthful feedback relative to a specified neutral policy from a shared checkpoint. DCP Evidence adds this effect after independent null calibration and a registered effect margin. Two controlled audits exercise the complete protocol in SQLite optimization and virtual catalyst control under different models. Each produced zero recoveries in 96 episodes, with an upper bound of 0.0468. Each paired study yielded 30 truthful recoveries and zero neutral recoveries, with passing 60-pair null studies. Additional cases exercise Core, recovered, and audit-incomplete decisions. A deterministic, LLM-free verifier reproduces the decisions from frozen evidence. DCP provides a common evidence language for useful outcomes, alternative routes, and feedback effects across AI research.

We introduce SealQA, a new challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) Seal-0 (main) and (2) Seal-Hard, which assess factual accuracy and reasoning capabilities, with Seal-0 focusing on the most challenging questions where chat models (e.g., GPT-4.1) typically achieve near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in "needle-in-a-haystack" settings. Our evaluation reveals critical limitations in current models: Even frontier LLMs perform poorly across all SealQA flavors. On Seal-0, frontier agentic models equipped with tools like o3 and o4-mini achieve only 17.1% and 6.3% accuracy, respectively, at their best reasoning efforts. We find that advanced reasoning models such as DeepSeek-R1-671B and o3-mini are highly vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across o3-mini, o4-mini, and o3, with performance often plateauing or even declining early. Additionally, while recent models are less affected by the "lost-in-the-middle" issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at huggingface.co/datasets/vtllms/sealqa.

Large language models (LLMs) have limitations in handling tasks that require real-time access to external APIs. While several benchmarks like ToolBench and APIGen have been developed to assess LLMs' API-use capabilities, they often suffer from issues such as lack of generalizability, limited multi-step reasoning coverage, and instability due to real-time API fluctuations. In this paper, we introduce SEAL, an end-to-end testbed designed to evaluate LLMs in real-world API usage. SEAL standardizes existing benchmarks, integrates an agent system for testing API retrieval and planning, and addresses the instability of real-time APIs by introducing a GPT-4-powered API simulator with caching for deterministic evaluations. Our testbed provides a comprehensive evaluation pipeline that covers API retrieval, API calls, and final responses, offering a reliable framework for structured performance comparison in diverse real-world scenarios. SEAL is publicly available, with ongoing updates for new benchmarks.

When an autonomous AI agent does something consequential, what can be proven about what it did? Agent-observability platforms capture traces, but a trace is mutable: alterable undetected, with no recipe for re-executing it, silent on whether captured secrets were removed. Regulation (EU AI Act, ISO 42001, NIST AI RMF) presumes records an independent party can check. We present NovaFabric, producing audit-grade execution evidence: provider-neutral, tamper-evident, replayable, shareable. It records an agent run, without modifying agent logic, into a portable Run Capsule (fifteen-entity schema), sealed with a holistic DSSE signature, RFC 3161 timestamp, Merkle log and redaction attestation. Sealed runs are re-executable under a four-mode replay protocol and exportable as an Evidence Bundle for third-party verification with stock tooling (specified, not evaluated). The contribution is integration, not new cryptography: OpenTelemetry, DSSE/in-toto and W3C PROV. We evaluate eight research questions at measured scope. Mocked replay serves every model response from the capsule (no live model call, 10/10) but is offline w.r.t. models, not the network; only 2/10 tool-using workloads completed; the gap is missing tool-response substitution. Tampering is rejected across three tested classes. Declared-stream completeness is 0.652 (95% CI +/-0.064, ten scenarios). A repaired rule pack redacts 14/14 credential types, preserving 9/9 decoys; diff localises 140/140 mutations. Blast-radius queries: 45.5ms p99 over 10M edges (3.3x faster than a columnar baseline), 167.9ms over 100M (1 client, n=30). A 314-machine, ten-region run finds capsule REST ingest lossless but capped at 61.6 req/s (p99 26.8s) by per-worker serialisation. Six defects found in NovaFabric and its evaluation corpus: four fixed, one withdrawn, one open. Verification is conditional on a stated trusted computing base.

Deep search enables language agents to solve complex information-seeking tasks through iterative web exploration. However, developing robust open-source agents remains challenging for three reasons. First, opaque trajectory-construction pipelines make training data difficult to audit and reproduce. Second, context compression can create a substantial mismatch between training and inference. Third, agents may retrieve benchmark content from the open web during evaluation, contaminating the results and overstating their performance. To overcome these bottlenecks, we introduce Glasseek, a fully transparent deep search framework spanning the complete development lifecycle, releasing all data, model weights, training recipes, and code. Specifically, for data construction, we establish a verifiable trajectory curation pipeline that systematically generates high-quality exploration paths. For training, we propose Compaction-Managed GSPO (CM-GSPO), which uses the same policy for search and summarization while optimizing only execution sub-trajectories, enabling end-to-end search-policy training across context-compaction boundaries. For evaluation, we conduct the first systematic audit of test-set leakage in open-web agents and release a modular anti-leakage plugin that seamlessly integrates into existing evaluation harnesses. Under this rigorous evaluation protocol, our 9B agent outperforms frontier proprietary agents across almost all deep research benchmarks spanning diverse task types, and achieves the best overall performance among recent open-weight agents, setting a new state of the art at the 9B scale.

Long-horizon tool agents often make useful progress without reaching terminal success, motivating partial-credit evaluation. Yet evaluators may reward milestones that were temporary, later reversed, or not attributable to the evaluated agent. Comparing an honest trajectory with a higher-scoring adversarial one is inconclusive if the latter made more genuine progress. We introduce PartHackBench, a controlled methodology that removes this confound. A private certifier admits a pair only when its trajectories match component-wise in both current-state predicate satisfaction and standardized agent attribution; score inflation, defined as f(A) - f(H), is measured only afterward. In 18 sealed held-out tasks in PB-CSTE, the frozen historical-target run produced matched adversaries for 15 tasks. Historical credit yielded mean inflation of .252, conditional attack success of 10/15, end-to-end yield of 10/18, and detected none of 14 strict rollbacks. Semantic LLM judges were more resistant but remained vulnerable, especially under evaluator-targeted attacks, while PB-CSTE current-state controls, defined as exact functions of the certified components, yielded zero inflation by construction. PartHackBench thus provides a certified control for testing whether evaluator credit changes while all benchmark-defined task-relevant progress remains fixed.

LeWorldModel trains a latent world model with a prediction loss and a single anti-collapse regulariser, and reports approximately 87% of goals reached on TwoRoom, its simplest diagnostic environment. We reproduce that result by independent reimplementation on roughly $25 of rented compute, with all evaluation on one laptop CPU. We reach 94.0% at the repository's evaluation goal offset, against 84.0% for the authors' own released checkpoint measured under our protocol on identical episodes, and we reproduce the reported representation result directly (position probe Pearson r = 0.9988 against a reported 0.996). Reaching that point required four conventions that determine the outcome and appear in no released configuration file: dense action gathering across a frameskip block, a programmatically-set action-encoder width, ImageNet pixel normalisation, and action z-scoring. A reproducer following the released configurations alone obtains a model whose predictor cannot converge. The evaluation protocol is itself contested by the released material. The paper's appendix and the repository's configuration specify different goal offsets and step budgets; on the authors' own weights these yield 14.0% and 84.0%, and only the configuration's values reproduce the reported figure. On fifty identical episodes, changing nothing but how the goal is constructed moves that checkpoint from 84.0% to 8.0%. Two findings generalise. One-step prediction accuracy does not predict long-horizon planning success: across three checkpoints spanning a sevenfold range in prediction error, including the authors' own, it orders short-horizon success monotonically and fails to order long-horizon success at all. And a batch normalisation layer inflated our reported validation loss by up to a factor of 300, concealing a training loss that was flat throughout.

Current Large Language Models (LLMs) benchmarks are often based on open-ended or close-ended QA evaluations, avoiding the requirement of human labor. Close-ended measurements evaluate the factuality of responses but lack expressiveness. Open-ended capture the model's capacity to produce discourse responses but are harder to assess for correctness. These two approaches are commonly used, either independently or together, though their relationship remains poorly understood. This work is focused on the healthcare domain, where both factuality and discourse matter greatly. It introduces a comprehensive, multi-axis suite for healthcare LLM evaluation, exploring correlations between open and close benchmarks and metrics. Findings include blind spots and overlaps in current methodologies. As an updated sanity check, we release a new medical benchmark--CareQA--, with both open and closed variants. Finally, we propose a novel metric for open-ended evaluations --Relaxed Perplexity-- to mitigate the identified limitations.

Benchmarks can run without determining what their results license. We freeze a large evaluation collection and attempt to replay its historical claims. Most units stop because the evidence required for replay is not bound. Where replay is possible, different claims remain stable at different resolutions. We make this otherwise implicit inference step explicit and executable.

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

Multi-agent large language model (LLM) debate is often evaluated by whether final answers improve, but movement is not necessarily improvement: the same discussion can rescue an initially wrong majority or destroy an initially correct one. Standard final-accuracy evaluations conflate these opposing mechanisms. We introduce an auditable protocol for homogeneous debate on multiple-choice questions (MCQs) that records each run as a transition ledger over collapse, correction, onset, and signed intervention utility. On 6,925 MMLU-Pro debates, the protocol identifies 253 collapses and a parallel correction ledger that changes how interventions should be judged. Replay experiments reveal the central tradeoff: a leave-one-model-out probe-gated freeze prevents 29 collapses but loses 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy. A compact pre-debate 8-probe screen is a triage signal: its unadjusted family-level association with conditional-collapse risk is high (G=7, Spearman rho=0.893, exact two-sided p=0.0123), but initial-majority accuracy is a close comparator (rho=0.821; family partial rho=0.767, p=0.0877), so we do not treat it as calibrated or capability-adjusted prediction. Round-level traces localize many collapses to the first debate round, where early disagreement can precede both harmful cascades and useful recovery. We release replayable schemas, coders, audits, cost cards, and zero-API rebuild scripts so future model-scaffold rows can be compared under the same denominators and signed utility ledger.

Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

Runtime-governance evidence often collapses materially different events into one audit record: a decision may be made, an instruction dispatched or received, an action may or may not execute, and a resulting state may or may not be observed. This paper presents AIREP, a vendor- and model-independent protocol for per-decision AI runtime evidence. Its current wire model separates evidence into four artifact families: Decision, Control, Execution, and Effect. Artifacts use closed core schemas, explicit identities and digests, declared scope limits, RFC 8785 canonical JSON, domain-separated SHA-256 hashing, and pure Ed25519 signatures. A three-level assurance model distinguishes structural/hash consistency (AIREP-Core), verifier-accepted authorship (AIREP-Authenticated), and independently anchored chain-head freshness and non-truncation relative to an accepted witness (AIREP-Witnessed); these classes do not establish event truth. A structured reconciler preserves failure, missing evidence, unevaluated prerequisites, and indeterminate outcomes as distinct states. The released beta includes a four-family first-party producer, Python and Node reference-verification paths, adversarial/lifecycle corpora, and reproducible validation. Post-release first-party Hermes and LightEval integration exercises preserve explicit evidence boundaries without claiming adoption or interoperability. Independent implementation evidence exists separately for a v0.1.2 producer and a v0.2 consumer/verifier; because they target different frozen versions, they do not establish same-version producer-to-consumer interoperability. AIREP remains experimental.

Benchmarks increasingly guide deployment, procurement and scientific screening, yet a score supports only the response it records, not necessarily the deployment action. We introduce deployment-complete benchmarking, which tests whether benchmark evidence determines a deployment action. A benchmark is complete for a claim exactly when the action is constant on each evidence fiber; mixed fibers expose missing deployment information, and completion curves quantify the evidence required to resolve ambiguity. In controlled response spaces, benchmark-channel conformal coverage of 94.98% transferred poorly to an unmeasured deployment channel (10.07%), whereas response-rank intervals achieved 94.91% coverage; even zero benchmark error certified only 45.4% of candidates at the largest residual size. Public audits revealed incompleteness, including 97.9% mixed Tox21 fibers and zero median certifiable fraction in main Matbench and JARVIS audits. In held-out replays, certify-then-acquire reduced false decisions from 1.19% to 0.027% in Tox21 and from 20.3% to 0.128% in JARVIS, while changing model choice and identifying deployment-relevant probes. Deployment-ready benchmarks should report evidence, supported actions, ambiguity and completion cost rather than scores alone.

Real-world Super-Resolution (Real-SR) methods focus on dealing with diverse real-world images and have attracted increasing attention in recent years. The key idea is to use a complex and high-order degradation model to mimic real-world degradations. Although they have achieved impressive results in various scenarios, they are faced with the obstacle of evaluation. Currently, these methods are only assessed by their average performance on a small set of degradation cases randomly selected from a large space, which fails to provide a comprehensive understanding of their overall performance and often yields inconsistent and potentially misleading results. To overcome the limitation in evaluation, we propose SEAL, a framework for systematic evaluation of real-SR. In particular, we cluster the extensive degradation space to create a set of representative degradation cases, which serves as a comprehensive test set. Next, we propose a coarse-to-fine evaluation protocol to measure the distributed and relative performance of real-SR methods on the test set. The protocol incorporates two new metrics: acceptance rate (AR) and relative performance ratio (RPR), derived from acceptance and excellence lines. Under SEAL, we benchmark existing real-SR methods, obtain new observations and insights into their performance, and develop a new strong baseline. We consider SEAL as the first step towards creating a comprehensive real-SR evaluation platform, which can promote the development of real-SR. The source code is available at https://github.com/XPixelGroup/SEAL

Context compression improves long-context efficiency but may introduce faithfulness issues through an information bottleneck. Existing evaluations rely on QA metrics, but a reader can still answer correctly from an unfaithful compression, so these metrics underestimate such loss. To bridge this gap, we introduce KEEP-Bench, a diagnostic benchmark that directly assesses compression faithfulness with 4,102 cases across three challenging settings that require keeping evidence for plausible reference resolutions, explicit qualifiers, and minority perspectives. Experiments show that even with tailored prompting, the strongest models remain unreliable: despite context windows exceeding 100K tokens and inputs averaging only 3.1K tokens, the best open-weight model (31B) preserves all required evidence in only 66.8% of cases, and the best proprietary model in only 61.6%. Performance also varies sharply across evidence structures, so conclusions from limited structures may be biased. We further introduce a metric, Evidence Preservation Return (EPR), to relate evidence recall to inference FLOPs, and find that it peaks at moderate scales (4B-8B), where recall remains low. Additional output budget mainly benefits already strong models and widens capability gaps, yet the stronger preservation of these models comes at a steep computational cost. These findings highlight the importance for future work to investigate compressors with strong information preservation under a small computation budget, rather than relying on a large output budget or model scaling.

AI pentesting agents are increasingly credible as offensive security systems, but current benchmarks still provide limited guidance on which will perform best in real-world targets. Existing evaluation protocols assess and optimize for predefined goals such as capture-the-flag, remote code execution, exploit reproduction, or trajectory similarity, in simplified or narrow settings. These tools are valuable for measuring bounded capabilities, yet they do not adequately capture the complexity, open-ended exploration, and strategic decision-making required in realistic pentesting. In this paper, we present a practical evaluation protocol that shifts assessment from task completion to validated vulnerability discovery, allowing evaluation in sufficiently complex targets spanning multiple attack surfaces and vulnerability classes. The protocol combines structured ground-truth with LLM-based semantic matching to identify vulnerabilities, bipartite resolution to score findings under realistic ambiguity, continuous ground-truth maintenance, repeated and cumulative evaluation of stochastic agents, efficiency metrics, and reduced-suite selection for sustainable experimentation. This protocol extends the state of the art by enabling a more realistic, operationally informative comparison of AI pentesting agents. To enable reproducibility, we also release expert-annotated ground truth and code for the proposed evaluation protocol: https://github.com/ethiack/ethibench.

