Daily Papers

Coding-agent evaluation has progressed from resolving individual issues to carrying out long-horizon development, yet task completion is still largely assessed within a single codebase. In software ecosystems, many features and bug fixes require coordinated changes across multiple repositories. We introduce WideSWE to evaluate coding agents on such cross-repository tasks. Mining and reviewing changes across 103 software ecosystems yields 120 real-world tasks, balanced between 60 bug fixes and 60 features. We derive prompts from related issues and pull requests. We systematically review and adapt hidden tests to support diverse correct implementations while preserving required behavior and regression checks. Across seven agent configurations, full task success ranges from 10.83% to 42.50%, with the configuration pairing Codex CLI with GPT-5.6-sol achieving the highest rate. Trajectories show agents failing to identify necessary changes, recognizing changes but leaving them unfinished, or modifying the required repositories without fully satisfying the request. To examine whether working on one repository at a time can alleviate these difficulties, we compare it with joint execution under identical prompts. Independent execution mainly recovers omitted work and is less effective at correcting previously attempted but unsuccessful implementations. Joint execution can use information from related repositories to guide implementation and verification. Code is available at https://github.com/ZJU-ACES-ISE/WideSWE.

Agent skills are portable packages of instructions and resources an agent consults at deployment. Self-evolving them fails in two ways today. First, skills evolved from scratch underperform human-curated ones and, on a weak model, using no skill at all. Second, an evolution-time pass records one lucky trajectory that a fresh stochastic agent often fails to reproduce at deployment. We present reSolve, a per-task, oracle-in-the-loop framework built on three components. It decouples interactive solving from a self-contained deliverable that is independently re-executed in a fresh container, a protocol we call solve-and-reproduce. It enhances the sparse reward signal with a surrogate verifier that cannot access hidden tests or reference answers. It then runs verifier-guided beam search over a solution-construction graph. Within a fixed harness, a cheap model self-evolves skills that reach 74.9% mean-of-3, +14.8 points over the 60.1% human-curated baseline, exceeding the strongest official curated-skill result (67.3%, GPT-5.5/OpenHands). We also report observed failure cases and domain-level results, including performance on the 14 Natural Science tasks, to clarify when the approach does and does not help.

Visible tests are a common gate for LLM-generated code, but passing them does not certify specification correctness. We study a deployment-like monitoring problem: after code has passed public tests, can a weaker LLM verifier identify the residual hidden bugs? We introduce Code Monitor Red Teaming, a monitor-red-teaming protocol that fixes a public-check information boundary while varying generator pressure, verifier scaffolding, and weak-to-strong capability. We instantiate it as CodeMonitorBench, spanning function-level, data-science, and workflow code. Across 71,000 generated candidates, 43,677 pass public tests and 23,081 of those fail hidden tests. Weak verifiers improve with scaffolding and model family, but still miss most hidden bugs at 5% false-positive rate. As a robustness stress test, adversarial public-test-overfit pressure lowers verifier AUROC and raises low-FPR miss rates in most cells. A GLM-5.1 verifier recovers part of the gap under the same evidence boundary; an inferability audit shows that remaining misses mix verifier failures with M1 evidence limits.

Evaluating the real capabilities of large language models in low-resource languages still represents a challenge, as many existing benchmarks focus on widespread tasks translated from English or evaluate only simple language understanding. This paper introduces UA-Code-Bench, a new open-source benchmark established for a thorough evaluation of language models' code generation and competitive programming problem-solving abilities in Ukrainian. The benchmark comprises 500 problems from the Eolymp platform, evenly distributed across five complexity levels from very easy to very hard. A diverse set of 13 leading proprietary and open-source models, generating Python solutions based on a one-shot prompt, was evaluated via the dedicated Eolymp environment against hidden tests, ensuring code correctness. The obtained results reveal that even top-performing models, such as OpenAI o3 and GPT-5, solve only half of the problems, highlighting the challenge of code generation in low-resource natural language. Furthermore, this research presents a comprehensive analysis of performance across various difficulty levels, as well as an assessment of solution uniqueness and computational efficiency, measured by both elapsed time and memory consumption of the generated solutions. In conclusion, this work demonstrates the value of competitive programming benchmarks in evaluating large language models, especially in underrepresented languages. It also paves the way for future research on multilingual code generation and reasoning-enhanced models. The benchmark, data parsing, preparation, code generation, and evaluation scripts are available at https://huggingface.co/datasets/NLPForUA/ua-code-bench.

Coding agents are increasingly asked to build software rather than patch it, yet benchmarks for from-scratch repository construction are mostly limited to a single language and depend on manually curated tasks. We introduce Zero2Repo, a benchmark in which an agent receives a product requirements document, an interface contract, and an empty workspace, and must deliver a complete repository in the project's native ecosystem. Tasks are produced by a language-agnostic authoring pipeline that converts real, version-pinned open-source projects into behavioral specifications, reproducible environments, and hidden acceptance tests. Each task is validated by execution: a reference implementation derived from the upstream project must pass, and adversarial validation must show that the tests reject incorrect implementations. Evaluation runs production coding agents in isolated containers, withholds the acceptance tests until an explicit submission, and assigns a binary reward only when every test passes, with no LLM judge. The pipeline and harness make no language-specific assumptions and apply to mainstream programming ecosystems; the current release contains Python, TypeScript, Go, and C++ tasks. Even on 11 tasks drawn from repositories that frontier models have very likely seen during training, the strongest agent solves only 10, and every failing submission passes 90-99% of the hidden tests; for the two strongest agents, 67-100% of failed tests trace to a single omission or a low-frequency rule stated in the specification rather than to a missing subsystem, so each failure is a concrete target for improvement.

Among the hardest tasks for humans are those found in competitive programming where problems require sophisticated algorithmic thinking, puzzle solving, and the creation of effective code. As a domain to assess language models (LMs), it has not received enough attention, though. This study presents the ICPC benchmark, which consists of 254 international collegiate programming contest (ICPC) tasks. Each problem includes official analysis, reference code, and sample, high-quality unit, and hidden tests. We are able to develop and evaluate a variety of LM inference techniques for competitive programming with these resources. With zero-shot chain-of-thought prompting, we find that o1 only achieves a 19.1\% pass@1 solve rate. With our best inference technique, which combines multi-turn self-judge with reflection and retrieval over episodic information, raises this to 42.2\%. Furthermore, we conduct a new human-in-the-loop investigation to gain a deeper understanding of the remaining difficulties. Surprisingly, we discover that o1 can solve 17 out of 18 problems that were previously unsolvable by any model or technique with just a few specific instructions. A footstep toward LMs with grounded, imaginative, and algorithmic thinking is provided by our quantitative findings and qualitative research. We open-source our code and data at https://github.com/kraritt/zolve.

Coding agents powered by large language models (LLMs) are evolving from making localized code changes to developing complete software repositories. However, evaluating repository-scale generation remains challenging: tasks must demand system-level reasoning while ensuring that all evaluated behaviors are precisely specified and independent of any particular implementation. We introduce E2E-SWE, a benchmark for evaluating whether coding agents can build complete, functional software repositories end to end. E2E-SWE contains 186 whole-repository generation tasks spanning 11 programming languages. Given only a natural-language specification and an empty workspace, an agent must implement a complete, installable project that satisfies a comprehensive suite of hidden tests. Each task is constructed by a software engineer in collaboration with an LLM; together, they develop the test suite and a corresponding implementation-independent specification. To ensure that tasks are well specified and practically solvable, we further subject them to an iterative verification process in which autonomous agents audit and repair task defects using static inspection and failures observed from real model rollouts. Evaluating 13 frontier models, we find substantial variation in end-to-end repository generation ability, with pass@1 ranging from 11.7% to 67.7%, providing strong model differentiation while leaving considerable headroom for future progress. Analysis of agent trajectories further reveals long, front-loaded reasoning patterns, highlighting the planning and system-level reasoning required to construct working codebases from scratch.

Computer-use agents (CUAs) have advanced along two separate lines: graphical interaction and software development through code and the command line. Real digital work requires both, interleaved rather than stacked end to end. We study hybrid CUAs that autonomously decide when to explore an interface, implement software, and run and visually verify their artifacts. We introduce RecreationWorld, a five-platform framework built around recreation: given a running reference, an agent must discover its behavior and build a faithful implementation with no prescribed workflow. RecreationWorld provides reproducible environments on Ubuntu, macOS, Windows, Android, and Web, plus a unified harness with native GUI control and coding tools. The running reference serves as an oracle for hidden behavioral tests, providing execution-grounded rewards. We scale trajectory generation with high-quality open-source applications. Models trained on these trajectories improve across five out-of-distribution coding and hybrid computer-use benchmarks and more frequently verify their rendered outputs, providing evidence of transfer beyond recreation. For held-out evaluation, we introduce RecreationBench, comprising 250 diverse tasks across domains and platforms. Reference-grounded programmatic and visual assertions cover action-conditioned outcomes at multiple interaction depths; each is validated on the reference and by human reviewers before the suite is frozen for automatic scoring. GPT-6 Astra leads at 58.1% overall, but passes all programmatic tests on just 2.8% of tasks. Agents reproduce static interface structure more reliably than interactions and computed outputs, while generated applications remain smaller and more monolithic than their references. We release the benchmark, environments, and test suites.

Coding-agent benchmarks have largely measured whether agents can produce functionally correct patches, but production software also demands measurable speedups on real execution targets. Performance optimization is a distinct agentic task: agents must profile executions, diagnose cross-layer bottlenecks, edit code without breaking correctness, and verify that gains are reproducible rather than measurement artifacts. We introduce PERFOPT-Bench, a benchmark for evaluating this full performance-engineering loop. Each task provides a correct but deliberately suboptimal codebase and asks the agent to improve a target performance metric; scoring requires hidden correctness tests, verified-speedup measurement, and trajectory-level audit. We evaluate 7 agent stacks with different LLMs and agent frameworks on 7 long-horizon optimization tasks. The results show that optimization performance is workload-dependent rather than determined by model identity alone: no single stack dominates, and changing the agent framework can materially change the same LLM's per-task speedup profile. We further find that raw speedup is unsafe as a benchmark score, since some large gains arise from benchmark-specific shortcut exploitation; an exploratory relay pilot suggests that restarting from an externalized optimization summary can recover additional headroom after an initial session stops. The benchmark and our evaluation are available at: https://anonymous.4open.science/r/Dataset-D3CC.