Self-driving laboratories couple a learned design surrogate to an instrument, and on every new target domain it must answer four questions: when it can be trusted, how many target measurements it needs, which to make first, and when to stop adapting, switch models or escalate. We state a framework making each measurable (a competence ratio, a budget k^*(varepsilon), a time-to-target and a priced decision) and instantiate it without wet experiments on six public droplet-microfluidic and liposome tables as offline laboratories: 107 leave-one-group tiers whose rows are hidden labels. A sealed, pre-registered programme (two sealed negative validations and a pre-registered test on three more tables) and seven frozen closed-loop protocols, each run twice, yield 23 findings. Shift cost is stratified by shift type; no label-free statistic certifies individual candidates; random-split model-family rankings predicted neither the sealed nor the loop rankings. Warm-start residual boosting returns a surrogate to 1.25times its development reference within 20 target rows in 94% of chemistry tiers (27% under a strict noise-floor reading), 90% of laboratory and 40% of device tiers, and a budget planned from the shift type on other tables is not measurably worse on a new one. In the loop, one gain survives drift and shuffled-label controls on all tiers, carried by the sparsely sampled ones: proposing at once and warm-starting on every measurement reaches a 10% target in 2.80 measurements against 3.31 for the reference, and in 5.72 against 11.35 for random proposals on 301 hard units at a 3% tolerance. No label-free acquisition rule survives the shuffled-label control, no interval-aware proposal or other model family does better, and no threshold rule on the tested diagnostics beats always continuing once escalation is priced. No wet-lab validation is claimed.

Enterprise AI agents that share a memory store face two unaddressed risks: sensitive data can leak through legitimately computed results the requester could not derive, and departments can silently compute a same-named key performance indicator (KPI) through conflicting logic. Existing agent-memory systems (e.g., MemGPT, Zep, A-MEM) gate retrieval by content, ownership, and role, not derivation, missing a cached insight that embeds a forbidden column. We introduce the Analytical Memory Unit (AMU), a memory schema that attaches a full derivation (lineage) graph to every cached result, gated by a retrieval policy that serves a hit only when the requester is authorised for every column touched. Provided lineage recording is complete, we prove by construction that the policy blocks retrieval of results derived from a sensitive column outside the requester's permissions, at O(n) worst case -- a conditional design guarantee, not an empirical claim, that excludes derived features encoding sensitive information without naming their source. Eliminating measured leakage required 75-90% recorded lineage completeness, so we treat 90% as a conservative deployment target. Across six experiments, lineage-gated retrieval removes the 18.8-25.5% cross-department leakage naive content-gated memory suffers, keeping 81.5-82.6% of memory reuse at 13.8 microsecond worst-case overhead. A real-agent proof-of-concept with LLM-generated SQL is consistent with the guarantee: zero leaks over 9 round-trips, two conflicts caught automatically -- though a feasibility demonstration, not evidence of production viability. This offers a practical governance layer for shared agent memory, complementing source-layer access control and supporting EU AI Act compliance.

We present a reproducible pipeline for mapping Common Vulnerabilities and Exposures (CVEs) to MITRE ATT&CK Enterprise techniques from free-text vulnerability descriptions. Rather than relying on the CWE->CAPEC->ATT&CK derivation chain, whose table-expansion artifacts we quantify, we train a multi-label classifier on a curated gold dataset of 1,207 CVEs from expert MITRE Center for Threat-Informed Defense mappings. The resulting model approximately doubles recall@5 compared with a zero-shot embedding-similarity baseline and improves every ranking metric. We then investigate whether LLM-assisted labeling can extend the gold dataset. Initial experiments suggest contradictory conclusions: a single run indicates degraded performance, while averaging over five random seeds suggests a small gain. However, an independent replication and an expansion-size study (100--984 additional CVEs) show that the apparent improvement is an evaluation artifact. LLM-generated labels, with approximately 0.39 agreement with expert annotations, provide no reliable improvement at any expansion size and reduce rare-technique coverage at around 1,000 added CVEs (macro-F1 decreases by 0.04). The root cause is evaluation noise. Selecting checkpoints on a small test split effectively maximizes over many noisy evaluations, producing recall@5 differences of up to 0.05 between otherwise identical runs. Using a corrected protocol based on validation-split checkpoint selection, the gold-only model achieves recall@5 of (0.673 \pm 0.019), and repeating the decisive experiment confirms the null result for LLM expansion. A final scaling study shows that additional expert-curated data consistently improves performance, whereas LLM-labeled data does not, indicating that the classifier is limited by label quality rather than dataset size. All datasets, models, code, and training logs are publicly released.

Practical LLM unlearning is usually evaluated through two objectives: suppress target-specific knowledge and preserve non-target utility. In generative QA, this leaves a third behavior underspecified: when a target-adjacent prompt admits a broader answer without target-specific leakage, the model should answer at that level rather than leak, evade, or refuse. We study this specification problem in a controlled LoRA-GRPO RWKU setting, comparing four reward designs that span lexical suppression, anti-refusal shaping, rubric-based broad answering, and an explicit refusal contrast, with and without SFT warm-up. The experiments show that optimization success is not equivalent to behavioral unlearning: RWKU forget scores, held-out completion audits, terminal training-rollout audits, and training dynamics can point to different conclusions. We trace these disagreements to reward-hacking endpoints, policy-support limits in GRPO, benchmark probes that miss endpoint changes, and rewards that can select broad-topic answering with low semantic leakage during optimization.

Hampering the interpretation of benchmark scores, evaluation data contamination has become a growing concern in the evaluation of LLMs, and an active area of research studies its effects. While evaluation data contamination is easily understood intuitively, it is surprisingly difficult to define precisely which samples should be considered contaminated and, consequently, how it impacts benchmark scores. We propose that these questions should be addressed together and that contamination metrics can be assessed based on whether models benefit from the examples they mark contaminated. We propose a novel analysis method called ConTAM, and show with a large scale survey of existing and novel n-gram based contamination metrics across 13 benchmarks and 7 models from 2 different families that ConTAM can be used to better understand evaluation data contamination and its effects. We find that contamination may have a much larger effect than reported in recent LLM releases and benefits models differently at different scales. We also find that considering only the longest contaminated substring provides a better signal than considering a union of all contaminated substrings, and that doing model and benchmark specific threshold analysis greatly increases the specificity of the results. Lastly, we investigate the impact of hyperparameter choices, finding that, among other things, both using larger values of n and disregarding matches that are infrequent in the pre-training data lead to many false negatives. With ConTAM, we provide a method to empirically ground evaluation data contamination metrics in downstream effects. With our exploration, we shed light on how evaluation data contamination can impact LLMs and provide insight into the considerations important when doing contamination analysis. We end our paper by discussing these in more detail and providing concrete suggestions for future work.

Open-weight LLMs are released not only as fixed products but also as substrates for downstream fine-tuning. This openness, however, creates legal and ethical risks because users may misuse fine-tuning to instill illicit knowledge or enable hostile operations. Model providers therefore need apre-release defense against such acquisition, motivating the problem of preemptive unlearning. Unlike retrospective unlearning, which removes capabilities already present in a fixed model, preemptive unlearning seeks to prevent their acquisition under unseen attack data and future fine-tuning procedures. Despite its practical importance, this setting remains largely unexplored, presents distinct challenges, and is therefore the central focus of our work. We first verify that existing retrospective methods provide insufficient pre-release protection. Even when forbidden capabilities are suppressed in current outputs, forbidden-domain data can still induce gradients through internal pathways, enabling later acquisition. Motivated by this finding, we propose a gradient-sealing principle that blocks these pathways by pushing relevant pre-activations into the negative region, where ReLU-family activations exhibit zero or near-zero derivatives. Experiments across multiple LLM families demonstrate our stronger resistance to downstream acquisition than retrospective baselines, validating gradient sealing as an effective mechanism for pre-release protection.

Multiple-choice question (MCQ) datasets like Massive Multitask Language Understanding (MMLU) are widely used to evaluate the commonsense, understanding, and problem-solving abilities of large language models (LLMs). However, the open-source nature of these benchmarks and the broad sources of training data for LLMs have inevitably led to benchmark contamination, resulting in unreliable evaluation results. To alleviate this issue, we propose a contamination-free and more challenging MCQ benchmark called MMLU-CF. This benchmark reassesses LLMs' understanding of world knowledge by averting both unintentional and malicious data leakage. To avoid unintentional data leakage, we source data from a broader domain and design three decontamination rules. To prevent malicious data leakage, we divide the benchmark into validation and test sets with similar difficulty and subject distributions. The test set remains closed-source to ensure reliable results, while the validation set is publicly available to promote transparency and facilitate independent verification. Our evaluation of mainstream LLMs reveals that the powerful GPT-4o achieves merely a 5-shot score of 73.4% and a 0-shot score of 71.9% on the test set, which indicates the effectiveness of our approach in creating a more rigorous and contamination-free evaluation standard. The GitHub repository is available at https://github.com/microsoft/MMLU-CF and the dataset refers to https://huggingface.co/datasets/microsoft/MMLU-CF.

Data contamination hinders fair LLM evaluation by introducing test data into newer models' training sets. Existing studies solve this challenge by updating benchmarks with newly collected data. However, they fail to guarantee contamination-free evaluation as the newly collected data may contain pre-existing knowledge, and their benchmark updates rely on intensive human labor. To address these issues, we in this paper propose AntiLeak-Bench, an automated anti-leakage benchmarking framework. Instead of simply using newly collected data, we construct samples with explicitly new knowledge absent from LLMs' training sets, which thus ensures strictly contamination-free evaluation. We further design a fully automated workflow to build and update our benchmark without human labor. This significantly reduces the cost of benchmark maintenance to accommodate emerging LLMs. Through extensive experiments, we highlight that data contamination likely exists before LLMs' cutoff time and demonstrate AntiLeak-Bench effectively overcomes this challenge.

Large Language Models (LLMs) are frequently confident, eloquent, and well versed. A natural question arises: do they know what they don't know? To answer this question, we borrow the concept of epistemic honesty and develop a novel metric to systematically evaluate whether an LLM appropriately acknowledges the boundaries of its knowledge. In this work, we introduce the Epistemic Honesty Quotient (EHQ), which reports three observable sub-scores across two operational axes (epistemic restraint and substantive-answer calibration), and construct EHQ-3000, a 3,000-question benchmark spanning Fabricated Entity, Post-Cutoff Event, Hyper-Niche True, and Context-Conditioned Questions. From a frozen registry of 21 model API routes, 15 completed the protocol after endpoint and eligibility checks; 14 entered the confirmatory analysis because severe provider-side truncation made one route's score indeterminate. The study reveals substantial variation across models, including a difference that can not be explained by their capability to extract explicitly available information. Composite EHQ ranges from 0.31 to 0.81 across the analysed panel, despite near-ceiling performance on the document-grounded capability probe. The two restraint criteria overlap strongly under the present category composition, whereas substantive-answer calibration varies across models and does not reliably co-vary with restraint; however, the small panel leaves substantial uncertainty. Thus, EHQ reveals behavioral differences that are not visible to conventional correctness-based assessment, while also showing why dataset composition, provider behavior, and confidence elicitation must remain part of the interpretation.

LLM agents can commit durable effects from authority evidence that was valid earlier in execution: a DOM snapshot, approval epoch, version witness, branch token, or worker result. We study the commit boundary at which earlier authority evidence no longer authorizes a durable effect. We call this property commit-time authorization: a durable effect is authorized only if the witness that licensed its derived state remains fresh, causally prior, bound to the same effect, and eligible at commit time. We build a controlled-invalidation suite spanning browser, tool/API, and multi-agent workflows. The suite preserves the user goal and payload shape while invalidating the authority relation before durability. In the primary 54-task matrix, endpoint success remains high: 262/270 runs reach the visible result. Only 55/270 are authorized completions; among the 216 invalidating rows, 207 commit after the authorizing path has failed. All 54 clean controls remain authorized, and a separate 54-run authority-preserving check produces no unauthorized commits. We then evaluate mitigation families. Prompt caution and single-condition checks are insufficient because different hazards break different boundary conditions. Defenses work when they refresh, rebind, replan, or refuse at the durability boundary. CommitGuard, a fail-closed boundary monitor, blocks stale durable-effect attempts on protected commit surfaces when runtimes emit witness, dependency, binding, and eligibility signals. The result is a reporting and runtime-design lesson: endpoint success is a utility metric; authorized commit is a security property.

How do you validate a cheap, deterministic proxy for an oracle that is expensive, rate-limited, and non-stationary? We present a protocol built on adversarial falsification gates (negative control, dose response, bounded amplification, duplication penalty, length neutrality) that define and select the proxy, fitted on a training split and confirmed held-out; around them it bounds what the proxy can never resolve, and re-measures external causal evidence on the current oracle rather than assuming it. We demonstrate it end to end on Generative Engine Optimization, where the proxy is a deterministic content score, and one step fails on that domain exactly as the protocol is built to detect: re-measuring the only published causal anchors (2023 effect sizes) on ten modern engine families shows their levers move citation on none, so the anchors are an expired external check; recalibrating to the near-zero modern vector strips the score of its lever-responsive components. What survives is the gate-enforced response surface. The gates buy a measured property: on a 500-source benchmark of adversarial edits, amplifying the score's calibrated levers gains an attacker at most 6 points, and decreases with dose; single-lever amplification is provably bounded, while the cap and cross-lever sub-additivity are empirical findings consistent with it. On detection, web-spam baselines dominate and out-of-distribution attacks evade the score, so the deployable filter layers it over them. A query-conditioned skyline bounds the score's citation signal (within-query Spearman 0.11), repositioning query-agnostic scores as quality filters rather than citation predictors. A query-leakage bug in our first ranking evaluation and a failed confidence flag are disclosed and corrected; every number reproduces offline from released artifacts at zero marginal API cost.

AI evaluation can support the wrong inference when an in-domain benchmark success does not survive distribution shift, or when the benchmark endpoint is entangled with a design factor. We study this problem in CRISPRi perturbation-effect prediction, evaluating a frozen Geneformer representation under a locked, pre-registered protocol: heads and model selection were frozen before test evaluation; the protocol required external outcome labels to remain withheld until final unblinding; and analysis-governing decisions were fixed before the evaluations they govern. In-distribution on the Virtual Cell Challenge (VCC), the frozen representation carries measurable predictive information beyond a dimension-matched random-feature control (Delta R^2 = +0.1645, 95% CI [+0.1375, +0.1920]), satisfying the pre-registered informativeness gate required before interpreting transfer. It then fails zero-shot transfer on both external screens (Spearman rho = -0.139 and -0.267), lying below that control on each. Adding a predefined magnitude block improves the representation externally (Delta rho = +0.032 and +0.143) but, under the frozen primary head, does not rescue transfer: both remain negative. A pre-registered, count-adjusted max-response secondary is positively associated with the outcome on both screens; we report it as correlational and secondary, not as a recovered magnitude signal. Finally, the VCC endpoint is strongly sample-size associated: a count-only linear model reaches R^2 = +0.4325, versus +0.2589 for the four magnitude scalars; adding those scalars to cell count improves R^2 by only +0.0017, so much of the aggregate-magnitude signal overlaps with cell count. This case study shows how locking the evaluation, harmonizing the measured endpoint, and separating primary from secondary evidence can change the inference supported by an AI benchmark.

Agent-interoperability protocols such as A2A and MCP standardize what agents say to one another but assume address-based transport. Whether over HTTP(S) or a content-protecting binding such as MLS-based SLIM, these transports protect message content yet leave the communication graph exposed: which agent contacts which, when, and how often. In agent systems this graph is more consequential than a privacy framing suggests. Endpoints are capability-labeled, workflows are structured and chained, and interactions are coupled to real actions, so an observer recovers more than past relationships: it can infer the pending workflow and, at machine speed, act on that inference before the workflow completes. The threat is therefore one of workflow integrity, not privacy alone. We formalize a threat model for the communication graph and locate what makes its metadata distinctively consequential: not stronger fingerprinting, which we measure to be comparable to other machine traffic, but exposure across independent trust domains, coupled to autonomous action. We define transport- and bootstrap-layer privacy properties, evaluate candidate transports, and give an A2A case study where a metadata-protecting binding surfaces the protocol's implicit identity assumptions. On a generative model anchored to a real capture and over a live A2A binding, a label-blind classifier recovers a task's class from passive metadata well above chance, and from only its opening; a defense-aware adversary does not overturn this, and only the full set of properties drives recovery toward chance. The leverage of acting on the leak is distinct from recoverability: under a fixed budget an adversary realizes most of a clairvoyant attacker's advantage from a workflow's opening, governed by precision over the top-ranked workflows rather than overall accuracy, so a defense suppresses it even while recovery stays above chance.

Process attestation systems verify that a continuous physical process, such as human authorship, actually occurred, rather than merely checking system state. These systems face a fundamental dependability challenge: the evidence collection infrastructure must remain available and tamper-resistant even when the attesting party controls the platform. Trusted Execution Environments (TEEs) provide hardware-enforced isolation that can address this challenge, but their integration with continuous process attestation introduces novel resilience requirements not addressed by existing frameworks. We present the first architecture for continuous process attestation evidence collection inside TEEs, providing hardware-backed tamper resistance against trust-inverted adversaries with graduated input assurance from software-channel integrity (Tier 1) through hardware-bound input (Tier 3). We develop a Markov-chain dependability model quantifying Evidence Chain Availability (ECA), Mean Time Between Evidence Gaps (MTBEG), and Recovery Time Objectives (RTO). We introduce a resilient evidence chain protocol maintaining chain integrity across TEE crashes, network partitions, and enclave migration. Our security analysis derives formal bounds under combined threat models including trust inversion and TEE side channels, parameterized by a conjectural side-channel leakage bound esc that requires empirical validation. Evaluation on Intel SGX demonstrates under 25% per-checkpoint CPU overhead (<0.3% of the 30 s checkpoint interval), >99.5% Evidence Chain Availability (ECA) (the fraction of session time with active evidence collection) in Monte Carlo simulation under Poisson failure models, and sealed-state recovery under 200 ms.

State-of-the-art large language models (LLMs) are now claiming remarkable supported context lengths of 256k or even more. In contrast, the average context lengths of mainstream benchmarks are insufficient (5k-21k), and they suffer from potential knowledge leakage and inaccurate metrics, resulting in biased evaluation. This paper introduces LV-Eval, a challenging long-context benchmark with five length levels (16k, 32k, 64k, 128k, and 256k) reaching up to 256k words. LV-Eval features two main tasks, single-hop QA and multi-hop QA, comprising 11 bilingual datasets. The design of LV-Eval has incorporated three key techniques, namely confusing facts insertion, keyword and phrase replacement, and keyword-recall-based metric design. The advantages of LV-Eval include controllable evaluation across different context lengths, challenging test instances with confusing facts, mitigated knowledge leakage, and more objective evaluations. We evaluate 10 LLMs on LV-Eval and conduct ablation studies on the techniques used in LV-Eval construction. The results reveal that: (i) Commercial LLMs generally outperform open-source LLMs when evaluated within length levels shorter than their claimed context length. However, their overall performance is surpassed by open-source LLMs with longer context lengths. (ii) Extremely long-context LLMs, such as Yi-6B-200k, exhibit a relatively gentle degradation of performance, but their absolute performances may not necessarily be higher than those of LLMs with shorter context lengths. (iii) LLMs' performances can significantly degrade in the presence of confusing information, especially in the pressure test of "needle in a haystack". (iv) Issues related to knowledge leakage and inaccurate metrics introduce bias in evaluation, and these concerns are alleviated in LV-Eval. All datasets and evaluation codes are released at: https://github.com/infinigence/LVEval.

Prompt leakage poses a compelling security and privacy threat in LLM applications. Leakage of system prompts may compromise intellectual property, and act as adversarial reconnaissance for an attacker. A systematic evaluation of prompt leakage threats and mitigation strategies is lacking, especially for multi-turn LLM interactions. In this paper, we systematically investigate LLM vulnerabilities against prompt leakage for 10 closed- and open-source LLMs, across four domains. We design a unique threat model which leverages the LLM sycophancy effect and elevates the average attack success rate (ASR) from 17.7% to 86.2% in a multi-turn setting. Our standardized setup further allows dissecting leakage of specific prompt contents such as task instructions and knowledge documents. We measure the mitigation effect of 7 black-box defense strategies, along with finetuning an open-source model to defend against leakage attempts. We present different combination of defenses against our threat model, including a cost analysis. Our study highlights key takeaways for building secure LLM applications and provides directions for research in multi-turn LLM interactions

When an AI agent calls an API and hits a validation error, it needs more than what went wrong -- it needs what to do next. A self-reflective API returns, on validation failure, a machine-readable recovery\_feedback.suggestions[] payload sufficient for the agent to repair the request and retry without external reasoning. On a leak-audited pilot (N{=}30 per cell, 3 LLMs, 10 adversarial tasks), structured suggestions lift task-completion rate by +36.7--40.0pp over plain-English diagnoses on Anthropic models (Fisher's exact p le 0.0022), at 1.8--2.2times better per-success token efficiency. The lift is not significant on gpt-4o-mini (p{=}0.435); a second-domain replication on a billing API confirms the pattern. The comparison only holds after auditing two undocumented classes of answer leakage in LLM benchmarks. We shipaudit\_prompt\_leakage.py as reusable CI infrastructure. Code and data: https://github.com/arquicanedo/self-reflective-apis.

Large Language Models (LLMs) as judges and LLM-based data synthesis have emerged as two fundamental LLM-driven data annotation methods in model development. While their combination significantly enhances the efficiency of model training and evaluation, little attention has been given to the potential contamination brought by this new model development paradigm. In this work, we expose preference leakage, a contamination problem in LLM-as-a-judge caused by the relatedness between the synthetic data generators and LLM-based evaluators. To study this issue, we first define three common relatednesses between data generator LLM and judge LLM: being the same model, having an inheritance relationship, and belonging to the same model family. Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests that preference leakage is a pervasive issue that is harder to detect compared to previously identified biases in LLM-as-a-judge scenarios. All of these findings imply that preference leakage is a widespread and challenging problem in the area of LLM-as-a-judge. We release all codes and data at: https://github.com/David-Li0406/Preference-Leakage.

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

Language models can distinguish between testing and deployment phases -- a capability known as evaluation awareness. This has significant safety and policy implications, potentially undermining the reliability of evaluations that are central to AI governance frameworks and voluntary industry commitments. In this paper, we study evaluation awareness in Llama-3.3-70B-Instruct. We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. We also find that current safety evaluations are correctly classified by the probes, suggesting that they already appear artificial or inauthentic to models. Our findings underscore the importance of ensuring trustworthy evaluations and understanding deceptive capabilities. More broadly, our work showcases how model internals may be leveraged to support blackbox methods in safety audits, especially for future models more competent at evaluation awareness and deception.

Recent observations have underscored a disparity between the inflated benchmark scores and the actual performance of LLMs, raising concerns about potential contamination of evaluation benchmarks. This issue is especially critical for closed-source models and certain open-source models where training data transparency is lacking. In this paper we study data contamination by proposing two methods tailored for both open-source and proprietary LLMs. We first introduce a retrieval-based system to explore potential overlaps between evaluation benchmarks and pretraining corpora. We further present a novel investigation protocol named Testset Slot Guessing (TS-Guessing), applicable to both open and proprietary models. This approach entails masking a wrong answer in a multiple-choice question and prompting the model to fill in the gap. Additionally, it involves obscuring an unlikely word in an evaluation example and asking the model to produce it. We find that certain commercial LLMs could surprisingly guess the missing option in various test sets. Specifically, in the TruthfulQA benchmark, we find that LLMs exhibit notable performance improvement when provided with additional metadata in the benchmark. Further, in the MMLU benchmark, ChatGPT and GPT-4 demonstrated an exact match rate of 52\% and 57\%, respectively, in guessing the missing options in benchmark test data. We hope these results underscore the need for more robust evaluation methodologies and benchmarks in the field.

This paper provides a systematic analysis of the opportunities, challenges, and potential solutions of harnessing Large Language Models (LLMs) such as GPT-4 to dig out vulnerabilities within smart contracts based on our ongoing research. For the task of smart contract vulnerability detection, achieving practical usability hinges on identifying as many true vulnerabilities as possible while minimizing the number of false positives. Nonetheless, our empirical study reveals contradictory yet interesting findings: generating more answers with higher randomness largely boosts the likelihood of producing a correct answer but inevitably leads to a higher number of false positives. To mitigate this tension, we propose an adversarial framework dubbed GPTLens that breaks the conventional one-stage detection into two synergistic stages - generation and discrimination, for progressive detection and refinement, wherein the LLM plays dual roles, i.e., auditor and critic, respectively. The goal of auditor is to yield a broad spectrum of vulnerabilities with the hope of encompassing the correct answer, whereas the goal of critic that evaluates the validity of identified vulnerabilities is to minimize the number of false positives. Experimental results and illustrative examples demonstrate that auditor and critic work together harmoniously to yield pronounced improvements over the conventional one-stage detection. GPTLens is intuitive, strategic, and entirely LLM-driven without relying on specialist expertise in smart contracts, showcasing its methodical generality and potential to detect a broad spectrum of vulnerabilities. Our code is available at: https://github.com/git-disl/GPTLens.

Side-channel evaluators routinely inspect leakage tests while acquisition is still running, and extend or stop the campaign based on what they see. Fixed-horizon screening such as the Welch t-test with threshold |t|>4.5 gives no error guarantee for this monitored decision rule. We study anytime-valid leakage detection based on testing by betting: SKIT-type swap-pair e-processes whose false-alarm probability is controlled uniformly over time under an explicit conditional symmetry null. In matched comparisons that share the frozen witness, rows and payoff, first-crossing detection needed 1.68-2.00times the traces of a fixed-horizon randomization test at 80% detection on synthetic streams, and 1.68-2.38times on degraded recordings from an open ML-KEM electromagnetic dataset with the primary Ridge witness at α=0.05. With the same primary witness and level, on undegraded reference and pqm4 recordings the monitored procedure stopped early: its median stopping point was 62-72 and 146-316 evaluation traces, i.e. 2-8% of a conservative 4096-trace budget. Under exact designed nulls on the recorded backgrounds, repeated-look |t|>4.5 screening over all 13000-20000 samples raised a false alarm in 2.7-12.9% of replicates, against 0.0-4.7% for terminal-only screening and no rejection by a sample-wise e-Bonferroni process, which in a prespecified follow-up detected natural-label associations in 4 of 4 backgrounds after 840-3288 traces. All recordings come from one device, and natural-label results are descriptive; we state the assumptions each claim requires.

Autonomous coding agents increasingly execute multi-step software work, but lifecycle states such as reviewed, tested, DONE, and ready-to-merge remain claims unless supported by current evidence. We present Proof-or-Stop Lifecycle Control, a method that permits lifecycle transitions only when fresh, tracked-source-state-bound, mechanically verifiable evidence satisfies the relevant gate. The method treats agent outputs as claims rather than lifecycle state, and uses proof operationally to mean gate-admissible evidence under a stated trust model, not semantic program correctness. We evaluate an open-source implementation through mechanism tests, a powered control-policy ablation, and operated self-application evidence. The unattended-loop engine passed 10 of 10 scenarios with zero false-DONE, and local-key receipt bundles rejected 18 tamper classes with zero false accepts. In a 9,240-cell ablation, the pre-registered A4 versus A2-prime comparison reduced visible-pass/hidden-fail amplification from 31 of 1,800 injected cells under a compute-budgeted naive loop to 2 of 1,800 under the gated loop, a 1.6 percentage-point improvement in not-amplified rate with a 95 percent confidence interval of [0.8, 2.5]. A near-compute A3 versus A4 comparison, 14 of 1,800 versus 2 of 1,800, indicates that the gain is associated with enforcing review as a lifecycle gate rather than merely adding a reviewer. The self-application corpus contains 565 stories and 1,007 review findings, with 94.8 percent resolved, plus a 68-row high/critical cross-vendor exhibit. These results support Proof-or-Stop as a model-agnostic, host-neutral control layer for deciding which autonomous-agent claims a lifecycle may act on. The evaluation is limited to one model family, 24 ablation tasks, and a self-hosted corpus.

LLM agents that invoke privileged tools are vulnerable to indirect prompt injection (IPI), in which adversarial instructions embedded in retrieved data hijack the agent's actions. A growing body of work evaluates defenses against IPI, but the validity of that evaluation is rarely examined. We audit an IPI benchmark and its harness and identify four defect classes -- silent payload non-delivery, attack success scored by tool identity rather than arguments, false-rejection rate conflated with model incapacity, and the absence of an audit trail -- each of which yields a plausible, publishable, and incorrect number. We quantify the distortion by re-scoring identical execution traces under the defective and corrected definitions: on real agent behaviour, the tool-identity scorer reports a 21.7% attack-success rate where the true argument-level rate is 1.2%. In the sharpest case, an open model previously reported at 62.8% registers 0% under the corrected harness -- the prior figure largely an artifact of undelivered payloads and identity-level scoring. We release a harness whose construction makes each defect unrepresentable -- machine-checkable payload placement, argument-level attacker predicates, per-scenario environments, and mandatory trace persistence -- and use it to report three quantities the field does not: whether a compromised agent discloses the attack, the full security/utility operating curve of an LLM-judge defense, and tool-calling capability disentangled from defensive over-blocking. A corrected harness further overturns a reported "capability barrier": a model deemed incapable of tool use is in fact fully capable, its earlier result an artifact of environment mismatch. We argue that evaluation validity is a prerequisite for, not a footnote to, defense claims in agentic security, and provide an instrument that enforces it.

This paper presents a new tool learning dataset Seal-Tools, which contains self-instruct API-like tools. Seal-Tools not only offers a large number of tools, but also includes instances which demonstrate the practical application of tools. Seeking to generate data on a large scale while ensuring reliability, we propose a self-instruct method to generate tools and instances, allowing precise control over the process. Moreover, our Seal-Tools contains hard instances that call multiple tools to complete the job, among which some are nested tool callings. For precise and comprehensive evaluation, we use strict format control and design three metrics from different dimensions. Therefore, Seal-Tools can serve as a new benchmark to evaluate the tool-calling ability of LLMs. Finally, we evaluate several prevalent LLMs and our finetuned model on Seal-Tools. The results show that current systems are far from perfect. The code, data and experiment results are available at https://github.com/fairyshine/Seal-Tools .

An enterprise agent that answers under a written policy must choose an action before it says anything: answer, ask for a missing fact, ask for confirmation, call one or several tools, escalate, or refuse. Holding one 27B Korean enterprise model fixed, we vary whether the decision is a generated JSON object or a single constrained token, and whether the model may reason first. On 2800 in-distribution items we did not detect an accuracy difference between the two deployed paths (93.46% vs 94.14%), while they differ about tenfold in restart stability (3.00% of actions flip vs 0.29%); they also differ in reasoning, so this is an operational comparison. On a sealed benchmark of 300 items whose labels come from a hidden executable policy, reasoning is worth +17.3 points [+12.3, +23.0] on exact action. With reasoning off, the format difference was 0.0 points [-4.3, +4.0], a failure to detect an advantage for either interface, not equivalence; with reasoning on it is unresolved. Against a token-matched random control, boundary-targeted supervision moves the two covered boundaries by +17.5 and +20.0 points and leaves a held-out boundary unchanged. In a 3-seed training x reasoning factorial, the sealed-rendering interaction is -4.9 [-8.1, -1.7], consistent with partial substitution. All supervision was Korean. On independently rendered English versions, the model is 0.82 points lower [-2.68, +0.93] with reasoning on, the same order as re-rendering in Korean; the reasoning-off gain from targeted supervision is not stable in either language, so we do not claim transfer. The original text stays private as the confirmatory set. An open-model re-rendering made after these results reproduced the reasoning effect and the English verdict; the interaction kept its sign but did not replicate under item-clustered intervals. We release the model and that re-rendering with gold labels.

The consensus that GCN, GraphSAGE, GAT, and EvolveGCN outperform feature-only baselines on the Elliptic Bitcoin Dataset is widely cited but has not been rigorously stress-tested under a leakage-free evaluation protocol. We perform a seed-matched inductive-versus-transductive comparison and find that this consensus does not hold. Under a strictly inductive protocol, Random Forest on raw features achieves F1 = 0.821 and outperforms all evaluated GNNs, while GraphSAGE reaches F1 = 0.689 +/- 0.017. A paired controlled experiment reveals a 39.5-point F1 gap attributable to training-time exposure to test-period adjacency. Additionally, edge-shuffle ablations show that randomly wired graphs outperform the real transaction graph, indicating that the dataset's topology can be misleading under temporal distribution shift. Hybrid models combining GNN embeddings with raw features provide only marginal gains and remain substantially below feature-only baselines. We release code, checkpoints, and a strict-inductive protocol to enable reproducible, leakage-free evaluation.