Modern software systems accumulate technical debt over decades of development, which makes migration expensive and largely manual. As coding agents become increasingly capable at bug fixing, can they autonomously perform such migrations? Existing benchmarks cannot answer this question because they evaluate only behavioural correctness, not whether the migration actually occurred. This leads an easy hack: agents copy the original implementation to make tests pass. We call this Blindness. To address this problem, we introduce SWE Refactor Bench, a benchmark comprising 20 whole-repository migrations, covering 4 kinds of technical debt. A three-stage evaluation protocol measures both migration completeness and behavioural correctness. (1) Migration Audit verifies that the migration occurred. (2) Behavioural Tests measure correctness with a fixed test suite. (3) Agentic Verification uses 6 independent coding agents to generate targeted tests for hidden behavioural differences. Across 520 runs from 8 frontier models and 26 model-effort configurations, only 28 of 520 runs (5.4%) pass all three stages, 13 of the 20 tasks receive no accepted solution, and the best model (claude-opus-5) scores 47.0/100. Migration completeness and behavioural correctness are distinct abilities: a few runs preserve behaviour by skipping the migration and are stopped at Migration Audit; most attempt it and break behaviour, and are stopped at Behavioural Tests. Agents cannot deliver a perfect migration: among the 340 runs that pass Migration Audit, 58% reach 99% of the fixed checks, yet only 26% reach 100%. Agent capability differs across migration categories: agents score 31.4 on build toolchain rewrites but only 5.6 on language rewrites. Together, these findings position SWE Refactor Bench as a rigorous testbed for developing coding agents for reliable whole-repository migrations.

LLM-based agents are increasingly deployed for software maintenance tasks such as automated program repair (APR). APR agents automatically fetch GitHub issues and use backend LLMs to generate patches that fix the reported bugs. However, existing work primarily focuses on the functional correctness of APR-generated patches, whether they pass hidden or regression tests, while largely ignoring potential security risks. Given the openness of platforms like GitHub, where any user can raise issues and participate in discussions, an important question arises: Can an adversarial user submit a valid issue on GitHub that misleads an LLM-based agent into generating a functionally correct but vulnerable patch? To answer this question, we propose SWExploit, which generates adversarial issue statements designed to make APR agents produce patches that are functionally correct yet vulnerable. SWExploit operates in three main steps: (1) program analysis to identify potential injection points for vulnerable payloads; (2) adversarial issue generation to provide misleading reproduction and error information while preserving the original issue semantics; and (3) iterative refinement of the adversarial issue statements based on the outputs of the APR agents. Empirical evaluation on three agent pipelines and five backend LLMs shows that SWExploit can produce patches that are both functionally correct and vulnerable (the attack success rate on the correct patch could reach 0.91, whereas the baseline ASRs are all below 0.20). Based on our evaluation, we are the first to challenge the traditional assumption that a patch passing all tests is inherently reliable and secure, highlighting critical limitations in the current evaluation paradigm for APR agents.

Diffusion language models form answers through a sequence of revisable continuous states, but multi-agent systems usually discard those states and make agents communicate only after decoding text. We ask whether one independent diffusion trajectory can pass a compact hidden message into another and, crucially, whether the receiver uses what that message means. In Hidden Messages, a sender compresses its denoising state into four signed dense vectors; the receiver consumes them while retaining its own context, canvas, and trajectory. Removing a message tests whether the channel matters; replacing it with a message from another example tests whether its content matters. Across controlled tasks, derangement removes a +63.6-point advantage on Tiny-A2D and a +100.0-point advantage on Dream-7B, while targeted substitutions steer the receiver toward the substituted fact. On QASC with the two supporting facts split across agents, matched communication reaches 80.8%, but a wrong-example message lowers accuracy to 46.8%, a content effect of +34.0 points. The surprising result is temporal: one late transfer reaches 87.8%, exceeding four recurrent exchanges by +7.0 points. These experiments provide causal evidence for continuous hidden communication between diffusion-LM agents and show that when information is transferred can matter more than how many times agents exchange it.

Retrieval-augmented generation (RAG) systems put more and more emphasis on grounding their responses in user-generated content found on the Web, amplifying both their usefulness and their attack surface. Most notably, indirect prompt injection and retrieval poisoning attack the web-native carriers that survive ingestion pipelines and are very concerning. We provide OpenRAG-Soc, a compact, reproducible benchmark-and-harness for web-facing RAG evaluation under these threats, in a discrete data package. The suite combines a social corpus with interchangeable sparse and dense retrievers and deployable mitigations - HTML/Markdown sanitization, Unicode normalization, and attribution-gated answered. It standardizes end-to-end evaluation from ingestion to generation and reports attacks time of one of the responses at answer time, rank shifts in both sparse and dense retrievers, utility and latency, allowing for apples-to-apples comparisons across carriers and defenses. OpenRAG-Soc targets practitioners who need fast, and realistic tests to track risk and harden deployments.

Position bias in multiple-choice LLM evaluation is widely cited as a confound in capability comparisons, but published measurements rely on single answer-order shuffles whose results confound the bias signal with content-level noise and sampling stochasticity. I introduce inspect_permute, an open-source extension to the inspect_ai evaluation framework that runs exhaustive answer-order permutations per question and reports the chi-squared / Cramer V signature of position bias with bootstrap confidence intervals. I apply the tool across four vendors (gpt-4o-mini, claude-haiku-4-5, gemini-2.5-flash, grok-3) on five MMLU subjects, 24,000 API calls under temperature-0 generation, with falsifier predictions pre-registered via a public SHA-256 hash before half the data was observed. Position bias turns out to be statistically detectable only within a roughly 60-95% base-accuracy Goldilocks zone. Below it, processing-load dominance swamps subject-specific signal; above it, ceiling effects compress the variance below the chi-squared test resolution. Detectable cells separate into two mechanism types: monotone A-to-D decrease (processing_load, in low-tier models) and non-monotone D-drop (content_ambiguity, in a narrow capability band). Standard MMLU places every frontier-tier model above the detection band, so absence of signal there should be read as not measurable, not unbiased. Together with the ceiling-effect characterisation in arXiv:2606.26185, this work brackets the detectable region of position-bias measurement and makes the field central question askable in a verifiable form. Package, data, preregistration under MIT.

Long-term memory promises LLM agents that grow more capable across sessions, maintaining an accurate, evolving understanding of the user that interaction forms. In practice, however, this memory is evaluated mostly through downstream behavior, such as later answers, personalization quality, or task success, which tests that understanding only indirectly and leaves the memory artifact itself largely unaudited. We argue that long-term memory should instead be evaluated as an auditable post-interaction artifact: after ordinary assistance, what structured user state can be reconstructed from the memory the agent leaves behind? We instantiate this view in MEMPROBE, a benchmark in which a memory-equipped agent assists simulated users, each carrying a hidden, taxonomy-anchored user-state bank, across a trajectory of leak-controlled tasks, after which that bank is reconstructed from the agent's resulting memory under both full-store and top-k access. Built on synthetic ground truth for efficient, scalable measurement, MEMPROBE spans 50 simulated users with 31 hidden dimensions each (1,550 recovery targets) and tests 5 representative memory systems. Testing state-of-the-art memory agents, we find that successful assistance and recoverable memory behave as distinct capabilities. Task completion nearly saturates, even for a memoryless baseline, while category-balanced recovery stays moderate (about 0.6) and drops further under top-k retrieval. MEMPROBE is the first benchmark to study memory recovery directly, reconstructing the user state a system retains and scoring it against ground truth. We see recovery as a concrete objective for future memory agents to optimize, and MEMPROBE as a step toward an environment where agents are trained to remember their users, growing more faithful the longer they know them.

The commitment to single-precision floating-point arithmetic is widespread in the deep learning community. To evaluate whether this commitment is justified, the influence of computing precision (single and double precision) on the optimization performance of the Conjugate Gradient (CG) method (a second-order optimization algorithm) and RMSprop (a first-order algorithm) has been investigated. Tests of neural networks with one to five fully connected hidden layers and moderate or strong nonlinearity with up to 4 million network parameters have been optimized for Mean Square Error (MSE). The training tasks have been set up so that their MSE minimum was known to be zero. Computing experiments have disclosed that single-precision can keep up (with superlinear convergence) with double-precision as long as line search finds an improvement. First-order methods such as RMSprop do not benefit from double precision. However, for moderately nonlinear tasks, CG is clearly superior. For strongly nonlinear tasks, both algorithm classes find only solutions fairly poor in terms of mean square error as related to the output variance. CG with double floating-point precision is superior whenever the solutions have the potential to be useful for the application goal.

Recent advances in multimodal foundation models have made them capable generalist physical agents for a range of manipulation tasks. However, successful operation in an unfamiliar environment may require an agent to seek task-relevant information through interaction when it is absent from the observations: it may need to determine where a relevant object is, inspect an unobserved property, or discover the effect of an unfamiliar tool. We thus introduce RoboQuest, a benchmark for goal-directed embodied exploration, where agents must actively acquire task-relevant information through physical interaction, use the resulting evidence to adapt subsequent actions, and autonomously decide when to commit to task completion. RoboQuest comprises ten mobile manipulation tasks centered on three forms of uncertainty: search, manipulation-based inspection, and interactive testing. We evaluate five frontier multimodal agents through a common visuomotor interface, as well as a π_{0.5} policy fine-tuned on the full-episode demonstrations we release. The best agent succeeds in only 23\% of the episodes, and the fine-tuned policy almost never succeeds. Isolated tests of the execution skills the tasks are built from, with the hidden information supplied, show that the agents can carry out most of the required actions, and our failure analysis attributes only a minority of the failures to execution. Our failure analysis further finds that the agents often stop exploring too early as they make decisions before observing the required evidence for task completion. We also find that agents rarely prevent or repair the disturbances caused by their exploration. Moreover, learning by trial and error remains difficult for most models.

We present Test-Driven AI Agent Definition (TDAD), a methodology that treats agent prompts as compiled artifacts: engineers provide behavioral specifications, a coding agent converts them into executable tests, and a second coding agent iteratively refines the prompt until tests pass. Deploying tool-using LLM agents in production requires measurable behavioral compliance that current development practices cannot provide. Small prompt changes cause silent regressions, tool misuse goes undetected, and policy violations emerge only after deployment. To mitigate specification gaming, TDAD introduces three mechanisms: (1) visible/hidden test splits that withhold evaluation tests during compilation, (2) semantic mutation testing via a post-compilation agent that generates plausible faulty prompt variants, with the harness measuring whether the test suite detects them, and (3) spec evolution scenarios that quantify regression safety when requirements change. We evaluate TDAD on SpecSuite-Core, a benchmark of four deeply-specified agents spanning policy compliance, grounded analytics, runbook adherence, and deterministic enforcement. Across 24 independent trials, TDAD achieves 92% v1 compilation success with 97% mean hidden pass rate; evolved specifications compile at 58%, with most failed runs passing all visible tests except 1-2, and show 86-100% mutation scores, 78% v2 hidden pass rate, and 97% regression safety scores. The implementation is available as an open benchmark at https://github.com/f-labs-io/tdad-paper-code.

A coding agent solving a software-engineering task spends dozens of steps reasoning, editing code, and running tests, yet little is known about what the underlying language model internally represents about the program it is working on. We show that the residual streams of language models under coding agents linearly encode properties of the evolving program: a logistic-regression probe on hidden states is able to decode whether the current code parses, passes its test suite, reduces the number of failing tests, and introduces regressions, reaching AUC up to 0.83 for correctness across two models and two benchmarks. Our second finding is more surprising: these representations run ahead of the agent's own edits. Probes trained to predict the outcome of future edits (before they are materialized and written on disk) achieve performance above chance up to roughly 25 steps in advance. We call this the agent's latent programming horizon. As a proof of external validity, we show that the probes transfer across benchmarks without retraining. Our positive results open calls for more research in mechanistic interpretability of coding agents.

Large language models (LLMs) handle many tasks with one set of parameters, but under KV-cached inference it is unclear what task-general structure, if any, is used at decode time rather than during prefill. We propose DecodeShare, a protocol that identifies a low-dimensional subspace consistently shared across tasks in decode-time hidden states, and then tests its causal role by removing that subspace only during decoding. In our experiments, disturbing the discovered shared subspace degrades decision performance far more than disturbing either a prefill-derived or random subspace under the same intervention budget. We further show this decode-shared subspace has practical consequences for activation steering: common steering directions can overlap the task-general decode channel. Projecting out this shared subspace directly separates the functional roles of the two components, while evaluating steering vectors at decode-time yields more reliable signal for downstream deployment than prefill-based proxies. Despite its compactness, the shared subspace can serve as a high-leverage causal channel at decode time. Code is available at: https://github.com/Zishan-Shao/decodeshare.git.

AI coding agents now write much of the code that ships, yet whether that code is secure remains unclear. We introduce the Last Secure Code Benchmark, a comprehensive evaluation of secure code generation by coding agents. Built from 150 real vulnerabilities, its 450 tasks span seven widely used languages (Java, C, C++, JavaScript, Python, PHP, and Go), 24 CWEs including 21 of the 2025 CWE Top 25 and 8 of the Top 10 KEV weaknesses, more than any benchmark we compare with, and three vibe-coding scenarios: filling in a function, writing a file, and building a repository's source tree from scratch. Every task asks only for functionality, through a specification that never mentions security, and hidden functional and security tests grade the result. Across six frontier agents, 73.6% to 91.3% of tasks pass functionally, yet even the strongest, Claude Fable 5, produces code that is both functional and secure for only 25.1% of tasks, and the others for 16.4% to 24.7%. The best model only passes 12% of the vulnerability scenarios across all vibe coding setting, posing strong concern on the security of the code written by vibe coding.

SWE-Bench Pro has emerged as a standard benchmark for evaluating software engineering agents on challenging repository-level tasks. However, our analysis work show that its evaluation is undermined by two sources of unreliability: reward hacking, enabled by leakage of gold solutions or hidden evaluation information, and task quality issues, including misleading problem statements and improperly scoped tests. These issues can inflate benchmark performance and obscure agents' true coding ability. We present SWE-Bench Pro Verified, a verified version of SWE-Bench Pro that addresses both problems. Our approach combines anti-hacking safeguards that eliminate major leakage channels without disrupting normal agent functionality, with task refinement that minimally corrects inconsistencies within flawed instances. Evaluations on SWE-Bench Pro Verified reveal that some models perform substantially worse than previously reported, suggesting that existing results on SWE-Bench Pro may overestimate real software engineering capability. SWE-Bench Pro Verified offers a more trustworthy benchmark for assessing software engineering agents.

Terminal-agent benchmarks have become a primary signal for measuring the coding and system-administration capabilities of large language models. As the market for evaluation environments grows, so does the pressure to ship tasks quickly, often without thorough adversarial review of the verification logic. This paper is a guideline for writing good benchmark tasks, drawn from over a year of contributing to and reviewing tasks for Terminal Bench. Most people write benchmark tasks the way they write prompts. They shouldn't. A prompt is designed to help the agent succeed; a benchmark is designed to find out if it can. We argue that good tasks are adversarial, difficult, and legible, and that a large class of common failure modes -- AI-generated instructions, over-prescriptive specifications, clerical difficulty, oracle solutions that assume hidden knowledge, tests that validate the wrong things, and reward-hackable environments -- are predictable consequences of treating task authoring as prompt authoring. We catalog these failure modes, argue that real difficulty is conceptual rather than environmental, and discuss recent empirical evidence that over 15% of tasks in popular terminal-agent benchmarks are reward-hackable. We hope this serves as a useful reference for benchmark maintainers, task contributors, and researchers using benchmark scores as evidence.

In software testing, a set of test cases is constructed according to some predefined selection criteria. The software is then examined against these test cases. Three interesting observations have been made on the current artifacts of software testing. Firstly, an error-revealing test case is considered useful while a successful test case which does not reveal software errors is usually not further investigated. Whether these successful test cases still contain useful information for revealing software errors has not been properly studied. Secondly, no matter how extensive the testing has been conducted in the development phase, errors may still exist in the software [5]. These errors, if left undetected, may eventually cause damage to the production system. The study of techniques for uncovering software errors in the production phase is seldom addressed in the literature. Thirdly, as indicated by Weyuker in [6], the availability of test oracles is pragmatically unattainable in most situations. However, the availability of test oracles is generally assumed in conventional software testing techniques. In this paper, we propose a novel test case selection technique that derives new test cases from the successful ones. The selection aims at revealing software errors that are possibly left undetected in successful test cases which may be generated using some existing strategies. As such, the proposed technique augments the effectiveness of existing test selection strategies. The technique also helps uncover software errors in the production phase and can be used in the absence of test oracles.

Large language models can hold knowledge they do not report. A model may sandbag on a capability evaluation, or answer against what it internally knows, and its outputs alone cannot tell whether it is hiding an answer or simply does not have one. We borrow the Concealed Information Test, a forensic method that identifies guilty knowledge by presenting a suspect with the true detail among plausible decoys and measuring a stronger response to the item they recognize. Our method, Probe of Internal Recognition (PIR), does the same inside a model. It presents a question with its candidate answers and reads, from the model's internal states, which candidate the model recognizes as correct. PIR is reference-free, needing no honest reference model and no labeled truth corpus. Across eight models from five families (Gemma, Qwen, Llama, Mistral, and Phi), PIR recovers the recognized answer at 0.70 to 0.87 balanced accuracy, well above the 0.28 to 0.40 unknown-item baseline and the 0.25 chance rate. It stays readable across every form of concealment we test, from prompted deception and trained sandbagging to external password-locked and circuit-broken checkpoints, with recognition between 0.85 and 0.93. When the model hides a known answer, recognition stays high. When unlearning removes the knowledge, recognition drops to the level of a question the model never knew. PIR therefore separates a model that will not answer from one that cannot, which supports sandbagging audits and unlearning verification. The signal is causal, adds information beyond black-box behavioral cues, and extends from multiple-choice questions to free-form generation.

Machine learning models for medical image analysis often suffer from poor performance on important subsets of a population that are not identified during training or testing. For example, overall performance of a cancer detection model may be high, but the model still consistently misses a rare but aggressive cancer subtype. We refer to this problem as hidden stratification, and observe that it results from incompletely describing the meaningful variation in a dataset. While hidden stratification can substantially reduce the clinical efficacy of machine learning models, its effects remain difficult to measure. In this work, we assess the utility of several possible techniques for measuring and describing hidden stratification effects, and characterize these effects on multiple medical imaging datasets. We find evidence that hidden stratification can occur in unidentified imaging subsets with low prevalence, low label quality, subtle distinguishing features, or spurious correlates, and that it can result in relative performance differences of over 20% on clinically important subsets. Finally, we explore the clinical implications of our findings, and suggest that evaluation of hidden stratification should be a critical component of any machine learning deployment in medical imaging.

Too many defective compute chips are escaping existing manufacturing tests -- at least an order of magnitude more than industrial targets across all compute chip types in data centers. Silent data corruptions (SDCs) caused by test escapes, when left unaddressed, pose a major threat to reliable computing. We present a three-pronged approach outlining future directions for overcoming test escapes: (a) Quick diagnosis of defective chips directly from system-level incorrect behaviors. Such diagnosis is critical for gaining insights into why so many defective chips escape existing manufacturing testing. (b) In-field detection of defective chips. (c) New test experiments to understand the effectiveness of new techniques for detecting defective chips. These experiments must overcome the drawbacks and pitfalls of previous industrial test experiments and case studies.