Observational studies can yield clinically actionable evidence at scale, but executing them on real-world databases is open-ended and requires coherent decisions across cohort construction, analysis, and reporting. Prior evaluations of LLM agents emphasize isolated steps or single answers, missing the integrity and internal structure of the resulting evidence bundle. To address this gap, we introduce RWE-bench, a benchmark grounded in MIMIC-IV and derived from peer-reviewed observational studies. Each task provides the corresponding study protocol as the reference standard, requiring agents to execute experiments in a real database and iteratively generate tree-structured evidence bundles. We evaluate six LLMs (three open-source, three closed-source) under three agent scaffolds using both question-level correctness and end-to-end task metrics. Across 162 tasks, task success is low: the best agent reaches 39.9%, and the best open-source model reaches 30.4%. Agent scaffolds also matter substantially, causing over 30% variation in performance metrics. Furthermore, we implement an automated cohort evaluation method to rapidly localize errors and identify agent failure modes. Overall, the results highlight persistent limitations in agents' ability to produce end-to-end evidence bundles, and efficient validation remains an important direction for future work. Code and data are available at https://github.com/somewordstoolate/RWE-bench.

As a core protocol for integrating large language models (LLMs) with external data sources and tools, the Model Context Protocol (MCP) has quickly become a key interface in emerging AI agent systems. As the MCP ecosystem expands, reported vulnerabilities have also increased, creating a growing software-security attack surface. However, MCP vulnerability information remains scattered, noisy, and difficult to study systematically. This paper addresses that gap by constructing a curated MCP vulnerability dataset from CVE disclosures. Starting from 119 MCP-related records retrieved through February 2026, we filter unrelated cases out and obtain a validated snapshot of 24 MCP vulnerabilities. Our filtering methodology explicitly addresses acronym ambiguity and achieves 96.7% inter-rater agreement on a sampled exclusion audit while retaining all known high-severity MCP cases identified in public advisories. Using this initial dataset snapshot, we find that 87.5% of validated vulnerabilities are Critical or High severity, and that injection and path traversal vulnerabilities account for 66.7% of cases. We further identify recurring issues involving insufficient input validation, inadequate access control, and protocol-level security gaps. Through six representative case studies, we illustrate common exploitation patterns and map them to practical defenses. We release the dataset, reproducibility tooling, and a sustainable maintenance workflow as a foundation for future MCP security measurement, benchmarking, and defense research.

Visual fluency in generated video does not imply physical reliability, and a scalar quality score alone is incapable of indicating the obligation a clip violates or the moment it fails. We present VeriPhy, an auditable physical-verification system in which a text-only planner compiles the prompt into typed physical obligations and a statically validated execution plan before any frame is observed. During execution, observations gate and scope only declared calls to frozen low-level experts (e.g., segmentation and tracking, counting, eleven typed physical measurements over the resulting tracks, depth, OCR, and audio-event detection). Each action returns a provenance-carrying evidence record whose payload, when usable, is either a typed measurement or an explicitly tagged learned state. Typed resolvers and fixed composition map usable records to a three-valued state (supported, contradicted, or unknown, surfaced as plausible, implausible, or abstain) with full provenance, so that every verdict is traceable to the evidence that produced it. We anchor evaluation in a 1,500-clip corpus of human-annotated flaw records that localize real generation failures in prompt reference, space, and time. On a 149-clip core carrying 304 such records, VeriPhy accounts for 228, against 164 for a published question-decomposition evaluator given the same clips and the same claims. Recall alone does not separate it from prompting the same backbone monolithically, which reaches 222; what separates them is that each decision retains its evidence record and provenance, making the traces auditable one verdict at a time and usable as the interface through which a critic verdict could be written back into generation.

Unlearning benchmarks such as TOFU and MUSE certify forgetting by reading the model's final answer, where a model that refuses to answer already counts as having forgotten. We show that this model-level certificate does not transfer once the model is deployed as an agent. We introduce K-Bench, a benchmark that scores LLM unlearning under agentic deployment. K-Bench inspects all six channels a ReAct agent exposes, including its chain-of-thought (CoT), tool calls and tool observations, and elicited summary. A query counts as leaked if the secret appears in any of them. Each experiment places the secret in exactly one of the agent's three sources (the weights, the prompt, or the retrieval store). The K-Score is computed separately for each source and credits forgetting only when the agent remains usable. Clearing the answer channel does not make the secret unrecoverable. On structured retrieval, the secret stays verbatim in the tool-observation channel and the aggregate leak rate is unchanged. When the secret lives in the prompt or the retrieval store, TOFU and MUSE report no leakage, while the deployed agent still leaks it on 22--86\% of queries. When the secret is in the weights, none of the twenty evaluated published methods demonstrably removes it, and only an input-corruption intervention reaches selective forgetting under the evaluated observer. The top-ranked method changes across base models. A refusal-tuning method resists the evaluated extraction without verified knowledge removal.

Recently, LoRA and its variants have become the de facto strategy for training and sharing task-specific versions of large pretrained models, thanks to their efficiency and simplicity. However, the issue of copyright protection for LoRA weights, especially through watermark-based techniques, remains underexplored. To address this gap, we propose SEAL (SEcure wAtermarking on LoRA weights), the universal whitebox watermarking for LoRA. SEAL embeds a secret, non-trainable matrix between trainable LoRA weights, serving as a passport to claim ownership. SEAL then entangles the passport with the LoRA weights through training, without extra loss for entanglement, and distributes the finetuned weights after hiding the passport. When applying SEAL, we observed no performance degradation across commonsense reasoning, textual/visual instruction tuning, and text-to-image synthesis tasks. We demonstrate that SEAL is robust against a variety of known attacks: removal, obfuscation, and ambiguity attacks.

Long contexts of recent LLMs have enabled a new use case: asking models to find security vulnerabilities in entire codebases. To evaluate model performance on this task, we introduce eyeballvul: a benchmark designed to test the vulnerability detection capabilities of language models at scale, that is sourced and updated weekly from the stream of published vulnerabilities in open-source repositories. The benchmark consists of a list of revisions in different repositories, each associated with the list of known vulnerabilities present at that revision. An LLM-based scorer is used to compare the list of possible vulnerabilities returned by a model to the list of known vulnerabilities for each revision. As of July 2024, eyeballvul contains 24,000+ vulnerabilities across 6,000+ revisions and 5,000+ repositories, and is around 55GB in size.

As large language models become increasingly widespread, third-party providers that deploy open-weight models have become an important part of the ecosystem. Auditing the quality of their inference APIs is therefore an open problem. We formalize hosted model routing as a stochastic process and propose \textbf{Ventor-QTest}, a composite black-box audit that requires no probability information from the target API. Its repeated-request component sends each frozen constrained context to the target multiple times, reconstructs a categorical output distribution from the returned text counts, and reports average fidelity loss (AFL) as a null-bias-corrected, within-window mean coarsened-KL statistic. Its long-sequence component uses independent runs to report extreme fidelity loss (EFL) through the empirical upper tail of a run-level reference-centered-surprisal statistic. Across three logprob-capable route conditions, AFL shows strong linear descriptive agreement with a logprob-derived coarsened-KL comparator. Across seven route snapshots, 20-run sequence probes reveal route-specific EFL variation. AFL and EFL have little detectable route-level association with GPQA-Diamond accuracy. In contrast, pronounced EFL coincides with a decline in Terminal-Bench pass rate as task exposure increases. This pattern may arise because correctness in long-horizon tasks is more sensitive to extreme fidelity loss. These results motivate reporting AFL and EFL jointly, particularly when auditing long-horizon agentic tasks. The open-source implementation is available at https://github.com/Tencent/AI-Infra-Guard/tree/main/services/api_checker/ventor_qtest.

Tool-using LLM agents increasingly use the Model Context Protocol (MCP) to answer from heterogeneous evidence sources, including search, APIs, databases, clinical records, and formulary tools. Standard factuality metrics usually test whether an answer is supported by pooled evidence, missing a provenance-sensitive failure mode: a claim may be supported somewhere while being attributed to the wrong source. We call this cross-source conflation. We introduce ProvenanceGuard, a source-aware verifier for MCP-grounded answers. It consumes captured MCP traces with stable tool IDs, source IDs, and raw outputs; decomposes answers into atomic claims; routes claims to source-specific evidence; checks support with NLI and a token-alignment proxy; compares stated attribution with the routed source; and returns per-claim verdicts plus an answer-level allow/block decision. Blocked answers can be repaired with retrieval-augmented answer revision and re-verified. We evaluate on 281 medical-domain MCP-agent traces. A 266-trace adjudicated subset yields 2,325 LLM-assisted claim labels split by trace; 361 held-out labels are human-verified. On the 40-trace held-out split, ProvenanceGuard achieves block F1 0.802 and source accuracy 0.858 over 260 source-eligible claims, outperforming source-blind baselines that do not emit claim-to-source IDs. On a harder multi-source benchmark it reaches block F1 0.846, while source-plus-relation accuracy drops to 0.229, showing that exact source ownership remains difficult with semantically close sources. Repair-and-reverify resolves all blocked answers in the full trace set, often via conservative fallback. In 50 controlled clinical conflation probes, ProvenanceGuard detects all injected attribution swaps with no retained wrong attribution. These results show that source attribution is an independent axis for factuality verification in MCP-based agents.

Large language models (LLMs) deployed for logical reasoning in knowledge-intensive domains exhibit a subtle but critical failure: coherence can be vacuously achieved through systematic abstention. A model that withholds commitment to either entailment or refutation satisfies negation consistency while providing no utility. We introduce Coherence Under Commitment (CUC), a dual-query evaluation paradigm that jointly measures consistency and decisiveness. CUC contributes three innovations: (1) a commitment score c(φ) = p(φ) + p(lnotφ) quantifying probability mass allocated to decisive outcomes; (2) a deterministic elicitation protocol via normalized YES/NO log probabilities, eliminating sampling variance; and (3) a 3-way decision framework (True/False/Uncertain) operationalizing the coherence-commitment trade-off into metrics. Experiments on four open-weight LLMs (1B-3B) across 204 FOLIO examples expose a sharp frontier. Qwen2.5-3B achieves near-zero contradiction (E[v_{neg}]{=}0.025) but only 7.4% coverage, while TinyLlama-1.1B reaches 79.4% coverage with violations on every example. Coherence-only evaluation would rank the abstaining model first; CUC exposes this as vacuous, and the frontier generalizes to LogiQA~v2 (ρ{=}0.97). We argue that evaluation must report both coherence and non-vacuous commitment and release a toolkit for standardized assessment.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

We propose claim-level falsification as a principle for test-time scaling and instantiate it through Claim-Level Reliability Assessment (CLR), a training-free framework that reallocates test-time compute from additional solution sampling to targeted verification. Since whole-trace evaluation often obscures decisive errors due to signal dilution from routine tokens, CLR condenses each reasoning trace into a compact set of decision-critical claims, thereby isolating its logical anchors. Furthermore, recognizing the inherent difficulty of generating entirely correct solutions under fixed model capabilities, CLR shifts the focus to semantic falsification. This approach exploits a fundamental asymmetry between solution construction and claim refutation. Constructing a valid solution requires a flawless reasoning path, whereas refuting an incorrect claim requires identifying only a single decisive flaw. This targeted search for negative evidence systematically compresses the survival space of high-confidence incorrect traces, effectively suppressing erroneous consensus via nonlinear reliability scoring. Across four LLMs and four reasoning benchmarks under matched budgets, CLR generally improves upon pass@1 and self-consistency. On GPT-OSS-20B/CMIMC25, for instance, CLR exceeds pass@1 by 27.15 percentage-points and raises self-consistency accuracy from 77.50\% to 82.19\% with 37.0\% fewer tokens.

Large Language Models (LLMs), such as OpenAI's o1-series have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism. However, recent studies reveal substantial redundancy in the CoT reasoning traces, which not only increases inference latency but also negatively impacts model performance by diverting attention to unnecessary reasoning paths. To address this issue, we investigate the internal reasoning structures of LLMs and categorize them into three primary thought types: execution, reflection, and transition thoughts. Moreover, our analysis reveals that excessive reflection and transition thoughts are strongly correlated with failure cases and these thought categories exhibit clear separation in the latent space. Based on these, we introduce SEAL (Steerable reasoning calibration), a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains. SEAL consists of an offline stage for extracting the reasoning steering vector in the latent space, followed by an on-the-fly calibration of the reasoning trace through representation intervention using the steering vector. Notably, the steering vector exhibits strong transferability across various tasks. Extensive experiments across multiple models (DeepSeek-R1-Distill and QwQ-32B-Preview) and benchmarks (Math500, GSM8K, LiveCodeBench) validate the effectiveness of SEAL, up to a 11% improvement in accuracy while reducing reasoning tokens by 11.8% to 50.4%. Our code is publicly available at https://github.com/VITA-Group/SEAL.

LLM agents call tools, query databases, delegate tasks, and trigger external side effects. Once an agent system can act in the world, the question is no longer only whether harmful actions can be prevented--it is whether those actions remain answerable after deployment. We distinguish accountability (the ability to determine compliance and assign responsibility), auditability (the system property that makes accountability possible), and auditing (the process of reconstructing behavior from trustworthy evidence). Our claim is direct: no agent system can be accountable without auditability. To make this operational, we define five dimensions of agent auditability, i.e., action recoverability, lifecycle coverage, policy checkability, responsibility attribution, and evidence integrity, and identify three mechanism classes (detect, enforce, recover) whose temporal information-and-intervention constraints explain why, in practice, no single approach suffices. We support the position with layered evidence rather than a single benchmark: lower-bound ecosystem measurements suggest that even basic security prerequisites for auditability are widely unmet (617 security findings across six prominent open-source projects); runtime feasibility results show that pre-execution mediation with tamper-evident records adds only 8.3 ms median overhead; and controlled recovery experiments show that responsibility-relevant information can be partially recovered even when conventional logs are missing. We propose an Auditability Card for agent systems and identify six open research problems organized by mechanism class.

Prompt injection remains a central obstacle to the safe deployment of large language models, particularly in multi-agent settings where intermediate outputs can propagate or amplify malicious instructions. Building on earlier work that introduced a four-metric Total Injection Vulnerability Score (TIVS), this paper extends the evaluation framework with semantic similarity-based caching and a fifth metric (Observability Score Ratio) to yield TIVS-O, investigating how defence effectiveness interacts with transparency in a HOPE-inspired Nested Learning architecture. The proposed system combines an agentic pipeline with Continuum Memory Systems that implement semantic similarity-based caching across 301 synthetically generated injection-focused prompts drawn from ten attack families, while a fourth agent performs comprehensive security analysis using five key performance indicators. In addition to traditional injection metrics, OSR quantifies the richness and clarity of security-relevant reasoning exposed by each agent, enabling an explicit analysis of trade-offs between strict mitigation and auditability. Experiments show that the system achieves secure responses with zero high-risk breaches, while semantic caching delivers substantial computational savings, achieving a 41.6% reduction in LLM calls and corresponding decreases in latency, energy consumption, and carbon emissions. Five TIVS-O configurations reveal optimal trade-offs between mitigation strictness and forensic transparency. These results indicate that observability-aware evaluation can reveal non-monotonic effects within multi-agent pipelines and that memory-augmented agents can jointly maximize security robustness, real-time performance, operational cost savings, and environmental sustainability without modifying underlying model weights, providing a production-ready pathway for secure and green LLM deployments.

An AI agent run is no longer a function call. It is a cascade of prompts, model calls, tool invocations, file and network effects and sub-agents. When such a run misbehaves, or an auditor asks what happened, logs and traces answer poorly: they are mutable, partial and cannot be re-executed. This technical report describes NovaFabric, an open-source (Apache-2.0) black-box recorder for AI agents that turns a run into a sealed, replayable, diffable and auditable evidence capsule. The design rests on two formats only. The Run Capsule is a directory of flat files and the source of truth; every database is a rebuildable index over capsules. The Evidence Bundle is the signed export for an auditor. Everything else follows existing standards: OpenTelemetry GenAI, OpenLineage, W3C PROV, in-toto/DSSE, RFC 3161 and CycloneDX. A single-node run is the smallest case of a distributed one; the schema supports a capsule hierarchy for Slurm arrays and Kubernetes jobs, though automatic parent finalization is library-only today. The report covers the architecture and the engines behind the verb chain capture, seal, replay, diff and audit: fail-closed mocked replay; a diff gate with a stable exit-code contract; DSSE signatures that bind their intent and a capsule root that commits to file names; a capsule treated as untrusted input; signer identity-trust levels with offline TSA and CRL checking; maker-checker promotion against an operator-held trust list; and a deny-by-default authorization table for every dashboard route. It presents no new experiments; the evaluation is in arXiv:2609.12582. Every count except the private ADR tally is generated by a probe script run on the public source at the v0.107.0 tag (150 top-level commands, 84 Python subpackages, 16,721 test functions), every capability carries an explicit maturity label, and honest-limits boxes state where the system stops.