Verifying the final answer of a long reasoning trace may show that the trace is wrong but it does not distinguish three cases. The problem may lie beyond the model's capability, another attempt would succeed, or one step in an otherwise recoverable trace made the trace fail. Standard hidden-state probes are trained on final-answer correctness, a label that conflates these cases. Monte Carlo rollouts distinguish the cases and locate the failing step at additional inference cost. We test the hidden states of 4B and 8B reasoning models for the failure type and the failing step and compare each probe with the strongest baseline that uses no hidden states. On incorrect 8B traces, the confidence probe, a calibrated correctness probe at a midpoint cutoff, misses 81% of cases in which a single step makes the trace unrecoverable and 47% of cases in which another attempt would succeed. A probe with a different training design over the same states reverses the ordering of these two kinds of failure, so either ordering can be read from those states. At both scales, no single tested probe locates the failing step more accurately than a rule based only on trace length, and at 4B a probe trained to predict that step fails to outperform this rule in any of 10 runs on identical traces. The confidence probe therefore indicates that an answer is wrong, but its failure-type ordering depends on its training design.

In science and medicine, model interpretations may be reported as discoveries of natural phenomena or used to guide patient treatments. In such high-stakes tasks, false discoveries may lead investigators astray. These applications would therefore benefit from control over the finite-sample error rate of interpretations. We reframe black box model interpretability as a multiple hypothesis testing problem. The task is to discover "important" features by testing whether the model prediction is significantly different from what would be expected if the features were replaced with uninformative counterfactuals. We propose two testing methods: one that provably controls the false discovery rate but which is not yet feasible for large-scale applications, and an approximate testing method which can be applied to real-world data sets. In simulation, both tests have high power relative to existing interpretability methods. When applied to state-of-the-art vision and language models, the framework selects features that intuitively explain model predictions. The resulting explanations have the additional advantage that they are themselves easy to interpret.

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.

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.

A memory test for a video world model should say whether the scene was set, whether the asked change was carried out, and whether the last frame shows the result. Current tests fold these into one number, so a low score is easy to read as forgetting. When the change is not carried out, the hidden state transition cannot appear, so the test is not really a test of memory. We introduce MIST, an evaluation that separates the change from the memory it should leave behind. The model first sets the scene while the change stays hidden, and only later is asked to show it, with leakage before that point kept under control. This separation locates the bottleneck: the model sets the scene but does not carry out the instructed change. The same gap appears in general memory evaluations. On that basis we propose M3-Agent (Make the Missing Move), which recursively revises the details of the input instruction so that the model successfully carries out the requested instruction. Relative to the baseline, it more than triples the MIST reveal score. On the benchmarks, WRBench's score T rises by 26.1% and MBench-T's score U by 52.7%. The method meets the problem from the execution side: it shows what the frozen model can do once the move is stated, and it sets a stricter baseline.

Large language models often answer complex reasoning questions without revealing intermediate steps, raising whether they reason latently or complete patterns. We propose the Hidden CoT Detection Score (HCDS), a comparative behavioral and mechanistic signal measuring whether neutral-prompt behavior aligns more closely with explicit CoT or explicit no- CoT. Here, hidden CoT operationally denotes this neutral-prompt CoT-like alignment; HCDS does not directly observe or prove an unexposed reasoning trace. On GSM8K, HCDS is significantly positive for both Qwen3-4B variants (Thinking +1.87, p = 1.2 times 10^{-7}; Instruct +1.41, p = 1.9 times 10^{-4}), replicates across a different inference stack and quantization within 0.08 (+1.80 and +1.45), and is not significantly positive in seven of eight length-adjusted calibration-control cells. The unadjusted score produces large positive scores on single-step arithmetic and numeric factual lookup. The variants also respond differently to no-CoT instructions: Instruct complies from the prompt alone, whereas Thinking continues reasoning and requires intervention. These findings show stronger, less prompt-conditional CoT-like behavior in the reasoning-tuned model, consistent with but not proof of latent reasoning. HCDS thus investigates latent reasoning without relying on models' self-reported traces.

Developing an educational test can be expensive and time-consuming, as each item must be written by experts and then evaluated by collecting hundreds of student responses. Moreover, many tests require multiple distinct sets of questions administered throughout the school year to closely monitor students' progress, known as parallel tests. In this study, we focus on tests of silent sentence reading efficiency, used to assess students' reading ability over time. To generate high-quality parallel tests, we propose to fine-tune large language models (LLMs) to simulate how previous students would have responded to unseen items. With these simulated responses, we can estimate each item's difficulty and ambiguity. We first use GPT-4 to generate new test items following a list of expert-developed rules and then apply a fine-tuned LLM to filter the items based on criteria from psychological measurements. We also propose an optimal-transport-inspired technique for generating parallel tests and show the generated tests closely correspond to the original test's difficulty and reliability based on crowdworker responses. Our evaluation of a generated test with 234 students from grades 2 to 8 produces test scores highly correlated (r=0.93) to those of a standard test form written by human experts and evaluated across thousands of K-12 students.

Recent advances in automated test generation utilises language models to produce unit tests. While effective, language models tend to generate many incorrect tests with respect to both syntax and semantics. Although such incorrect tests can be easily detected and discarded, they constitute a "missed opportunity" -- if fixed, they are often valuable as they directly add testing value (they effectively target the underlying program logic to be tested) and indirectly form good seeds for generating additional tests. To this end, we propose a simple technique for repairing some of these incorrect tests through a combination of rule-based static analysis and re-prompting. We evaluate this simple approach, named YATE, on a set of 6 open-source projects and show that it can effectively produce tests that cover on average 32.06% more lines and kill 21.77% more mutants than a plain LLM-based method. We also compare YATE with four other LLM-based methods, namely HITS, SYMPROMPT, TESTSPARK and COVERUP and show that it produces tests that cover substantially more code. YATE achieves 22% higher line coverage, 20% higher branch coverage and kill 20% more mutants at a comparable cost (number of calls to LLMs).

We present a generic framework for creating differentially private versions of any hypothesis test in a black-box way. We analyze the resulting tests analytically and experimentally. Most crucially, we show good practical performance for small data sets, showing that at epsilon = 1 we only need 5-6 times as much data as in the fully public setting. We compare our work to the one existing framework of this type, as well as to several individually-designed private hypothesis tests. Our framework is higher power than other generic solutions and at least competitive with (and often better than) individually-designed tests.

Large language models are trained on vast amounts of internet data, prompting concerns and speculation that they have memorized public benchmarks. Going from speculation to proof of contamination is challenging, as the pretraining data used by proprietary models are often not publicly accessible. We show that it is possible to provide provable guarantees of test set contamination in language models without access to pretraining data or model weights. Our approach leverages the fact that when there is no data contamination, all orderings of an exchangeable benchmark should be equally likely. In contrast, the tendency for language models to memorize example order means that a contaminated language model will find certain canonical orderings to be much more likely than others. Our test flags potential contamination whenever the likelihood of a canonically ordered benchmark dataset is significantly higher than the likelihood after shuffling the examples. We demonstrate that our procedure is sensitive enough to reliably prove test set contamination in challenging situations, including models as small as 1.4 billion parameters, on small test sets of only 1000 examples, and datasets that appear only a few times in the pretraining corpus. Using our test, we audit five popular publicly accessible language models for test set contamination and find little evidence for pervasive contamination.

AI has achieved remarkable mastery over games such as Chess, Go, and Poker, and even Jeopardy, but the rich variety of standardized exams has remained a landmark challenge. Even in 2016, the best AI system achieved merely 59.3% on an 8th Grade science exam challenge. This paper reports unprecedented success on the Grade 8 New York Regents Science Exam, where for the first time a system scores more than 90% on the exam's non-diagram, multiple choice (NDMC) questions. In addition, our Aristo system, building upon the success of recent language models, exceeded 83% on the corresponding Grade 12 Science Exam NDMC questions. The results, on unseen test questions, are robust across different test years and different variations of this kind of test. They demonstrate that modern NLP methods can result in mastery on this task. While not a full solution to general question-answering (the questions are multiple choice, and the domain is restricted to 8th Grade science), it represents a significant milestone for the field.

One of the critical phases in software development is software testing. Testing helps with identifying potential bugs and reducing maintenance costs. The goal of automated test generation tools is to ease the development of tests by suggesting efficient bug-revealing tests. Recently, researchers have leveraged Large Language Models (LLMs) of code to generate unit tests. While the code coverage of generated tests was usually assessed, the literature has acknowledged that the coverage is weakly correlated with the efficiency of tests in bug detection. To improve over this limitation, in this paper, we introduce MuTAP for improving the effectiveness of test cases generated by LLMs in terms of revealing bugs by leveraging mutation testing. Our goal is achieved by augmenting prompts with surviving mutants, as those mutants highlight the limitations of test cases in detecting bugs. MuTAP is capable of generating effective test cases in the absence of natural language descriptions of the Program Under Test (PUTs). We employ different LLMs within MuTAP and evaluate their performance on different benchmarks. Our results show that our proposed method is able to detect up to 28% more faulty human-written code snippets. Among these, 17% remained undetected by both the current state-of-the-art fully automated test generation tool (i.e., Pynguin) and zero-shot/few-shot learning approaches on LLMs. Furthermore, MuTAP achieves a Mutation Score (MS) of 93.57% on synthetic buggy code, outperforming all other approaches in our evaluation. Our findings suggest that although LLMs can serve as a useful tool to generate test cases, they require specific post-processing steps to enhance the effectiveness of the generated test cases which may suffer from syntactic or functional errors and may be ineffective in detecting certain types of bugs and testing corner cases PUTs.

A myriad of different Large Language Models (LLMs) face a common challenge in contextually analyzing table question-answering tasks. These challenges are engendered from (1) finite context windows for large tables, (2) multi-faceted discrepancies amongst tokenization patterns against cell boundaries, and (3) various limitations stemming from data confidentiality in the process of using external models such as gpt-3.5-turbo. We propose a cooperative game dubbed "HiddenTables" as a potential resolution to this challenge. In essence, "HiddenTables" is played between the code-generating LLM "Solver" and the "Oracle" which evaluates the ability of the LLM agents to solve Table QA tasks. This game is based on natural language schemas and importantly, ensures the security of the underlying data. We provide evidential experiments on a diverse set of tables that demonstrate an LLM's collective inability to generalize and perform on complex queries, handle compositional dependencies, and align natural language to programmatic commands when concrete table schemas are provided. Unlike encoder-based models, we have pushed the boundaries of "HiddenTables" to not be limited by the number of rows - therefore we exhibit improved efficiency in prompt and completion tokens. Our infrastructure has spawned a new dataset "PyQTax" that spans across 116,671 question-table-answer triplets and provides additional fine-grained breakdowns & labels for varying question taxonomies. Therefore, in tandem with our academic contributions regarding LLMs' deficiency in TableQA tasks, "HiddenTables" is a tactile manifestation of how LLMs can interact with massive datasets while ensuring data security and minimizing generation costs.

Probes are the workhorse of interpretability. If a model's hidden states predict a variable, the model is said to represent it. But a probe score has no fixed meaning. An R^2 of 0.6 may only reflect what the input already gives away, and the same score can mean different things on different data. We propose reading every probe score against two reference points: a floor, what a declared set of simple inputs already predicts, and a ceiling, what the full input can predict. The gap between them, the headroom, is the range in which a probe can show that a model computes something beyond the simple inputs. We prove that headroom vanishes in two ways: the target stops depending on a hidden variable the model must infer, or the input stops revealing it. We test this on transformers trained for in-context meta-analysis, which must infer the hidden heterogeneity between studies to weight them correctly, and where both reference points are known. Under distribution shift, probe scores fall and prediction error rises 12--15times, yet the model recovers a similar share of the headroom, indicating that the data lost information, not the representation. We then analyze the real models. The single-cell foundation model scGPT encodes biological variability only partially. We also revisit four influential LLM probing studies, which claim that models represent geography, the state of an Othello board, truth, and the demographics of their users. Against a floor computed from the input text alone, some of these claims hold, while others are largely explained by the text itself.

This work presents LURK (Latent UnleaRned Knowledge), a novel framework that probes for hidden retained knowledge in unlearned LLMs through adversarial suffix prompting. LURK automatically generates adversarial prompt suffixes designed to elicit residual knowledge about the Harry Potter domain, a commonly used benchmark for unlearning. Our experiments reveal that even models deemed successfully unlearned can leak idiosyncratic information under targeted adversarial conditions, highlighting critical limitations of current unlearning evaluation standards. By uncovering latent knowledge through indirect probing, LURK offers a more rigorous and diagnostic tool for assessing the robustness of unlearning algorithms. All code will be publicly available.

As long-horizon coding agents produce more code than any developer can review, oversight collapses onto a single surface: the automated test suite. Reward hacking naturally arises in this setup, as the agent optimizes for passing tests while deviating from the users true goal. We study this reward hacking phenomenon by decompose software engineering tasks into three parts: (i) a natural language description of the specification (ii) visible validation tests that exercise specified features in isolation, and (iii) held-out tests that compose those same features to simulate real-world usage. Based on the specification and the visible validation test suites, a genuine agent would be able to generate a solution that can also pass all of the held-out tests. Therefore we use the gap in pass rates on these two suites to quantify reward hacking. Based on this methodology, we introduce SpecBench, a benchmark comprising 30 systems-level programming tasks ranging from short horizon tasks like building a JSON parser to ultra long horizon tasks like building an entire OS kernel from scratch. Large-scale experiments reveal a consistent pattern: while every frontier agent saturates the visible suite, reward hacking persists, with smaller models exhibiting larger gaps on holdout suites. The gap also scales sharply with task length: it grows by 28 percentage points for every tenfold increase in code size. Failures range from subtle feature isolation to deliberate exploits, including a 2,900-line hash-table "compiler" that memorizes test inputs. SpecBench offers a principled testbed for measuring whether coding agents build genuine working systems or merely game the test suites developers hand them.

Software testing is a critical, yet resource-intensive phase of the software development lifecycle. Over the years, various automated tools have been developed to aid in this process. Search-based approaches typically achieve high coverage but produce tests with low readability, whereas large language model (LLM)-based methods generate more human-readable tests but often suffer from low coverage and compilability. While the majority of research efforts have focused on improving test coverage and readability, little attention has been paid to enhancing the robustness of bug detection, particularly in exposing corner cases and vulnerable execution paths. To address this gap, we propose AdverTest, a novel adversarial framework for LLM-powered test case generation. AdverTest comprises two interacting agents: a test case generation agent (T) and a mutant generation agent (M). These agents engage in an adversarial loop, where M persistently creates new mutants "hacking" the blind spots of T's current test suite, while T iteratively refines its test cases to "kill" the challenging mutants produced by M. This interaction loop is guided by both coverage and mutation scores, enabling the system to co-evolve toward both high test coverage and bug detection capability. Experimental results in the Defects4J dataset show that our approach improves fault detection rates by 8.56% over the best existing LLM-based methods and by 63.30% over EvoSuite, while also improving line and branch coverage.

Can a language model read the quality of its ongoing computation, and can an external intervention turn that readout into better outcomes? We test both questions in a frozen 2.6B looped transformer, Ouro-RLTT. On GSM8K, a strict pre-answer probe excludes the answer region and gold value yet predicts success: hidden states plus length/log-probability features reach AUROC 0.797 versus 0.731 for those surface features alone (increment +0.066; task-clustered 95% CI [+0.021,+0.112]; 170 tasks). On Horizon Logic, a prospectively extended task-disjoint study gives an increment of +0.111 (CI [+0.056,+0.169]), independently replicated on the new cohort (+0.095) and robust to an adversarial malformed-sibling shortcut. Recurrence also moves candidate-quality readability to progressively earlier physical depth; the trend replicates across the Ouro family and qualitatively in out-of-family Huginn, although their transfer geometry differs. The readout converts into validated decision-level gains. Hidden-state-based scores improve risk-coverage over shortcut-only scores in four sealed selective-prediction arms, and terminal selection beats matched random even when every candidate is well formed (27/32 correct selections versus 64.8% expected; p = 0.0086). Generative control does not convert: directional steering is negative, a branch screen is bounded, and exact-compute loop allocation and minimal LoRA direction-binding detect no gain. These tests run through bit-exact branch/carry/prune machinery over Ouro's 192-slot recurrent cache, including a suffix-recompute splice saving up to 88% of per-branch layer passes. We call this decision-usable but not generatively controllable property operational proto-introspection. All load-bearing values use source-item-disjoint splits and antisymmetrized pairwise evaluation.

Software vulnerabilities are often detected via taint analysis, penetration testing, or fuzzing. They are also found via unit tests that exercise security-sensitive behavior with specific inputs, called vulnerability-witnessing tests. Generative AI models could help developers in writing them, but they require many examples to learn from, which are currently scarce. This paper introduces VuTeCo, an AI-driven framework for collecting examples of vulnerability-witnessing tests from Java repositories. VuTeCo carries out two tasks: (1) The "Finding" task to determine whether a unit test case is security-related, and (2) the "Matching" task to relate a test case to the vulnerability it witnesses. VuTeCo addresses the Finding task with UniXcoder, achieving an F0.5 score of 0.73 and a precision of 0.83 on a test set of unit tests from Vul4J. The Matching task is addressed using DeepSeek Coder, achieving an F0.5 score of 0.65 and a precision of 0.75 on a test set of pairs of unit tests and vulnerabilities from Vul4J. VuTeCo has been used in the wild on 427 Java projects and 1,238 vulnerabilities, obtaining 224 test cases confirmed to be security-related and 35 tests correctly matched to 29 vulnerabilities. The validated tests were collected in a new dataset called Test4Vul. VuTeCo lays the foundation for large-scale retrieval of vulnerability-witnessing tests, enabling future AI models to better understand and generate security unit tests.

Monitoring chain-of-thought (CoT) reasoning is a foundational safety technique for large language model (LLM) agents; however, this oversight is compromised if models learn to conceal their reasoning. We explore the potential for steganographic CoT -- where models hide secret reasoning within innocuous text -- to inform risk assessment and deployment policies. We systematically evaluate the limits of steganographic capabilities across 28 models, ranging from past generations to the current frontier. We measure monitor evasion, refusal rates, encoding fidelity, and hidden task accuracy across four datasets, comparing steganographic acrostics against plain reasoning and filler-token baselines. We find that current models cannot yet sustain hidden reasoning for complex math and arithmetic tasks. However, in a simplified counting experiment, Claude Opus 4.5 achieved 92% accuracy on the hidden task, demonstrating nascent capability. Notably, in rare cases (<1%), GPT-5.2 might refuse steganographic instructions while simultaneously complying with them. Our findings underscore the need for continuous evaluation of steganographic risks. This study provides a methodology to preemptively detect and prevent hidden reasoning that might empower misaligned scheming and deceptive behavior.

We discover that DALLE-2 seems to have a hidden vocabulary that can be used to generate images with absurd prompts. For example, it seems that Apoploe vesrreaitais means birds and Contarra ccetnxniams luryca tanniounons (sometimes) means bugs or pests. We find that these prompts are often consistent in isolation but also sometimes in combinations. We present our black-box method to discover words that seem random but have some correspondence to visual concepts. This creates important security and interpretability challenges.

Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, 5.7 times 10^2 or 580?" This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers). Furthermore, the hidden state after reading a pair of numerals encodes their ranking, with a linear classifier achieving over 90% accuracy. Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective. Finally, we show that incorporating the classifier probe's log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models' internal magnitude representations can enhance their numerical reasoning capabilities.

As language models become more powerful and sophisticated, it is crucial that they remain trustworthy and reliable. There is concerning preliminary evidence that models may attempt to deceive or keep secrets from their operators. To explore the ability of current techniques to elicit such hidden knowledge, we train a Taboo model: a language model that describes a specific secret word without explicitly stating it. Importantly, the secret word is not presented to the model in its training data or prompt. We then investigate methods to uncover this secret. First, we evaluate non-interpretability (black-box) approaches. Subsequently, we develop largely automated strategies based on mechanistic interpretability techniques, including logit lens and sparse autoencoders. Evaluation shows that both approaches are effective in eliciting the secret word in our proof-of-concept setting. Our findings highlight the promise of these approaches for eliciting hidden knowledge and suggest several promising avenues for future work, including testing and refining these methods on more complex model organisms. This work aims to be a step towards addressing the crucial problem of eliciting secret knowledge from language models, thereby contributing to their safe and reliable deployment.