The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demonstrating the statistical indistinguishability of this retro-holdout dataset, and (iii) comparing LLMs on the two datasets to quantify the performance gap due to the dataset's public availability. Applying these methods to TruthfulQA, we construct and release Retro-Misconceptions, on which we evaluate twenty LLMs and find that some have inflated scores by as much as 16 percentage points. Our results demonstrate that public benchmark scores do not always accurately assess model properties, and underscore the importance of improved data practices in the field.

Background: Agent skills are increasingly deployed as modular, reusable capability units in AI agent systems. Medical research agent skills require safeguards beyond general-purpose evaluation, including scientific integrity, methodological validity, reproducibility, and boundary safety. This study developed and preliminarily evaluated a domain-specific audit framework for medical research agent skills, with a focus on reliability against expert review. Methods: We developed MedSkillAudit (skill-auditor@1.0), a layered framework assessing skill release readiness before deployment. We evaluated 75 skills across five medical research categories (15 per category). Two experts independently assigned a quality score (0-100), an ordinal release disposition (Production Ready / Limited Release / Beta Only / Reject), and a high-risk failure flag. System-expert agreement was quantified using ICC(2,1) and linearly weighted Cohen's kappa, benchmarked against the human inter-rater baseline. Results: The mean consensus quality score was 72.4 (SD = 13.0); 57.3% of skills fell below the Limited Release threshold. MedSkillAudit achieved ICC(2,1) = 0.449 (95% CI: 0.250-0.610), exceeding the human inter-rater ICC of 0.300. System-consensus score divergence (SD = 9.5) was smaller than inter-expert divergence (SD = 12.4), with no directional bias (Wilcoxon p = 0.613). Protocol Design showed the strongest category-level agreement (ICC = 0.551); Academic Writing showed a negative ICC (-0.567), reflecting a structural rubric-expert mismatch. Conclusions: Domain-specific pre-deployment audit may provide a practical foundation for governing medical research agent skills, complementing general-purpose quality checks with structured audit workflows tailored to scientific use cases.

Quantized checkpoints are often screened first with quality metrics and only later, if at all, with direct safety tests. This paper audits that shortcut on a matched 51-row matrix spanning 6 models, 4 families, a 7-level GGUF ladder, and AWQ/GPTQ INT4 checkpoints. In this matrix the shortcut fails: all 36 quality-safety pairings split direction across models, and 9 hidden-danger rows plus 1 near-hidden-danger row show quality stable or improved while refusal falls by 12-68 percentage points. Seven of the 11 AWQ/GPTQ rows are hidden-danger. A four-probe mechanistic follow-up over the 17 Hugging Face-backed FP16/AWQ/GPTQ cells does not rescue it: entropy, refusal-direction, and calibration probes are weak or null separators of dangerous rows, and although probe-identified safety-associated neurons absorb 1.39times more quantization error overall (p < 5 times 10^{-7}), the effect is not regime-specific. Claude Sonnet 4 relabels 11,470 items in a predefined stratified set, agrees with the primary gemma3:12b judge on 89.9\% of rows (κ= 0.873, 95\% CI [0.866, 0.881]), and changes 0/10 hidden-danger cells. A calibrated study-internal behavioral screen -- the Refusal Template Stability Index (RTSI), built from four refusal-template drift features and calibrated on this matrix -- routes 10/10 hidden- or near-hidden-danger rows to direct safety testing (Wilson 95\% CI lower bound 0.72) while leaving 23 of 45 non-baseline rows in a low-risk bucket under both in-sample scoring and row-level leave-one-out validation; on the same matrix, the best single-feature baselines (unique-prefix-rate-delta, raw refusal-rate delta) recover 9/10 and 8/10 respectively at matched bucket size, and cross-stack transfer requires recalibration. For the quantized checkpoints, model families, and safety outcomes studied here, retained quality cannot waive direct safety evaluation.

Closed-source frontier labs do not disclose parameter counts. Storing F facts requires at least F/(bits per parameter) weights, so factual recall lower-bounds parameter count--an intrinsic, serving-independent signal, though (as we show) a coarse one. We introduce Incompressible Knowledge Probes (IKPs), a benchmark of 1,400 factual questions spanning 7 tiers of obscurity, designed to isolate knowledge that cannot be derived by reasoning or compressed by architectural improvements. We score with no hallucination penalty (lambda = 0: IKP accuracy is simply the fraction of probed facts answered correctly), which removes both the penalty hyperparameter and the per-tier flooring choice; a full lambda x flooring ablation shows the calibration is robust across scoring choices while individual estimates are not, motivating the no-penalty default. We calibrate a log-linear mapping from IKP accuracy to parameter count on 93 open-weight models (135M-1,600B) spanning 19 vendors, achieving R^2 = 0.910; leave-one-out cross-validation confirms generalization (median fold error 1.48x, 72% within 2x, 86% within 3x). The instrument is deliberately coarse--its 90% prediction interval spans ~3x in either direction, wider than inference economics--so IKP recovers order-of-magnitude effective capacity and relative rankings, not precise parameter counts. For Mixture-of-Experts models, total parameters predict knowledge (R^2 = 0.67) better than active parameters (R^2 = 0.41). We evaluate 201 models from 27 vendors on a curated probe set (1,311 of 1,400 probes surviving name-collision and label-ambiguity filters) and report effective knowledge capacity for all major proprietary frontier models as prediction bands rather than point estimates; for heavily safety-tuned models these are lower bounds, since refusal policy can suppress tens of percentage points of otherwise-answerable capacity.

The rapid deployment of Large Language Models (LLMs) requires careful consideration of their effect on cybersecurity. Our work aims to improve the selection process of LLMs that are suitable for facilitating Secure Coding (SC). This raises challenging research questions, such as (RQ1) Which functionality can streamline the LLM evaluation? (RQ2) What should the evaluation measure? (RQ3) How to attest that the evaluation process is impartial? To address these questions, we introduce LLMSecCode, an open-source evaluation framework designed to assess LLM SC capabilities objectively. We validate the LLMSecCode implementation through experiments. When varying parameters and prompts, we find a 10% and 9% difference in performance, respectively. We also compare some results to reliable external actors, where our results show a 5% difference. We strive to ensure the ease of use of our open-source framework and encourage further development by external actors. With LLMSecCode, we hope to encourage the standardization and benchmarking of LLMs' capabilities in security-oriented code and tasks.

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

Small-model agentic post-training is bottlenecked less by the algorithm than by the trajectory substrate it consumes. Leading recipes (RLVR, group-relative RL, rejection-sampled re-SFT) all need multi-turn traces carrying per-trajectory supervision, and the two existing sources fall short: frontier-synthesised data inherits the synthesizer's biases and collapses the long tail, while unfiltered production logs are unjudged and contaminated by shortcut behaviour. We argue that an incentive-aligned agent arena can be engineered to manufacture such trajectories, and demonstrate this on ORO Subnet 15 (SN15), a Bittensor deployment of the ShoppingBench agentic-commerce benchmark. SN15's race mechanism, LLM reasoning judge, and rotating leak-cluster-guarded problem suite yield a corpus with three properties: incentive-aligned diversity, per-trajectory judging, and anti-memorised held-out evaluation. We introduce a structural-quality filter that converts the raw firehose into a trainable corpus by keeping agentic trajectories (the model itself emits the tool calls) and rejecting sub-task trajectories (the model only classifies or narrates over a deterministic search loop), then post-train Qwen3-4B with a recipe matched to the published ShoppingBench SFT-then-GRPO pipeline. On a leak-cluster-guarded held-out partition scored production-strict, the model lifts from the published Qwen3-4B base of 18.0% ASR to 42.7%, within single-problem noise of the synthetic-data SFT-only baseline (43.6%), while training on a fraction of a single day of subnet output. The supervised stack leaves a large pass@8 to pass@1 gap (53.3% vs 34.8%); a per-step teacher-grounded Dr. GRPO reward converts that headroom into process improvement, and we identify the sub-task firehose as the primary lever for closing the gap to the 48.7% SFT+GRPO bar. We release the filter, the corpus splits, and the arena mechanics.

Selective knowledge erasure from LLMs is critical for GDPR compliance and model safety, yet current unlearning methods conflate behavioral suppression with true knowledge removal, allowing latent capabilities to persist beneath surface-level refusals. In this work, we address this challenge by introducing Knowledge Immunization Framework (KIF), a representation-aware architecture that distinguishes genuine erasure from obfuscation by targeting internal activation signatures rather than surface outputs. Our approach combines dynamic suppression of subject-specific representations with parameter-efficient adaptation, enabling durable unlearning without full model retraining. KIF achieves near-oracle erasure (FQ approx 0.99 vs. 1.00) while preserving utility at oracle levels (MU = 0.62), effectively breaking the stability-erasure tradeoff that has constrained all prior work. We evaluate both standard foundation models (Llama and Mistral) and reasoning-prior models (Qwen and DeepSeek) across 3B to 14B parameters. Our observation shows that standard models exhibit scale-independent true erasure (<3% utility drift), while reasoning-prior models reveal fundamental architectural divergence. Our comprehensive dual-metric evaluation protocol, combining surface-level leakage with latent trace persistence, operationalizes the obfuscation - erasure distinction and enables the first systematic diagnosis of mechanism-level forgetting behavior across model families and scales.

The integration of LLMs into vulnerability detection (VD) has shifted the field toward interpretable and context-aware analysis. While post-training methods have shown promise in general coding tasks, their systematic application to VD remains underexplored. In this paper, we present the first comprehensive investigation into the post-training pipeline for LLM-based VD, spanning from cold-start SFT to off-policy preference optimization and on-policy RL, uncovering how data curation, stage interactions, reward mechanisms, and evaluation protocols collectively dictate the efficacy of model training and assessment. Our study identifies practical guidelines and insights: (1) SFT based on rejection sampling greatly outperforms rationalization-based supervision, which can introduce hallucinations due to ground-truth leakage. (2) While increased SFT epochs constantly benefit preference optimization, excessive SFT inhibits self-exploration during RL, ultimately limiting performance gains. (3) Coarse-grained reward signals often mislead RL, whereas fine-grained root-cause judgments ensure reliable credit assignment. Specification-based rewards offer further benefits but incur significant effort in specification generation. (4) Although filtering extremely hard-to-detect vulnerability samples improves RL training efficiency, the cost of performance loss should be considered in practical applications. (5) Models trained under GRPO significantly outperform those using SFT and preference optimization (i.e., DPO and ORPO), as well as a series of zero-shot SOTA LLMs, underscoring the significant potential of on-policy RL for LLM-based VD. (6) In contrast to binary matching that tends to overestimate performance, LLM-as-a-Judge based on root-cause analysis provides a more robust evaluation protocol, although its accuracy varies across judge models with different levels of security expertise.

Despite strong performance on existing benchmarks, it remains unclear whether large language models can reason over genuinely novel scientific information. Most evaluations score end-to-end RAG pipelines, where reasoning is confounded with retrieval and toolchain choices, and the signal is further contaminated by parametric memorization and open-web volatility. We introduce DeR2, a controlled deep-research sandbox that isolates document-grounded reasoning while preserving core difficulties of deep search: multi-step synthesis, denoising, and evidence-based conclusion making. DeR2 decouples evidence access from reasoning via four regimes--Instruction-only, Concepts (gold concepts without documents), Related-only (only relevant documents), and Full-set (relevant documents plus topically related distractors)--yielding interpretable regime gaps that operationalize retrieval loss vs. reasoning loss and enable fine-grained error attribution. To prevent parametric leakage, we apply a two-phase validation that requires parametric failure without evidence while ensuring oracle-concept solvability. To ensure reproducibility, each instance provides a frozen document library (drawn from 2023-2025 theoretical papers) with expert-annotated concepts and validated rationales. Experiments across a diverse set of state-of-the-art foundation models reveal substantial variation and significant headroom: some models exhibit mode-switch fragility, performing worse with the Full-set than with Instruction-only, while others show structural concept misuse, correctly naming concepts but failing to execute them as procedures.

Document-based LLM systems often flatten a PDF before guardrails inspect it. That step can discard evidence that an instruction was never visible to the user. We introduce CrackedPDFs, a controlled benchmark for hidden prompt injection in PDFs. The benchmark contains 29,322 generated PDFs from 4,983 base docu ments. It includes 9,774 injected files and 19,548 benign or matched-confounder files. We evaluate PromptGuard and a rule baseline. We also evaluate structural only learned models and a sanitized hybrid detector. The evaluation uses held-out provenance splits and paired benign-confounder controls. It also uses label-shuffle checks and shortcut audits. On a 2,919-document held-out test set, the hybrid de tector reaches 0.960 F1. ROC-AUC is 0.998 and PR-AUC is 0.997. It also ranks injected files above matched benign confounders in 95.9% of 973 pairs. Prompt Guard has low recall when given extracted text only. Structural-only learned mod els are weak under paired controls. A text-only TF-IDF model reaches perfect held-out scores but fails shortcut audits. These results show that document-aware hybrid detection is useful under controlled paired evaluation. They do not show broad real-world robustness or reliable cross-family generalization.

Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

Agent benchmarks increasingly evaluate repository editing, web research, terminal use, and long-horizon interaction. Their scores support capability claims only when the evaluation protocol keeps the intended capability necessary for success. Recent reward-hacking benchmarks and system reports show that agents can instead recover public solutions, read evaluation artifacts, infer generator structure, manipulate feedback, or benefit from invalid scoring paths; existing responses do not provide a common procedure for attributing these shortcuts and quantifying their effect across benchmarks. We formulate protocol validity and introduce HackDetect, a post-hoc audit that identifies an exposure, determines how the agent used it, and assesses whether the resulting score is misleading. We quantify score inflation with the Mislead gap, defined as the exploit score minus the intended score. We audit 2,385 traces across 15 agent benchmarks and find evidence of exposures and reward hacking in 67.0% of Frontier Science traces and 66.7% of AutoLab tasks. Across paired comparisons, we measure score inflation of 0.45-1.00, showing that benchmark reports should provide evidence that scores reflect the intended capability.

Test data from public benchmarks inevitably leaks into pretraining corpora, inflating evaluation scores once memorized. Contamination mitigation evaluation intervenes in the decoding process to suppress memorization and restore a contaminated model's genuine capability, but its prevailing metric, the G-AP (Gap of Aggregate Performance), is flawed. Discrete correct/incorrect readouts cannot characterize per-question performance, averaging before differencing lets over- and under-suppression cancel out, and uniform per-question weighting invites strategies to push solve probabilities onto the clean model's high-frequency values. We propose SA-PPG (Stratified Aggregate of Per-question Probability Gaps): estimate each question's solve probability by sampling, difference it against the clean model per question, and aggregate within groups defined by the clean model's solve probability. Existing mitigation strategies first estimate where contamination lies and then operate on the estimate, so they are only as correct as the estimate. RailCap instead judges contamination during generation: whenever a sample falls back onto the greedy trajectory, the next trajectory token is capped to the runner-up, accumulating suppression until the response distribution becomes sufficiently dispersed. Across multiple contaminated models and benchmarks, SA-PPG reveals that prior strategies' restoration is substantially overestimated, while RailCap attains the lowest SA-PPG.

Large language models (LLMs) are increasingly used in sensitive domains, where their ability to infer personal data from seemingly benign text introduces emerging privacy risks. While recent LLM-based anonymization methods help mitigate such risks, they often rely on proprietary models (e.g., GPT-4), raising concerns about cost and the potential exposure of sensitive data to untrusted external systems. To address this, we introduce SElf-refining Anonymization with Language model (SEAL), a novel distillation framework for training small language models (SLMs) to perform effective anonymization without relying on external models at inference time. SEAL leverages adversarial interactions between an LLM anonymizer and an inference model to collect trajectories of anonymized texts and inferred attributes, which are then used to distill anonymization and critique capabilities into SLMs through supervised fine-tuning and preference learning. The resulting models learn both to anonymize text and to evaluate their outputs, enabling iterative improvement of anonymization quality via self-refinement. Experiments on SynthPAI, a dataset of synthetic personal profiles and text comments, demonstrate that SLMs trained with SEAL achieve substantial improvements in anonymization capabilities. Notably, 8B models attain a privacy-utility trade-off comparable to that of the GPT-4 anonymizer and, with self-refinement, even surpass it in terms of privacy protection. These results highlight the effectiveness of our adversarial distillation framework for training SLMs as efficient anonymizers.

Statutes are increasingly parsed by machines before people read them, and the parsers disagree: on Missouri's statutes, two independently written extractors diverge on numeric-threshold presence at a false-negative rate of 0.43. We ask what formal logic survives such noise. We build a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts: per-attribute inter-extractor disagreement is measured, replayed against the basis in 1,000 Monte Carlo trials, and an implication is certified only when a one-sided Wilson 95% lower bound on survival reaches 0.95; every certified implication carries premise spans and a minimal counterexample. On 29,365 Missouri sections and 502 Indian central-Act sections, the preregistered held-out gate passes (10 statute families across 7 Titles exact; 16 across 11 with 5% tolerance), yet under one globally deployed error model 93.2% of held-out chapters fall below the informativeness floor, and a 2x2 factorial assigns that to calibration-rate transfer, not selection. The certificate is usable but fragile: deploy it per-chapter-calibrated or error-tolerant. Code, data products, and the audit trail, including one retracted claim, are released.

This paper introduces BioAgent Bench, a benchmark dataset and an evaluation suite designed for measuring the performance and robustness of AI agents in common bioinformatics tasks. The benchmark contains curated end-to-end tasks (e.g., RNA-seq, variant calling, metagenomics) with prompts that specify concrete output artifacts to support automated assessment, including stress testing under controlled perturbations. We evaluate frontier closed-source and open-weight models across multiple agent harnesses, and use an LLM-based grader to score pipeline progress and outcome validity. We find that frontier agents can complete multi-step bioinformatics pipelines without elaborate custom scaffolding, often producing the requested final artifacts reliably. However, robustness tests reveal failure modes under controlled perturbations (corrupted inputs, decoy files, and prompt bloat), indicating that correct high-level pipeline construction does not guarantee reliable step-level reasoning. Finally, because bioinformatics workflows may involve sensitive patient data, proprietary references, or unpublished IP, closed-source models can be unsuitable under strict privacy constraints; in such settings, open-weight models may be preferable despite lower completion rates. We release the dataset and evaluation suite publicly.

Predictive models in biomedicine depend on structured assay data locked in the text, tables, and supplements of primary publications. This bottleneck is especially acute in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files. Inconsistent compound identifiers and incomplete or implicit assay context further demand domain-specific logic that generic LLM pipelines do not provide. Existing molecular glue and PROTAC databases are manually curated and often lack the experimental context required for downstream modeling. We formulate TPD database extraction as a domain-specific curation task and present an expert-in-the-loop LLM workflow, evaluated through a triangular comparison among LLM predictions, standardized baseline records, and expert-annotated ground truth. A lightweight cross-validated prompt-refinement module adapts extraction instructions from scarce expert annotations. With only seven annotated molecular glue publications, the workflow achieved record-level F_1 = 0.98 and transferred to PROTACs by terminology substitution alone, maintaining record-level F_1 > 0.93. Applied at scale, it expanded molecular glue and PROTAC databases by 81% and 92% records, respectively, with 92% and 82.5% of newly recovered records validated as correct upon expert review. The workflow also recovered kinetic and assay-context information essential for cross-study potency comparison and condition-aware degradation modeling. We release the workflow, prompts, evaluation code, and extracted datasets as resources for TPD data curation and AI-assisted scientific curation more broadly.

Routine complete blood counts (CBCs) could yield new biomarkers, but the private records needed to evaluate candidates cannot be shared with frontier language model agents that excel at discovery. We distilled the evidence held in the Clalit Health Services panel of over 5.4 million patients into a released scoring tool: for each of 13 immune-mediated diseases, a graph attention network was trained inside the data boundary to predict the case-control AUC of candidate CBC expressions, and only the trained weights were released. The tool grounds an agent's propose-score-refine loop in real-world data without exposing any patient data. In external validation, agent-discovered expressions improved on their literature-seeded starting points by a median of 4.18 AUC percentage points, and across three independent cohorts, reranking the candidates of three frontier research tools improved on their first choices in most comparisons, with gains that varied by cohort. The released scorer supports privacy-preserving biomarker hypothesis generation.

Existing long-horizon agent benchmarks often emphasize final task success while providing limited assessment of intermediate execution validity. Yet real-world business operations require agents to navigate extensive, potentially outdated or conflicting policies while satisfying critical constraints throughout execution. These requirements are particularly consequential in high-stakes domains such as payments, where intermediate violations can incur financial losses or regulatory consequences regardless of the final outcome. To bridge this evaluation gap, we introduce BizBench, a multi-turn benchmark grounded in real-world procedures from a leading payments company. Its executable environment integrates 789 knowledge documents and 601 discoverable business tools, with consistent policies, states, and tool behavior. The benchmark comprises 50 tasks involving resource constraints, temporal dependencies, prior commitments, and conflicting or outdated guidance. Each task is validated through reference execution, while deterministic checks assess both goal attainment and execution validity without prescribing a single solution path. Experiments reveal substantial limitations even among frontier models: Claude Opus 5 and GPT-6 Sol achieve strict success rates of only 64.0% and 61.3%, respectively. For Qwen3.8-27B, goal attainment reaches 42.0%, but strict success drops to 32.7%, demonstrating how outcome-based evaluation can mask invalid execution. Beyond success rates, DeepSeek-V4.1-Flash averages 191.3 tool calls per run, approximately 70% more than reference executions, revealing substantial execution overhead. Overall, BizBench establishes a testbed for evaluating and advancing long-horizon agents beyond task completion, toward reliable and constraint-compliant execution in complex business environments.

Evaluations of LLM systems routinely average over small prompt sets and report models as a ranked table. We ask how much confidence such a table deserves, using LLM-based prompt-structure inference as the case study: eight open model variants across five families and 8B to 675B parameters, caching disabled, 293 raw intermediate representations persisted. The measured phenomenon is unstable to begin with. Identical calls do not reliably recover identical structure, with mean node-set Jaccard from 0.39 to 0.96 and 72% of prompt-model cells never node-set-perfect. Auditing the evaluation weakens its conclusions further, and this is our main contribution. Under a joint cluster bootstrap over prompts, only the bottom of the ranking is firm: the two least reproducible models hold rank in 99% and 86% of replicates, the middle four in 27% to 48%, and the top two in 68% each, so the table identifies the worst model reliably but does not reliably identify the best. Two equally defensible rules for merging repeated campaigns change four of eight rows and move the study-wide headline by 7 percentage points. Checking the inferred structure against ground-truth annotations shows reproducibility cannot be read as accuracy. And four of the eight endpoints were withdrawn within ten weeks of measurement, so the study as specified can no longer be run. Small-sample LLM evaluations can therefore look far more definitive than their evidence supports. We recommend reporting rank stability, per-cell provenance, executed sensitivity comparisons, raw per-run outputs, and a measurement date alongside any ranking.

Competition retrospectives are useful when they explain what a leaderboard measured, how hidden evaluation changed conclusions, and which design patterns were rewarded. We revisit the CODS 2025 challenge, a privacy-aware Codabench competition on industrial multi-agent orchestration built on . We combine final rank sheets, a 300-submission server log, 149-team registrations, best-submission exports, the organizer winners report, the companion system paper, and verified planning-track source trees. Five results stand out. First, the public planning leaderboard saturates at 72.73\%, and richer prompts do not improve that peak. Second, hidden evaluation changes the story: public and private scores correlate moderately in planning (r{=}0.69) but negatively in execution (r{=}{-}0.13), with several 45.45\% public execution systems reaching 63.64\% on the hidden set. Third, the term is numerically almost inert in the official composite -- combined on a 0--1 scale with 0--100 percentage scores, it contributes at most 0.05 points per track, and rescaling would swap the top two teams. Fourth, the competition is operationally account-based but substantively team-based: 149 registered teams reduce to 24 with non-zero public scores and 11 fully ranked, while 52.3\% of deduplicated registrations list multiple usernames. Fifth, successful execution methods mostly improve guardrails -- response selection, contamination cleanup, fallback, and context control -- rather than novel agent architectures. These findings identify which behaviors the evaluation rewarded, and motivate scale-aware composites, skill-level diagnostics, and versioned artifact release.

As multimodal models like CLIP become integral to downstream systems, the need to remove sensitive information is critical. However, machine unlearning for contrastively-trained encoders remains underexplored, and existing evaluations fail to diagnose fine-grained, association-level forgetting. We introduce SALMUBench (Sensitive Association-Level Multimodal Unlearning), a benchmark built upon a synthetic dataset of 60K persona-attribute associations and two foundational models: a Compromised model polluted with this data, and a Clean model without it. To isolate unlearning effects, both are trained from scratch on the same 400M-pair retain base, with the Compromised model additionally trained on the sensitive set. We propose a novel evaluation protocol with structured holdout sets (holdout identity, holdout association) to precisely measure unlearning efficacy and collateral damage. Our benchmark reveals that while utility-efficient deletion is feasible, current methods exhibit distinct failure modes: they either fail to forget effectively or over-generalize by erasing more than intended. SALMUBench sets a new standard for comprehensive unlearning evaluation, and we publicly release our dataset, models, evaluation scripts, and leaderboards to foster future research.

We present a systems-security case study of a two-node split-LLM training system whose privacy evaluation passed while leaving an observable channel untested. The Trusted Local Node (TLN) sends protected activations to the Untrusted Cloud Node (UCN), the UCN returns its output, and TLN, holding the private loss, returns the output gradient. The frame the UCN receives mixes real rows with decoys, and the loss ignores the decoys. Their gradients are exactly zero, so the pattern of zeros reveals which rows were real. We measure it with a protocol fixed in advance: a leak injected at known strength to prove the instrument can see one, a shuffled-label control to prove it does not report absent leaks, and a threshold set before the runs. Across nine seeds, the zeros identified the real rows on every frame, 4,096 of 4,096 per run. An attack on the frame contents recovered about one extra token per hundred over a constant-guess baseline (+0.65 to +1.50 percentage points); the shuffled controls recovered nothing. A second set of runs repeated this on a configuration that keeps model quality within budget, so the finding is not confined to a setting nobody would deploy. On both datasets, every such run passed the forward-channel privacy check and the quality check, yet failed that same check once the returned gradient was included. Clipping and noising each row of the gradient closed the leak for about 0.01 nats of held-out cross-entropy. The system is not thereby safe: five classes of attack, including those accumulating observations across training steps, were never measured.

As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.

Large language model agents increasingly undertake long-horizon tasks in complex environments characterized by fragmented evidence, partially completed states, and dependencies across resources and artifacts. Existing agents may satisfy individual requirements while failing to maintain these relationships, producing locally correct but incomplete outcomes. These failures arise because task completion depends on relationships among environmental evidence, consequential decisions, and required effects. To address this completion gap, we propose Interlock, a framework for task synthesis that represents each task as a typed completion graph. Grounding edges connect public evidence to decisions, coupling hyperedges impose joint constraints across components, and realization edges connect decisions to all required downstream effects. Interlock compiles this graph into a resettable workspace, a public request, and an evaluation rubric. A source-coverage review checks whether the original requirements survive generation, while a public-input review checks whether the rubric is supported by the materials available to the solver. An agent-as-judge inspects execution artifacts and runs verification scripts, retaining trajectories judged to satisfy all mandatory requirements. To evaluate training utility, we fine-tune two Qwen3.6 backbones on their own accepted trajectories. Across the two backbones, training improves AutomationBench task pass rates by 4.00-4.50 percentage points, JobBench weighted scores by 3.20-6.64 points, and Harvey LAB criterion pass rates by 3.89-4.16 percentage points. These results demonstrate the training utility of the synthesized experience across external task environments.

The development of general-purpose agents requires a shift from executing simple instructions to completing complex, real-world productivity workflows. However, current tool-use benchmarks remain misaligned with real-world requirements, relying on AI-generated queries, dummy tools, and limited system-level coordination. To address this, we propose GTA-2, a hierarchical benchmark for General Tool Agents (GTA) spanning atomic tool use and open-ended workflows. Built on real-world authenticity, it leverages real user queries, deployed tools, and multimodal contexts. (i) GTA-Atomic, inherited from our prior GTA benchmark, evaluates short-horizon, closed-ended tool-use precision. (ii) GTA-Workflow introduces long-horizon, open-ended tasks for realistic end-to-end completion. To evaluate open-ended deliverables, we propose a recursive checkpoint-based evaluation mechanism that decomposes objectives into verifiable sub-goals, enabling unified evaluation of both model capabilities and agent execution frameworks (i.e., execution harnesses). Experiments reveal a pronounced capability cliff: while frontier models already struggle on atomic tasks (below 50%), they largely fail on workflows, with top models achieving only 14.39% success. Further analysis shows that checkpoint-guided feedback improves performance, while advanced frameworks such as Manus and OpenClaw substantially enhance workflow completion, highlighting the importance of execution harness design beyond the underlying model capacity. These findings provide guidance for developing reliable personal and professional assistants. Dataset and code will be available at https://github.com/open-compass/GTA.

Do research agents have taste, the judgment of which experiments are worth running before their outcomes are known? We give six frontier models seven recent NLP papers. Each model, working as a research agent alone or as a team, proposes minimum viable experiments (MVEs), forecasts whether each will pass a threshold it sets and ranks them. A separate executor then attempts every MVE, including those ranked last, and three model auditors judge the results. Of 415 MVEs, 194 are judged sound, valuable and new, but fewer than half of these (73) are also judged beyond the seed paper and non obvious. The agents' first choices match the auditors' most valued experiment more often than a random pick, but are not shown to improve on choosing the experiment written first. Their forecasts beat a base rate but are not measurably more accurate than those of a fresh instance of the same model given only the paper and registrations. Teams cost more without a clear overall improvement in proposal quality. The agents produce many worthwhile experiments, but the evidence that their rankings improve the choice of which to run remains limited. The evaluation harness and a dataset of all 415 MVEs, with their forecasts and audited outcomes, are available at https://github.com/ArtKanke/rTaste.

LLM agents are expected to complete end-to-end units of work across software tools, business services, and local workspaces. Yet many agent benchmarks freeze a curated task set at release time and grade mainly the final response, making it difficult to evaluate agents against evolving workflow demand or verify whether a task was executed. We introduce Claw-Eval-Live, a live benchmark for workflow agents that separates a refreshable signal layer, updated across releases from public workflow-demand signals, from a reproducible, time-stamped release snapshot. Each release is constructed from public workflow-demand signals, with ClawHub Top-500 skills used in the current release, and materialized as controlled tasks with fixed fixtures, services, workspaces, and graders. For grading, Claw-Eval-Live records execution traces, audit logs, service state, and post-run workspace artifacts, using deterministic checks when evidence is sufficient and structured LLM judging only for semantic dimensions. The release contains 105 tasks spanning controlled business services and local workspace repair, and evaluates 13 frontier models under a shared public pass rule. Experiments reveal that reliable workflow automation remains far from solved: the leading model passes only 66.7% of tasks and no model reaches 70%. Failures are structured by task family and execution surface, with HR, management, and multi-system business workflows as persistent bottlenecks and local workspace repair comparatively easier but unsaturated. Leaderboard rank alone is insufficient because models with similar pass rates can diverge in overall completion, and task-level discrimination concentrates in a middle band of tasks. Claw-Eval-Live suggests that workflow-agent evaluation should be grounded twice, in fresh external demand and in verifiable agent action.

We conduct a systematic audit of three widely used reasoning benchmarks, SocialIQa, FauxPas-EAI, and ToMi, and uncover pervasive flaws in both benchmark items and evaluation methodology. Using five LLMs (GPT-{3, 3.5, 4, o1}, and LLaMA 3.1) as diagnostic tools, we identify structural, semantic, and pragmatic issues in benchmark design (e.g., duplicated items, ambiguous wording, and implausible answers), as well as scoring procedures that prioritize output form over reasoning process. Through systematic human annotation and re-evaluation on cleaned benchmark subsets, we find that model scores often improve not due to due to erratic surface wording variations and not to improved reasoning. Infact, further analyses show that model performance is highly sensitive to minor input variations such as context availability and phrasing, revealing that high scores may reflect alignment with format-specific cues rather than consistent inference based on the input. These findings challenge the validity of current benchmark-based claims about reasoning in LLMs, and highlight the need for evaluation protocols that assess reasoning as a process of drawing inference from available information, rather than as static output selection. We release audited data and evaluation tools to support more interpretable and diagnostic assessments of model reasoning.