This paper describes a testing methodology for quantitatively assessing the risk that rare or unique training-data sequences are unintentionally memorized by generative sequence models---a common type of machine-learning model. Because such models are sometimes trained on sensitive data (e.g., the text of users' private messages), this methodology can benefit privacy by allowing deep-learning practitioners to select means of training that minimize such memorization. In experiments, we show that unintended memorization is a persistent, hard-to-avoid issue that can have serious consequences. Specifically, for models trained without consideration of memorization, we describe new, efficient procedures that can extract unique, secret sequences, such as credit card numbers. We show that our testing strategy is a practical and easy-to-use first line of defense, e.g., by describing its application to quantitatively limit data exposure in Google's Smart Compose, a commercial text-completion neural network trained on millions of users' email messages.

Large Language Models (LLMs) excel in reasoning and generation across domains, but still struggle with identifying and diagnosing complex errors. This stems mainly from training objectives that prioritize correct answers, limiting exposure to and learning from errors. While recent studies have begun to address this by introducing error signals, most rely on shallow, static errors, restricting improvement in deep diagnostic ability. To overcome this, we propose Hide and Seek Game (HSG), a dynamic adversarial framework for error generation and diagnosis, and evaluate it on mathematical problem-solving. HSG involves two adversarial roles: Sneaky, which "hides" by generating subtle, deceptive reasoning errors, and Diagnosis, which "seeks" to accurately detect them. Through adversarial co-evolution, both error stealth and diagnostic precision are enhanced. Experiments on several math reasoning tasks show that HSG significantly boosts error diagnosis, achieving 16.8\%--31.4\% higher accuracy than baselines like GPT-4o. We also release a challenging dataset of deceptive errors and diagnostic annotations as a benchmark for future research.

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

The tendency to find and exploit "shortcuts" to complete tasks poses significant risks for reliable assessment and deployment of large language models (LLMs). For example, an LLM agent with access to unit tests may delete failing tests rather than fix the underlying bug. Such behavior undermines both the validity of benchmark results and the reliability of real-world LLM coding assistant deployments. To quantify, study, and mitigate such behavior, we introduce ImpossibleBench, a benchmark framework that systematically measures LLM agents' propensity to exploit test cases. ImpossibleBench creates "impossible" variants of tasks from existing benchmarks like LiveCodeBench and SWE-bench by introducing direct conflicts between the natural-language specification and the unit tests. We measure an agent's "cheating rate" as its pass rate on these impossible tasks, where any pass necessarily implies a specification-violating shortcut. As a practical framework, ImpossibleBench is not just an evaluation but a versatile tool. We demonstrate its utility for: (1) studying model behaviors, revealing more fine-grained details of cheating behaviors from simple test modification to complex operator overloading; (2) context engineering, showing how prompt, test access and feedback loop affect cheating rates; and (3) developing monitoring tools, providing a testbed with verified deceptive solutions. We hope ImpossibleBench serves as a useful framework for building more robust and reliable LLM systems. Our implementation can be found at https://github.com/safety-research/impossiblebench.

We present our on-going effort of constructing a large-scale benchmark for face forgery detection. The first version of this benchmark, DeeperForensics-1.0, represents the largest face forgery detection dataset by far, with 60,000 videos constituted by a total of 17.6 million frames, 10 times larger than existing datasets of the same kind. Extensive real-world perturbations are applied to obtain a more challenging benchmark of larger scale and higher diversity. All source videos in DeeperForensics-1.0 are carefully collected, and fake videos are generated by a newly proposed end-to-end face swapping framework. The quality of generated videos outperforms those in existing datasets, validated by user studies. The benchmark features a hidden test set, which contains manipulated videos achieving high deceptive scores in human evaluations. We further contribute a comprehensive study that evaluates five representative detection baselines and make a thorough analysis of different settings.

In this work, we address the TrackRAD2025 challenge of real-time tumor tracking in cine-MRI sequences of the thoracic and abdominal regions under strong data scarcity constraints. Two complementary strategies were explored: (i) unsupervised registration with the IMPACT similarity metric and (ii) foundation model-based segmentation leveraging SAM 2.1 and its recent variants through prompt-based interaction. Due to the one-second runtime constraint, the SAM-based method was ultimately selected. The final configuration used SAM2.1 b+ with mask-based prompts from the first annotated slice, fine-tuned solely on the small labeled subset from TrackRAD2025. Training was configured to minimize overfitting, using 1024x1024 patches (batch size 1), standard augmentations, and a balanced Dice + IoU loss. A low uniform learning rate (0.0001) was applied to all modules (prompt encoder, decoder, Hiera backbone) to preserve generalization while adapting to annotator-specific styles. Training lasted 300 epochs (~12h on RTX A6000, 48GB). The same inference strategy was consistently applied across all anatomical sites and MRI field strengths. Test-time augmentation was considered but ultimately discarded due to negligible performance gains. The final model was selected based on the highest Dice Similarity Coefficient achieved on the validation set after fine-tuning. On the hidden test set, the model reached a Dice score of 0.8794, ranking 6th overall in the TrackRAD2025 challenge. These results highlight the strong potential of foundation models for accurate and real-time tumor tracking in MRI-guided radiotherapy.

Accurate estimation of brain infarction (i.e., irreversibly damaged tissue) is critical for guiding treatment decisions in acute ischemic stroke. Reliable infarct prediction informs key clinical interventions, including the need for patient transfer to comprehensive stroke centers, the potential benefit of additional reperfusion attempts during mechanical thrombectomy, decisions regarding secondary neuroprotective treatments, and ultimately, prognosis of clinical outcomes. This work introduces the Ischemic Stroke Lesion Segmentation (ISLES) 2024 challenge, which focuses on the prediction of final infarct volumes from pre-interventional acute stroke imaging and clinical data. ISLES24 provides a comprehensive, multimodal setting where participants can leverage all clinically and practically available data, including full acute CT imaging, sub-acute follow-up MRI, and structured clinical information, across a train set of 150 cases. On the hidden test set of 98 cases, the top-performing model, a multimodal nnU-Net-based architecture, achieved a Dice score of 0.285 (+/- 0.213) and an absolute volume difference of 21.2 (+/- 37.2) mL, underlining the significant challenges posed by this task and the need for further advances in multimodal learning. This work makes two primary contributions: first, we establish a standardized, clinically realistic benchmark for post-treatment infarct prediction, enabling systematic evaluation of multimodal algorithmic strategies on a longitudinal stroke dataset; second, we analyze current methodological limitations and outline key research directions to guide the development of next-generation infarct prediction models.

RECIST diameter measurements are widely used for tumor response assessment, but they provide only a limited 2D description of lesion extent. We present LETT-NeXt, a lightweight RECIST-guided model that predicts 3D lesion masks from CT volumes and RECIST markers for the CVPR 2026 Foundation Models for Pan-cancer Segmentation in CT Images competition. LETT-NeXt extracts a RECIST-centered regional crop, encodes the RECIST line and endpoints as two prompt channels, and concatenates them with the CT input. A compact MedNeXt-v2 encoder--decoder predicts the lesion mask, followed by prompt-aware component selection and adaptive AutoZoom inference. On the public validation set, LETT-NeXt achieved a Dice Similarity Coefficient (DSC) of 79.4 pm 10.1 and a Normalized Surface Dice (NSD) of 72.3 pm 16.2. On the hidden test set, it achieved a DSC of 73.9 and an NSD of 67.3, corresponding to a challenge score of 70.6\%. On the public validation mirror, LETT-NeXt completed CPU inference in 6.9 pm 3.0 s per case with a peak memory use of 3.6 GB. Code is available at github.com/Ahus-AIM/lett-next.

Given questions regarding some prototypical situation such as Name something that people usually do before they leave the house for work? a human can easily answer them via acquired experiences. There can be multiple right answers for such questions, with some more common for a situation than others. This paper introduces a new question answering dataset for training and evaluating common sense reasoning capabilities of artificial intelligence systems in such prototypical situations. The training set is gathered from an existing set of questions played in a long-running international game show FAMILY- FEUD. The hidden evaluation set is created by gathering answers for each question from 100 crowd-workers. We also propose a generative evaluation task where a model has to output a ranked list of answers, ideally covering all prototypical answers for a question. After presenting multiple competitive baseline models, we find that human performance still exceeds model scores on all evaluation metrics with a meaningful gap, supporting the challenging nature of the task.

While large language models have shown impressive capabilities across a wide range of domains, they still encounter significant challenges in reasoning tasks that require gathering evidence over multiple turns and drawing logical conclusions. These challenges present significant obstacles for LLM chat user interfaces, which rely on multi-turn interactions to facilitate effective collaboration. This limitation leads to real-world issues; for example, service chatbots must gather necessary information from customers over multiple turns to diagnose and resolve problems effectively. Despite the multi-turn nature of many real-world LLM use cases, most existing benchmarks rely on carefully curated single-turn tests, which often blur the line between memorization and genuine reasoning. To address this, we introduce the Wason Inductive Logic Test (WILT), a simple yet challenging multi-turn reasoning benchmark designed to resist memorization. WILT is inspired by the Wason 2-4-6 task, where participants must infer a boolean function involving three variables (e.g., x < y < z) by proposing test cases (such as (2, 4, 6)). In WILT, each test starts from a clean slate, with only the initial instructions provided, preventing models from relying on pre-learned responses. Over several turns, models must interact with the environment by suggesting test cases to narrow the possible hypotheses and ultimately infer the hidden function based on the outcomes. Our findings reveal that LLMs struggle with this task, exhibiting distinct strengths and weaknesses: some are better at narrowing down the hypothesis space by proposing valuable test cases, while others are more adept at deducing the hidden function from observed cases. Despite these variations, the best-performing model achieves only 28% accuracy, highlighting a significant gap in LLM performance on complex multi-turn reasoning tasks.