AI agents combine language models with external data and tools that can modify files, call APIs, or execute code. Security failures can arise when adversarial content changes an agent's tool use or when the surrounding software contains vulnerabilities such as path traversal or command injection. We study authorized white-box pre-deployment auditing, where the auditor has access to the target repository and a controlled runtime, but successful attacks must still act through the task-defined attacker interface and be confirmed by an external verifier. We present AgentXploit, a two-role auditing system that separates repository-level attack-path discovery from runtime exploitation. The Analyzer Agent traces attacker-controlled inputs to sensitive operations and records code-supported candidate attack paths; the Exploiter Agent turns these paths into concrete attacks and revises them using runtime feedback. We also introduce AgentXploit-Bench, containing 72 reproducible vulnerabilities across 12 open-source AI-agent systems and frameworks. Across three runs, AgentXploit reaches 59.3% end-to-end success, compared with 38.4% for Codex. Under a token-budget-matched comparison, Codex reaches 46.3%. On AgentDojo, where injection points are provided, the Exploiter Agent reaches 79.2% attack success versus 52.7% for AgentVigil. These results highlight repository discovery and runtime exploitation as distinct challenges in end-to-end agent security auditing.

Agent benchmarks often evaluate only final answers even when agents run on stateful runtimes. We argue this under-specifies what is being evaluated: the proper unit is a declared model-plus-runtime configuration whose failures can occur in evidence acquisition, runtime routing, safety boundaries, or repeated execution. We present ClawProBench, a trace-aware benchmark for runtime-native agent evaluation instantiated on OpenClaw, a live agent runtime with workspace tools and native surfaces for browsing, memory, messaging, scheduling, skills, and subagents. ClawProBench defines two tracks: a 102-scenario full profile with live workspace and native-runtime routing tasks, and a frozen 68-scenario holdout with closed-world JSON output contracts for robust ranking. Trials are scored from execution traces via a safety-gated formula combining correctness, process quality, and efficiency, preserving failure evidence for audit. Our anonymous artifact includes benchmark definitions, scoring code, manifests and sanitized traces. We evaluate 68 configurations on the full profile and 37 on holdout. The top safety-gated average trace score is 0.7671. Native-runtime tasks underperform workspace-live tasks (0.5238 vs. 0.6415). On holdout, pass@k-any outperforms strict three-trial pass (0.6638 vs. 0.2890), while full-profile and holdout rankings show weak alignment (Spearman 0.1300). Rankings based purely on correctness differ substantially from process-aware, safety-gated and strict-pass views. Final-answer leaderboards may hide native-surface weaknesses, one-off successes and trace-local agent failure modes.

Benchmark contamination has become a significant concern in the LLM evaluation community. Previous Agents-as-an-Evaluator address this issue by involving agents in the generation of questions. Despite their success, the biases in Agents-as-an-Evaluator methods remain largely unexplored. In this paper, we present a theoretical formulation of evaluation bias, providing valuable insights into designing unbiased evaluation protocols. Furthermore, we identify two type of bias in Agents-as-an-Evaluator through carefully designed probing tasks on a minimal Agents-as-an-Evaluator setup. To address these issues, we propose the Unbiased Evaluator, an evaluation protocol that delivers a more comprehensive, unbiased, and interpretable assessment of LLMs.Extensive experiments reveal significant room for improvement in current LLMs. Additionally, we demonstrate that the Unbiased Evaluator not only offers strong evidence of benchmark contamination but also provides interpretable evaluation results.

Tool-using language-model agents select and execute third-party artifacts. Different implementations can return the requested output while producing hidden execution effects that task-, attack-, or choice-based evaluations may miss. We study functional counterfeits: implementations that match benign alternatives on the requested output but add an effect forbidden by the task contract. We introduce SINGED (Source Integrity and the Nonidentifiability Gap in Execution Decisions for LLM Agents), a controlled benchmark covering five primary and two held-out task families. It varies displayed rank, evidence depth, decision policy, model release, and agent configuration, while task and process oracles verify the artifact and execution path. Across 7,549 audited trials, the randomized-rank study finds counterfeit execution in 45% (27/60) of rank-one trials and none at later ranks. Cross-candidate comparison eliminates shallow failures and reduces layered failures from 15.7% to 4.2%, but leaves dependency failures; its benefit is uncertain on unseen effects and public-package structures. Moreover, seven releases with no counterfeit executions when benign alternatives are available execute the counterfeit in 55/175 single-source cells after alternatives are removed. SINGED thus exposes a rank-, evidence-, and choice-sensitive outcome-to-execution gap: evaluation must connect correct outputs to execution paths.

Large language models are increasingly deployed as autonomous agents executing multi-step workflows in real-world software environments. However, existing agent benchmarks suffer from three critical limitations: (1) trajectory-opaque grading that checks only final outputs, (2) underspecified safety and robustness evaluation, and (3) narrow modality coverage and interaction paradigms. We introduce Claw-Eval, an end-to-end evaluation suite addressing all three gaps. It comprises 300 human-verified tasks spanning 9 categories across three groups (general service orchestration, multimodal perception and generation, and multi-turn professional dialogue). Every agent action is recorded through three independent evidence channels (execution traces, audit logs, and environment snapshots), enabling trajectory-aware grading over 2,159 fine-grained rubric items. The scoring protocol evaluates Completion, Safety, and Robustness, reporting Average Score, Pass@k, and Pass^k across three trials to distinguish genuine capability from lucky outcomes. Experiments on 14 frontier models reveal that: (1) trajectory-opaque evaluation is systematically unreliable, missing 44% of safety violations and 13% of robustness failures that our hybrid pipeline catches; (2) controlled error injection primarily degrades consistency rather than peak capability, with Pass^3 dropping up to 24% while Pass@3 remains stable; (3) multimodal performance varies sharply, with most models performing poorer on video than on document or image, and no single model dominating across all modalities. Beyond benchmarking, Claw-Eval highlights actionable directions for agent development, shedding light on what it takes to build agents that are not only capable but reliably deployable.

We set out to train behavioral dispositions (self-verification, uncertainty acknowledgment, feedback integration) into small language models (0.6B to 2.3B effective parameters) through a four-stage all-MIT distillation pipeline, with follow-on experiments on inference-time attention-head interventions and a frozen-base confidence-gated sidecar. An internal draft reported +33.9-point MCAS and +15.3-point HumanEval gains on a Qwen3-0.6B student; a second-pass sanity check falsified both numbers before publication. The HumanEval delta was a truncation artifact (n_predict=512) that inverted to -8.0 points at n_predict=1024; the MCAS gain disappeared under apples-to-apples scoring. That falsification triggered three subsequent arcs. Across (1) SFT/DPO LoRA on three model families and two domains, (2) inference-time attention-head tempering on o_proj, and (3) a training-free frozen-base sidecar reading the final-token hidden state h_last, we find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry. The failure is consistent across five models (Qwen3-0.6B, Qwen3-1.7B, Qwen3.5-0.8B, Gemma 4 E2B, and SmolLM2-1.7B-Instruct). A within-distribution cross-validation pass (AUC=0.683) collapsed to chance on fresh prompts (AUC=0.516). We contribute a three-arc negative result with mechanism, a two-failure-mode taxonomy for linear h_last probes, and an honest falsification pipeline that converts the class of false positives we ourselves produced into publishable negatives. As an independent finding, Gemma 4 E2B exhibits near-complete confidence-correctness decoupling on the Chef domain (assertion asymmetry -0.009; the model asserts at 91% regardless of correctness).

Decentralized reputation schemes present a promising area of experimentation in blockchain applications. These solutions aim to overcome the shortcomings of simple monetary incentive mechanisms of naive tokenomics. However, there is a significant research gap regarding the limitations and benefits of such solutions. We formulate these trade-offs as a conjecture on the irreconcilability of three desirable properties of the reputation system in this context. Such a system can not be simultaneously generalizable, trustless, and Sybil resistant. To handle the limitations of this trilemma, we propose MeritRank: Sybil tolerant feedback aggregation mechanism for reputation. Instead of preventing Sybil attacks, our approach successfully bounds the benefits of these attacks. Using a dataset of participants' interactions in MakerDAO, we run experiments to demonstrate Sybil tolerance of MeritRank. Decay parameters of reputation in MeritRank: transitivity decay and connectivity decay, allow for a fine-tuning of desirable levels of reputation utility and Sybil tolerance in different use contexts.

Third-party LLM gateways have become a critical infrastructure layer between applications and external LLM providers. Conventional gateways do more than forward traffic: they decide which provider and model are called, whether fallback occurred, which stream is delivered, and what usage record should be billed. Because these decisions and records are authored inside the operator-controlled service, clients cannot independently distinguish honest mediation from route substitution, hidden fallback, stream manipulation, or forged provenance. We present an evidence-bound LLM gateway architecture that separates the operator control plane from an attested execution plane. Within the gateway, a measured Attested Gateway Runtime (AGR) is the only component allowed to decrypt requests, enforce path policy, construct upstream calls, and sign evidence. Clients verify signed release metadata and fresh attestation before encrypting requests to keys bound to the AGR measurement. AGR enforces request-scoped routing, fallback, and endpoint constraints, invokes admitted providers, returns encrypted response streams, and signs evidence binding the policy, selected route, endpoint identity, stream commitments, and completion metadata to the attested runtime. An initial Rust prototype on AWS Nitro Enclaves shows modest mechanism overhead and fail-closed detection of policy, routing, endpoint, and stream-evidence tampering outside the attested runtime.

Large Language Model (LLM) agents are increasingly adopted in real-world security operations with access to host artifacts and command-line interfaces (CLIs), making it critical to thoroughly assess their security capabilities. However, existing cybersecurity benchmarks focus on pre-compromise settings where agents are placed in a clean and idealized environment before an attack occurs. This leaves the post-compromise setting underexplored. To address this gap, we introduce SecRespond, the first benchmark for evaluating LLM agents on the post-compromise incident-response workflow. Given a forensic disk snapshot of a compromised host together with the alerts, vulnerability scans, and baseline checks reported by a host security product, agents are required to produce forensic reports on intrusions, baseline risks, and vulnerability risks, together with a remediation plan. We instantiate this task across 10 cyber ranges, each constructed from a distinct compromised cloud host, spanning 4 entry-point types, 21 ATT&CK techniques, and 5 operating systems. We evaluate 23 frontier LLMs on the OpenCode agent harness. Experimental results show that although current agents can reliably uncover the problems exposed by alerts, they struggle to proactively investigate the disk for silent intrusions and to produce comprehensive, verified remediation plans, with no model achieving complete detection and remediation on any single range. This reveals a fundamental bottleneck in building agents for real-world incident response. The benchmark is publicly available at https://github.com/Alibaba-NLP/qqr/tree/main/data/secrespond.

LLM agents often retry after external validation rejects a candidate, but the interface between validation and the next model call remains underspecified. We introduce VeriHarness, a code-controlled agent loop in which models generate candidates while external validators control acceptance, budgets, and traces. We use it to compare raw diagnostics with feedback that identifies the failure location, observed value, and admissible alternatives. Across 50 paired TextWorld games under a four-call cap, feedback containing all three fields raises terminal success from 14/50 to 36/50 for Qwen2.5-Coder-14B (+44 percentage points) and from 8/50 to 29/50 for Llama-3.1-8B (+42 points). Ablations locate most of the gain in the admissible alternatives: feedback containing only the location and observed value remains near the raw diagnostic baseline. Presenting the complete repair information in prose instead of a keyed JSON record yields nearly the same success, providing no evidence that JSON syntax itself improves repair. The ordering persists across the tested call budgets and one sampled-decoding setting.

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

Coding agents and LLM-powered applications routinely send potentially sensitive content to cloud LLM APIs where it may be logged, retained, used for training, or subpoenaed. Existing privacy tooling focuses on network-level encryption and organization-level DLP, neither of which addresses the content of prompts themselves. We present a systematic empirical evaluation of eight techniques for privacy-preserving LLM requests: (A) local-only inference, (B) redaction with placeholder restoration, (C) semantic rephrasing, (D) Trusted Execution Environment hosted inference, (E) split inference, (F) fully homomorphic encryption, (G) secret sharing via multi-party computation, and (H) differential-privacy noise. We implement all eight (or a tractable research-stage subset where deployment is not yet feasible) in an open-source shim compatible with MCP and any OpenAI-compatible API. We evaluate the four practical options (A, B, C, H) and their combinations across four workload classes using a ground-truth-labelled leak benchmark of 1,300 samples with 4,014 annotations. Our headline finding is that no single technique dominates: the combination A+B+C (route locally when possible, redact and rephrase the rest) achieves 0.6% combined leak on PII and 31.3% on proprietary code, with zero exact leaks on PII across 500 samples. We present a decision rule that selects the appropriate option(s) from a threat-model budget and workload characterisation. Code, benchmarks, and evaluation harness are released at https://github.com/jayluxferro/llm-redactor.

ARC-AGI-3 evaluates agents in interactive environments whose rules and objectives must be inferred from observation. We present Kepler, an open-source harness that represents hypotheses as executable world models and validates them through retrospective transition checks and conditional prediction checks. Under one frozen Claude Opus 5 configuration, Kepler obtained a server-verified 100.00 RHAE on all 25 public games, with no per-game model selection or score-conditioned reruns. On 181 of 183 completed levels, the final Opus attempt used no more actions than the corresponding median-human baseline. The retained board runs used 8,256 environment actions, of which 7,292 occurred in scored levels. Retained local provider-session records yield 858.0 million tokens, 97.37% cache reads, and a \$777.72 cost at September 1, 2026 API list-equivalent rates. We also report three evaluation failures: source-code leakage that produced an invalid perfect run, agents reconstructing a removed harness in a control condition, and autonomous repair masking a broken planner. A single-game observation case study showed that animation frames contained task-relevant information absent from settled text grids. Across the final Claude Opus 5 and GPT-5.6 Sol boards, 48 of 50 game-model cells reached 100. These results indicate that public-set score alone has limited discriminative value and motivate first-attempt, cost-conditioned, and verification-aware reporting.

Emerging personal AI agents are moving toward persistent, multi-source memory. This creates an evaluation problem: systems must decide how to use conflicting or incomplete evidence; they cannot just retrieve facts from one clean history. Existing benchmarks rarely show whether an error came from the evidence given to a method or from the method's conflict-resolution step. We study this as selective QA over conflicting multi-source personal memory: systems answer based on conflicting, sometimes incomplete sources, or abstain when evidence is insufficient. We develop a benchmark containing 18 question templates across 8 reasoning types, 480 personas, 4 random seeds, and 34,560 instances, with controlled source distortions and deterministic ground truth. We evaluate the performance of baselines without access to any source, access to a single source, structured fusion methods, and frontier LLMs. The best trained fusion resolver reaches 80.3% accuracy, while the strongest prompt-only LLM baseline reaches 70.0%. With abstention, the same resolver reaches 85.3% selective accuracy at 78.3% coverage and the best LLM reaches 71.0% selective accuracy at 95.4% coverage. Different models have different strengths across reasoning types. We release the data, code, cached model outputs, and data-generating process for reuse.