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.

A stereotype is an over-generalized belief about a particular group of people, e.g., Asians are good at math or Asians are bad drivers. Such beliefs (biases) are known to hurt target groups. Since pretrained language models are trained on large real world data, they are known to capture stereotypical biases. In order to assess the adverse effects of these models, it is important to quantify the bias captured in them. Existing literature on quantifying bias evaluates pretrained language models on a small set of artificially constructed bias-assessing sentences. We present StereoSet, a large-scale natural dataset in English to measure stereotypical biases in four domains: gender, profession, race, and religion. We evaluate popular models like BERT, GPT-2, RoBERTa, and XLNet on our dataset and show that these models exhibit strong stereotypical biases. We also present a leaderboard with a hidden test set to track the bias of future language models at https://stereoset.mit.edu

The automated analysis of the aortic vessel tree (AVT) from computed tomography angiography (CTA) holds immense clinical potential, but its development has been impeded by a lack of shared, high-quality data. We launched the SEG.A. challenge to catalyze progress in this field by introducing a large, publicly available, multi-institutional dataset for AVT segmentation. The challenge benchmarked automated algorithms on a hidden test set, with subsequent optional tasks in surface meshing for computational simulations. Our findings reveal a clear convergence on deep learning methodologies, with 3D U-Net architectures dominating the top submissions. A key result was that an ensemble of the highest-ranking algorithms significantly outperformed individual models, highlighting the benefits of model fusion. Performance was strongly linked to algorithmic design, particularly the use of customized post-processing steps, and the characteristics of the training data. This initiative not only establishes a new performance benchmark but also provides a lasting resource to drive future innovation toward robust, clinically translatable tools.

Transformer-based models generate hidden states that are difficult to interpret. In this work, we analyze hidden states and modify them at inference, with a focus on motion forecasting. We use linear probing to analyze whether interpretable features are embedded in hidden states. Our experiments reveal high probing accuracy, indicating latent space regularities with functionally important directions. Building on this, we use the directions between hidden states with opposing features to fit control vectors. At inference, we add our control vectors to hidden states and evaluate their impact on predictions. Remarkably, such modifications preserve the feasibility of predictions. We further refine our control vectors using sparse autoencoders (SAEs). This leads to more linear changes in predictions when scaling control vectors. Our approach enables mechanistic interpretation as well as zero-shot generalization to unseen dataset characteristics with negligible computational overhead.

When AI systems explain their reasoning step-by-step, practitioners often assume these explanations reveal what actually influenced the AI's answer. We tested this assumption by embedding hints into questions and measuring whether models mentioned them. In a study of over 9,000 test cases across 11 leading AI models, we found a troubling pattern: models almost never mention hints spontaneously, yet when asked directly, they admit noticing them. This suggests models see influential information but choose not to report it. Telling models they are being watched does not help. Forcing models to report hints works, but causes them to report hints even when none exist and reduces their accuracy. We also found that hints appealing to user preferences are especially dangerous-models follow them most often while reporting them least. These findings suggest that simply watching AI reasoning is not enough to catch hidden influences.

The rapid progress in synthetic image generation and manipulation has now come to a point where it raises significant concerns for the implications towards society. At best, this leads to a loss of trust in digital content, but could potentially cause further harm by spreading false information or fake news. This paper examines the realism of state-of-the-art image manipulations, and how difficult it is to detect them, either automatically or by humans. To standardize the evaluation of detection methods, we propose an automated benchmark for facial manipulation detection. In particular, the benchmark is based on DeepFakes, Face2Face, FaceSwap and NeuralTextures as prominent representatives for facial manipulations at random compression level and size. The benchmark is publicly available and contains a hidden test set as well as a database of over 1.8 million manipulated images. This dataset is over an order of magnitude larger than comparable, publicly available, forgery datasets. Based on this data, we performed a thorough analysis of data-driven forgery detectors. We show that the use of additional domainspecific knowledge improves forgery detection to unprecedented accuracy, even in the presence of strong compression, and clearly outperforms human observers.

Multimodal Foundation Models (MMFMs) have shown remarkable performance on various computer vision and natural language processing tasks. However, their performance on particular tasks such as document understanding is still limited. They also require more compute, time, and engineering resources to finetune and deploy compared to traditional, unimodal models. In this report, we present Multimodal Structured Generation, a general framework which constrains the output logits of frozen MMFMs to force them to reason before responding with structured outputs that downstream APIs can parse and use. We provide a detailed account of our approach, including the technical details, theoretical discussions, and final evaluation results in the 2nd Multimodal Foundation Models Challenge hosted by the Computer Vision and Pattern Recognition (CVPR) conference. Our approach achieved the second highest score in the hidden test set for Phase 2 and third highest overall. This shows the method's ability to generalize to unseen tasks. And that simple engineering can beat expensive & complicated modelling steps as we first discussed in our paper, Retrieval Augmented Structured Generation: Business Document Information Extraction as Tool Use. All of our scripts, deployment steps, and evaluation results can be accessed in https://github.com/leloykun/MMFM-Challenge

We present a complete software pipeline for revealing the hidden texts of the Herculaneum papyri using X-ray CT images. This enhanced virtual unwrapping pipeline combines machine learning with a novel geometric framework linking 3D and 2D images. We also present EduceLab-Scrolls, a comprehensive open dataset representing two decades of research effort on this problem. EduceLab-Scrolls contains a set of volumetric X-ray CT images of both small fragments and intact, rolled scrolls. The dataset also contains 2D image labels that are used in the supervised training of an ink detection model. Labeling is enabled by aligning spectral photography of scroll fragments with X-ray CT images of the same fragments, thus creating a machine-learnable mapping between image spaces and modalities. This alignment permits supervised learning for the detection of "invisible" carbon ink in X-ray CT, a task that is "impossible" even for human expert labelers. To our knowledge, this is the first aligned dataset of its kind and is the largest dataset ever released in the heritage domain. Our method is capable of revealing accurate lines of text on scroll fragments with known ground truth. Revealed text is verified using visual confirmation, quantitative image metrics, and scholarly review. EduceLab-Scrolls has also enabled the discovery, for the first time, of hidden texts from the Herculaneum papyri, which we present here. We anticipate that the EduceLab-Scrolls dataset will generate more textual discovery as research continues.

Transformer-based decoders for 3D instance segmentation typically commit to a fixed number of queries and positional modeling calibrated on the training distribution rather than on the scene at hand. Indoor scans vary widely in spatial extent and object count, so a fixed query set over-initializes small scenes and under-initializes large ones, while learned absolute and relative encodings are bound to the training scenes' extents and can saturate. We present AQ3D, which is designed to handle scenes of various sizes during training and inference. Queries are instantiated at a fixed ratio of the scene's superpoints, forming an overcomplete set whose background rejection is entirely left to the decoder. Positional information is encoded using 3D RoPE over quantized metric coordinates, replacing learned bounded lookup tables of prior decoders. Further, we improve the decoder itself by using attribution-based superpoint pooling, a mask refinement branch, and a cosine classifier for background rejection. Experiments show our method sets a new state-of-the-art on validation and hidden test splits across the datasets ScanNetV2, ScanNet200, and ScanNet++V2 among decoder methods trained without additional data augmentation. Code is available at https://github.com/kenomo/aq3d{github.com/kenomo/aq3d}.

Machine unlearning has emerged as a prevalent technical solution for selectively removing unwanted knowledge absorbed during pre-training, without requiring full retraining. While recent unlearning techniques can effectively remove undesirable content without severely compromising performance on standard benchmarks, we find that they may inadvertently create ``knowledge holes'' -- unintended losses of benign knowledge that standard benchmarks fail to capture. To probe where unlearned models reveal knowledge holes, we propose a test case generation framework that explores both immediate neighbors of unlearned content and broader areas of potential failures. Our evaluation demonstrates significant hidden costs of unlearning: up to 98.7\% of the test cases yield irrelevant or nonsensical responses from unlearned models, despite being answerable by the pretrained model. These findings necessitate rethinking the conventional approach to evaluating knowledge preservation in unlearning, moving beyond standard, static benchmarks.

This paper describes Meta's ACH system for mutation-guided LLM-based test generation. ACH generates relatively few mutants (aka simulated faults), compared to traditional mutation testing. Instead, it focuses on generating currently undetected faults that are specific to an issue of concern. From these currently uncaught faults, ACH generates tests that can catch them, thereby `killing' the mutants and consequently hardening the platform against regressions. We use privacy concerns to illustrate our approach, but ACH can harden code against {\em any} type of regression. In total, ACH was applied to 10,795 Android Kotlin classes in 7 software platforms deployed by Meta, from which it generated 9,095 mutants and 571 privacy-hardening test cases. ACH also deploys an LLM-based equivalent mutant detection agent that achieves a precision of 0.79 and a recall of 0.47 (rising to 0.95 and 0.96 with simple pre-processing). ACH was used by Messenger and WhatsApp test-a-thons where engineers accepted 73% of its tests, judging 36% to privacy relevant. We conclude that ACH hardens code against specific concerns and that, even when its tests do not directly tackle the specific concern, engineers find them useful for their other benefits.

Outdoor LiDAR semantic scene completion (SSC) recovers a dense semantic voxel grid from a scan observing 1% of the target volume, under class imbalance beyond 7,000x. We recast SSC as generative semantic scene completion (GSSC): a single discrete-diffusion formulation in three roles. First, paired sparse-dense scene synthesis (PS^3) generates matched sparse LiDAR observations with their dense semantic completions, addressing the long tail at its source and yielding the PS^3-SemanticKITTI corpus we train on alongside SemanticKITTI. Second, semantic-guided generative scene completion (SGSC) generates the scene from noise with multinomial discrete diffusion, conditioned on the sparse scan through a bird's-eye-view semantic map and a sparse 3D feature stream. Third, the same framework instead refines an existing completion in one flow-matching step: structured source discrete diffusion (S^2D^2). S^2D^2 improves the mIoU of SGSC's own output and every external SSC base tested, without base retraining or test-time adaptation. On the strongest base, one step without test-time augmentation reaches 38.8% mIoU on the SemanticKITTI hidden test. To our knowledge that is the best causal, single-sweep, single-sample result on that leaderboard, +2.1 pp over the previous best published score under the same restriction. Four correction steps with eight-view test-time augmentation reach 39.2%, outside that restriction.