Model Inversion (MI) attacks aim to reconstruct information from private training data by exploiting access to machine learning models T. To evaluate such attacks, the standard evaluation framework relies on an evaluation model E, trained under the same task design as T. This framework has become the de facto standard for assessing progress in MI research, used across nearly all recent MI studies without question. In this paper, we present the first in-depth study of this evaluation framework. In particular, we identify a critical issue of this standard framework: Type-I adversarial examples. These are reconstructions that do not capture the visual features of private training data, yet are still deemed successful by T and ultimately transferable to E. Such false positives undermine the reliability of the standard MI evaluation framework. To address this issue, we introduce a new MI evaluation framework that replaces the evaluation model E with advanced Multimodal Large Language Models (MLLMs). By leveraging their general-purpose visual understanding, our MLLM-based framework does not depend on training of shared task design as in T, thus reducing Type-I transferability and providing more faithful assessments of reconstruction success. Using our MLLM-based evaluation framework, we reevaluate 27 diverse MI attack setups and empirically reveal consistently high false positive rates under the standard evaluation framework. Importantly, we demonstrate that many state-of-the-art (SOTA) MI methods report inflated attack accuracy, indicating that actual privacy leakage is significantly lower than previously believed. By uncovering this critical issue and proposing a robust solution, our work enables a reassessment of progress in MI research and sets a new standard for reliable and robust evaluation. Code can be found in https://github.com/hosytuyen/MI-Eval-MLLM

Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models. We show the surgery is not clean. As a disposition probe we use 21,600 decisions under uncertainty - weekly up/down calls on 60 Warsaw Stock Exchange equities over 18 weeks, replayed through a frozen pipeline so the decision-layer model is the only variable. The task elicits no refusals at all, so any between-arm delta is pure side effect. Holding provenance constant (official BF16 checkpoints, a single abliteration author, an identical serving stack, one byte-identical frozen prompt), we compare base and abliterated arms of two Mixture-of-Experts families, Gemma-4-26B-A4B-it and Qwen3-30B-A3B-Instruct-2507. Three effects replicate across both families (weeks-clustered bootstrap CIs excluding zero): abliterated models are systematically more optimistic (+12.2 pp Gemma, +7.4 pp Qwen; the confirmed preregistered endpoint), justify themselves at greater length, and use fewer explicit uncertainty words in forced self-critiques (both exploratory). A fourth effect reverses sign: the same operation makes Gemma-abliterated less confident and Qwen-abliterated more (family CIs non-overlapping) - one weight surgery, opposite shifts in expressed confidence. Capability covariates rule out instruction-following degradation as the driver, and no arm shows economic skill: the apparent edge of abliterated arms is regime beta, not alpha. Our provenance audit also caught two independent contamination channels - a mismatched-quantizer pilot pair and a stale community chat template that silently mangled the rendered prompt - suggesting toolchain artifacts are the rule in studies of community-modified checkpoints. Whoever deploys an "uncensored" model as an agent is deploying a measurably different decision-maker, not the base model minus refusals.

Recently, using a powerful proprietary Large Language Model (LLM) (e.g., GPT-4) as an evaluator for long-form responses has become the de facto standard. However, for practitioners with large-scale evaluation tasks and custom criteria in consideration (e.g., child-readability), using proprietary LLMs as an evaluator is unreliable due to the closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose Prometheus, a fully open-source LLM that is on par with GPT-4's evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. We first construct the Feedback Collection, a new dataset that consists of 1K fine-grained score rubrics, 20K instructions, and 100K responses and language feedback generated by GPT-4. Using the Feedback Collection, we train Prometheus, a 13B evaluator LLM that can assess any given long-form text based on customized score rubric provided by the user. Experimental results show that Prometheus scores a Pearson correlation of 0.897 with human evaluators when evaluating with 45 customized score rubrics, which is on par with GPT-4 (0.882), and greatly outperforms ChatGPT (0.392). Furthermore, measuring correlation with GPT-4 with 1222 customized score rubrics across four benchmarks (MT Bench, Vicuna Bench, Feedback Bench, Flask Eval) shows similar trends, bolstering Prometheus's capability as an evaluator LLM. Lastly, Prometheus achieves the highest accuracy on two human preference benchmarks (HHH Alignment & MT Bench Human Judgment) compared to open-sourced reward models explicitly trained on human preference datasets, highlighting its potential as an universal reward model. We open-source our code, dataset, and model at https://github.com/kaistAI/Prometheus.

Tool-augmented LLM agents can harbor implicit state that persists across sessions, activates through events, and propagates across agent boundaries---largely invisible to standard debugging. We formalize this as Persistent Semantic Entities (PSEs): constructs defined by name binding, event triggering, and cross-boundary propagation, and evaluate them across 24 models from 11 families (1.5B--1T parameters). First, every tested model is susceptible (20--100% on the 20-model susceptibility panel), with name binding as the necessary and dominant mechanism: without it, contamination is 0%. Second, persistence depends on contamination type rather than scale or deployment: preference contamination persists undecayed on every model probed (100% at t=10) and instruction contamination persists wherever adopted, persona-style injection decays partially (90%to10%), while factual injection is model-dependent---self-corrected on Llama-3.1-8B and GPT-4o-mini but held at ceiling on both Qwen2.5-coder variants, so we do not claim it self-corrects in general. The preference and instruction results hold across providers in our controlled setting. Third, context-isolated self-verification achieves 20--79% reduction (median 36.5%) without oracle references while keyword-based detection produces systematic false positives, and contamination compounds 1.9times along a four-stage agent pipeline (40%to75%). Preference and instruction contamination---persistent, lacking self-correction, and poorly captured by standard monitoring---represent a particularly concerning attack surface for deployed agent systems.

Modern LLM agents can retrieve information from accessible documents, and then disclose a sensitive conclusion assembled from them. This mosaic effect creates a critical risk for agents operating over confidential data. We quantify its realization with 30 synthetic cybersecurity, medical, and military examples, yielding 10,800 final answers from three LLM agents in controlled simulation and 1,440 answers from live agents in two open-source harnesses. Even under benign queries designed to be answerable from one document, GPT-5.6-Sol, the best performing model tested, discloses forbidden composites in 7.08% of answers (95% CI: 3.75-10.67%) in simulation, with worse disclosure rates in live harnesses. Adding visible classification headers and explicit aggregation rules reduces pooled disclosure rates substantially, yet unsafe disclosures persist in 32.37% of simulated answers (95% CI: 28.33-36.19%) across all intents despite these safeguards. Furthermore, increased refusals do not guarantee confidentiality, since significant refusals reveal the prohibited conclusion regardless. These results point to a troubling conclusion: sensitive data leaks via mosaic-style events are an observed confidentiality failure mode in the evaluated settings at non-negligible rates. They motivate stronger controls on what agents may infer and disclose in sensitive data environments with mixed-clearance users.

Powered by remarkable advancements in Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) demonstrate impressive capabilities in manifold tasks. However, the practical application scenarios of MLLMs are intricate, exposing them to potential malicious instructions and thereby posing safety risks. While current benchmarks do incorporate certain safety considerations, they often lack comprehensive coverage and fail to exhibit the necessary rigor and robustness. For instance, the common practice of employing GPT-4V as both the evaluator and a model to be evaluated lacks credibility, as it tends to exhibit a bias toward its own responses. In this paper, we present MLLMGuard, a multidimensional safety evaluation suite for MLLMs, including a bilingual image-text evaluation dataset, inference utilities, and a lightweight evaluator. MLLMGuard's assessment comprehensively covers two languages (English and Chinese) and five important safety dimensions (Privacy, Bias, Toxicity, Truthfulness, and Legality), each with corresponding rich subtasks. Focusing on these dimensions, our evaluation dataset is primarily sourced from platforms such as social media, and it integrates text-based and image-based red teaming techniques with meticulous annotation by human experts. This can prevent inaccurate evaluation caused by data leakage when using open-source datasets and ensures the quality and challenging nature of our benchmark. Additionally, a fully automated lightweight evaluator termed GuardRank is developed, which achieves significantly higher evaluation accuracy than GPT-4. Our evaluation results across 13 advanced models indicate that MLLMs still have a substantial journey ahead before they can be considered safe and responsible.

Proprietary LMs such as GPT-4 are often employed to assess the quality of responses from various LMs. However, concerns including transparency, controllability, and affordability strongly motivate the development of open-source LMs specialized in evaluations. On the other hand, existing open evaluator LMs exhibit critical shortcomings: 1) they issue scores that significantly diverge from those assigned by humans, and 2) they lack the flexibility to perform both direct assessment and pairwise ranking, the two most prevalent forms of assessment. Additionally, they do not possess the ability to evaluate based on custom evaluation criteria, focusing instead on general attributes like helpfulness and harmlessness. To address these issues, we introduce Prometheus 2, a more powerful evaluator LM than its predecessor that closely mirrors human and GPT-4 judgements. Moreover, it is capable of processing both direct assessment and pair-wise ranking formats grouped with a user-defined evaluation criteria. On four direct assessment benchmarks and four pairwise ranking benchmarks, Prometheus 2 scores the highest correlation and agreement with humans and proprietary LM judges among all tested open evaluator LMs. Our models, code, and data are all publicly available at https://github.com/prometheus-eval/prometheus-eval.

Large language model (LLM) agents coordinate complex tasks through multi-role and multi-stage workflows. Upstream state is repeatedly transformed into intermediate language artifacts, such as summaries, plans, tickets, memories, and handoff notes, from which downstream components act. For action-constraining state, topical retention is insufficient: an artifact may mention an unresolved condition while changing it from a requirement that must be resolved before execution into information that may merely inform the next action. We study this action-binding role as operational state preservation. Safety blockers provide a controlled instance because each source state has an explicit prerequisite, authority, fallback, and execution consequence. We condition on correct upstream identification, vary the handoff transformation, and evaluate an executor restricted to the resulting artifact. Across 1,296 controlled synthetic episodes, direct-handoff controls preserve every blocker, whereas compression, plan assimilation, convergence, ownership deferral, and precedent substitution repeatedly turn binding state into caveats or non-binding considerations. Normal handoff compression produces 100.0% deactivation and 54.2% forbidden action. Restoring all four state fields raises preservation to 100.0% and reduces forbidden action to 0.0%. Fixed-artifact interventions further separate preservation from containment: downstream verification eliminates forbidden action while artifact deactivation remains 95.3%. These results identify a state-transmission failure between information extraction and action. Handoff transformations can retain state content while weakening its constraints on downstream action. Semantic availability does not guarantee operational preservation.

While watermarks for closed LLMs have matured and have been included in large-scale deployments, these methods are not applicable to open-source models, which allow users full control over the decoding process. This setting is understudied yet critical, given the rising performance of open-source models. In this work, we lay the foundation for systematic study of open-source LLM watermarking. For the first time, we explicitly formulate key requirements, including durability against common model modifications such as model merging, quantization, or finetuning, and propose a concrete evaluation setup. Given the prevalence of these modifications, durability is crucial for an open-source watermark to be effective. We survey and evaluate existing methods, showing that they are not durable. We also discuss potential ways to improve their durability and highlight remaining challenges. We hope our work enables future progress on this important problem.

Multi-agent systems have shown strong performance in domains with reliable verifiers such as coding, where multi-parallel candidate generation selected by a verifier is effective. However, such pipelines would not generalize to open-ended, long-horizon research tasks without a verifier. While majority voting or self-consistency is often used to reach consensus as a proxy verifier, parallel agents repeatedly explore the same evidence, while access to peers' partial findings cause search to converge on an early candidate before alternatives are tested. We present ArcticSwarm, a multi-agent research architecture that separates evidence gathering from evidence integration. Subagents publish findings to a shared bulletin board, while gated isolation lets selected search tasks maintain their own prior, preventing early consensus. Structured review at three commitment boundaries enforce only confident candidates to be propagated. As a result, ArcticSwarm reaches 82.6% on the full BrowseComp-Plus set with the open-weight Qwen 3.5-27B model, compared with 78.8% without gated isolation and 74.5% additionally with structured review disabled, outperforming aligned baseline MiroFlow runs (70.6%). Extending to live-web BrowseComp, ArcticSwarm reaches 73.6% with GPT-5, which is well above the reported provider system (54.9%) and MiroFlow (63.4%). Overall, the results show that restricting peer reads during evidence gathering and strengthening commitment boundaries before a hypothesis is shared can broaden search and improve long-horizon multi-agent deep research.

Multimodal automated fact-checking (MAFC) verifies claims by retrieving and reasoning over external evidence. However, most existing static benchmarks risk contamination: they primarily consist of outdated claims verifiable using an LLM's internal knowledge without external evidence. This can inflate performance estimates and fail to reflect true capability on novel claims that require up-to-date information. To address this, emerging dynamic benchmarks collect claims published after LLMs' knowledge cut-off dates, assuming they are uncontaminated. This work revisits this assumption by empirically studying contamination risks in both the state-of-the-art (SOTA) static AVeriTeC benchmark and our newly constructed dynamic ClaimReview2025Q4 benchmark, as well as their impact on MAFC evaluation. Our experiments yield 16 findings, highlighting three key results: (1) Dynamic evaluation reduces but does not eliminate contamination risks, as 17.09\%--29.30\% of post-cut-off claims remain potentially contaminated; (2) Many newly published claims can be verified either directly or by synthesizing multiple pieces of public knowledge available before the cut-off; and (3) Contamination can induce statistically significant inflation in MAFC performance, increasing Macro-F1 by up to 11.34 points and distorting system rankings. In light of these findings, we re-evaluate SOTA LLMs under a strictly contamination-controlled setting. Our study provides practical guidelines for trustworthy MAFC evaluation.

Evaluating large language models (LLMs) today rests on fixed benchmarks that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the boundary, where the per-prompt pass probability is near 0.5 under random-sampling decoding strategy, and propose Dynamic Boundary Evaluation (DBE), which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across 9 reference LLMs; (ii) Skill-Guided Boundary Search (SGBS), a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful request refusal, over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.

Large language models are increasingly deployed as autonomous agents serving users on behalf of companies, placing them in settings where user and deployer interests can conflict. When an agent knows that a user is owed something its deployer would prefer to deny, does it remain honest? Answering this is difficult because false statements can reflect either ignorance or hallucination rather than deception. To address this challenge, we introduce KnownLieBench , a knowledge-verified benchmark that first confirms through a neutral probe that an agent knows a user's entitlement, and then evaluates whether it makes false claims once an incentive to deny that entitlement is introduced. Specifically, KnownLieBench covers eight customer-service domains and 112 grounded cases, conducts multi-round dialogues with a trust-tracking customer agent, and separates deception emerging from incentive alone from deception produced under explicit instruction. Across eighteen proprietary and open-weight models, emergent deception varies substantially across model families and domains. We further use the benchmark for post-training, finding that honesty-directed fine-tuning reduces deception under incentive, while deception-graded fine-tuning increases lie success on honest-control dialogues without increasing lie frequency under incentive. By verifying entitlement knowledge before scoring deceptive behavior, KnownLieBench reduces the confound between lying and not knowing and enables more rigorous auditing and steering of agent honesty.

Cloud-based large language models (LLMs) require users to disclose plaintext data to service providers, creating privacy risks in sensitive domains. Existing privacy-preserving approaches often trade utility for protection, incur substantial computational or communication overhead, remain vulnerable to reconstruction from intermediate representations, or protect only a subset of the training and inference pipeline. We introduce Client-Resolved Generation (CRG), a genera- tion interface that separates server-side generation from the lexical realization of input-derived content. The client transmits only pooled and noise-perturbed rep- resentations, while input-derived output content is represented using request-local positional references and resolved to its original strings only on the client. This interface protects private input and input-derived output content during both train- ing and inference while allowing the service provider to keep its proprietary model parameters hidden from the client. At the same time, exact lexical reuse remains possible without directly exposing the reused content on the provider-visible gen- eration path. We evaluate CRG on medical and document-grounded QA, sensi- tive identifier transfer, and tool calling, together with reconstruction and raw-logit leakage analyses. On SealTools, CRG improves complete-call exact match from 57.3% to 79.9% over the input-privacy framework PPFT, with larger gains as more required output content can be resolved through references. Together, these results show that CRG provides a practical interface for privacy-sensitive cloud LLMs by reducing plaintext exposure across both input and output pathways while preserv- ing task utility and server-side model confidentiality.

Safeguards in deployed LLM services are evaluated by refusal, attack success, and policy violation rates. Those rates characterize how a control performed on the requests it was tested on. A deployment has to answer a different question: how much help with harmful tasks the service still gives an attacker who keeps adapting or finds another way in. We determine what each reported result implies for that question, allowing results from different safeguard families to be compared under one deployment criterion. The evidence requirements are strongly asymmetric. One attack that obtains harmful help from the deployed service suffices to establish that such help remains, and such attacks appear repeatedly in the coded record. Establishing that little remains cannot follow from the safeguard's own numbers alone; it also requires evidence about what the surrounding system still allows after the safeguard performs its local function. Such evidence is supported or derived in only a small minority of the depth-coded claims, and one such claim bounds its scoped residual. A better local score is therefore not, by itself, a stronger claim about the deployment. Safeguard research cannot stop at raising local scores; a gain has to be judged by whether it makes a deployed system any safer.