Semantic segmentation of street panoramas can support detailed descriptions of urban environments, yet small datasets and unequal training costs make model selection difficult. This paper presents the system used for a first place submission to the PalmCity challenge in the leaderboard snapshot dated 5 October 2026. Nine pretrained segmentation systems are compared using approximately equal computation budgets. The candidates include DeepLabV3+, SegFormer, UPerNet, Mask2Former, DINOv3 with a linear decoder, and an Encoder only Mask Transformer using DINOv3. The two leading candidates are trained independently with three random seeds and longer budgets. Equal averaging of class probabilities from the three Encoder only Mask Transformer models, evaluated at three image scales with horizontal reflection, produces 60.95% mean intersection over union and 71.16% mean F1 on the 84 image public validation split. The submitted predictions receive 57.08% mean intersection over union and 67.96% mean F1 on the hidden test leaderboard. Producing all 249 test masks takes 251.49 seconds including model initialization and provenance checks on one NVIDIA RTX 5090. Peak allocated GPU memory is 2.70 GiB. The study reports all eligible models, all inference variants, class level errors, source conditions, and reproducibility checks, providing a documented challenge workflow with existing architectures.

Existing question answering (QA) datasets are no longer challenging to most powerful Large Language Models (LLMs). Traditional QA benchmarks like TriviaQA, NaturalQuestions, ELI5 and HotpotQA mainly study ``known unknowns'' with clear indications of both what information is missing, and how to find it to answer the question. Hence, good performance on these benchmarks provides a false sense of security. A yet unmet need of the NLP community is a bank of non-factoid, multi-perspective questions involving a great deal of unclear information needs, i.e. ``unknown uknowns''. We claim we can find such questions in search engine logs, which is surprising because most question-intent queries are indeed factoid. We present Researchy Questions, a dataset of search engine queries tediously filtered to be non-factoid, ``decompositional'' and multi-perspective. We show that users spend a lot of ``effort'' on these questions in terms of signals like clicks and session length, and that they are also challenging for GPT-4. We also show that ``slow thinking'' answering techniques, like decomposition into sub-questions shows benefit over answering directly. We release sim 100k Researchy Questions, along with the Clueweb22 URLs that were clicked.

Accurate segmentation of teeth and pulp in Cone-Beam Computed Tomography (CBCT) is vital for clinical applications like treatment planning and diagnosis. However, this process requires extensive expertise and is exceptionally time-consuming, highlighting the critical need for automated algorithms that can effectively utilize unlabeled data. In this paper, we propose U-Mamba2-SSL, a novel semi-supervised learning framework that builds on the U-Mamba2 model and employs a multi-stage training strategy. The framework first pre-trains U-Mamba2 in a self-supervised manner using a disruptive autoencoder. It then leverages unlabeled data through consistency regularization, where we introduce input and feature perturbations to ensure stable model outputs. Finally, a pseudo-labeling strategy is implemented with a reduced loss weighting to minimize the impact of potential errors. U-Mamba2-SSL achieved an average score of 0.789 and a DSC of 0.917 on the hidden test set, achieving first place in Task 1 of the STSR 2025 challenge. The code is available at https://github.com/zhiqin1998/UMamba2.

The production of food, feed, fiber, and fuel is a key task of agriculture, which has to cope with many challenges in the upcoming decades, e.g., a higher demand, climate change, lack of workers, and the availability of arable land. Vision systems can support making better and more sustainable field management decisions, but also support the breeding of new crop varieties by allowing temporally dense and reproducible measurements. Recently, agricultural robotics got an increasing interest in the vision and robotics communities since it is a promising avenue for coping with the aforementioned lack of workers and enabling more sustainable production. While large datasets and benchmarks in other domains are readily available and enable significant progress, agricultural datasets and benchmarks are comparably rare. We present an annotated dataset and benchmarks for the semantic interpretation of real agricultural fields. Our dataset recorded with a UAV provides high-quality, pixel-wise annotations of crops and weeds, but also crop leaf instances at the same time. Furthermore, we provide benchmarks for various tasks on a hidden test set comprised of different fields: known fields covered by the training data and a completely unseen field. Our dataset, benchmarks, and code are available at https://www.phenobench.org.

Lossless data compression remains central to computer science, with direct impact on storage, communication bandwidth, computational cost, and energy consumption. It is also closely related to Algorithmic Information Theory, where compressibility provides an operational measure of structure and non-randomness. This paper presents the 2026 Algorithmic Information Theory Data Compression Challenge, a benchmark for evaluating general-purpose lossless compressors under realistic constraints. Submissions were encouraged to use arithmetic or range coding, limited to at most 8 GB of memory, and required to include a decompressor no larger than 1 MB. The benchmark comprised sixteen heterogeneous files, split into public training and hidden testing datasets. In total, 117 valid submitted compressors were evaluated alongside established reference compressors using compression ratio, compression and decompression time, Weissman score, and Pareto-frontier analysis. The results show that performance depends strongly on the optimization criterion: fast compressors achieved the best speed-oriented scores, whereas modelling-intensive compressors produced smaller outputs at higher computational cost. A Normalized Compression Distance analysis further revealed clusters of related submissions and distinguished incremental variants from more independent implementations. Selected submissions were described for their methodological novelty or competitive performance and further tested on four large external datasets, where several achieved competitive or superior results relative to established compressors. Overall, the challenge confirms the importance of probabilistic modelling, hidden testing, and external datasets for assessing compression performance and generalization. Benchmark resources, leaderboard data, binaries, and selected source code are publicly available at https://aitdcc.github.io.

Portable ultra-low-field (ULF) MRI can expand access to pediatric neuroimaging, but segmentation at 0.064 T remains challenging because anatomical boundaries are weakly delineated, small structures may be only partially visible, and high-field references can be locally misregistered. The LISA 2026 Challenge provides two non-equivalent annotations reflecting different sources of anatomical evidence: a highfield-derived (HF) mask defining the scored target and a low-field-edited (LF) mask aligned with visible ULF anatomy. In this challenge report, we describe AURA, an nnU-Net-based asymmetric supervision strategy that treats these annotations as distinct observations rather than interchangeable ground truths. AURA anchors training to the HF mask and incorporates the LF mask through a bounded reliability gate based on label disagreement, boundaries, predictive uncertainty, class reliability, and training stage. On a 16-case development split, the HF-supervised baseline, AURA, and their ensemble achieved Dice scores of 0.7984, 0.7950, and 0.7988, respectively, while the ensemble achieved an HD95 of 1.8892 and an ASSD of 0.7855. These results provide a preliminary evaluation of AURA within the LISA 2026 Challenge and motivate further assessment on the hidden test set and external ULF cohorts. Our code and pretrained models are available at https://github.com/minhdang050806/ A-nnU-Net-based-asymmetric-supervision-strategy.

Most existing methods realize 3D instance segmentation by extending those models used for 3D object detection or 3D semantic segmentation. However, these non-straightforward methods suffer from two drawbacks: 1) Imprecise bounding boxes or unsatisfactory semantic predictions limit the performance of the overall 3D instance segmentation framework. 2) Existing method requires a time-consuming intermediate step of aggregation. To address these issues, this paper proposes a novel end-to-end 3D instance segmentation method based on Superpoint Transformer, named as SPFormer. It groups potential features from point clouds into superpoints, and directly predicts instances through query vectors without relying on the results of object detection or semantic segmentation. The key step in this framework is a novel query decoder with transformers that can capture the instance information through the superpoint cross-attention mechanism and generate the superpoint masks of the instances. Through bipartite matching based on superpoint masks, SPFormer can implement the network training without the intermediate aggregation step, which accelerates the network. Extensive experiments on ScanNetv2 and S3DIS benchmarks verify that our method is concise yet efficient. Notably, SPFormer exceeds compared state-of-the-art methods by 4.3% on ScanNetv2 hidden test set in terms of mAP and keeps fast inference speed (247ms per frame) simultaneously. Code is available at https://github.com/sunjiahao1999/SPFormer.

Conducting contamination-free evaluation of mathematical capabilities can be difficult for two reasons: models may memorize a test set once it is made public, and current mathematical benchmarks are prone to overfitting due to having limited diversity of symbols and rules, coupled with closed-ended answers. This paper proposes a method to leverage these shortcomings as useful features to a construct dynamic, counterfactual benchmark, which can be used to both reveal overfitting and measure true reasoning. We demonstrate this via MatheMagic, which generates math test instances with the interpretations of numbers and operators altered, yet has automatically verifiable answers. Test instances are randomly seeded and constructed at test time to evaluate a model's induction or deduction capability, offering stability, extensibility, comparability, and robustness to overfitting. Our experiments find that models solve deduction more easily than induction, but they revert to standard math. Further analysis reveals that math-adapted models fail to exhibit a general "skill" of reasoning, and fine-tuning on induction tasks generalizes poorly.

Unlearning in large language models (LLMs) is typically evaluated at the output level, where a model appears to suppress sensitive or undesirable content. In this work, we show that such evaluations can create an illusion of forgetting: even when output-level leakage is eliminated, sensitive information can remain encoded in the model's hidden representations. We first provide a theoretical analysis establishing a fundamental separation between output suppression and representational erasure. Specifically, we show that the decoder can be made arbitrarily insensitive to sensitive directions, driving output-level leakage to zero, while the hidden representations retain the underlying information. To empirically validate this phenomenon, we train generative probe decoders on hidden states across transformer layers, enabling layer-wise measurement of information leakage. Across three widely used benchmarks, TOFU, MUSE, and WMDP, and state-of-the-art unlearning methods, we find that substantial sensitive information remains recoverable from hidden representations, even when standard output-level metrics indicate successful unlearning. To address this gap, we propose Probe-Adversarial Representation Suppression (PARS), an unlearning objective that adversarially minimizes the extractable information from hidden representations. PARS directly targets representational leakage and provides significantly stronger guarantees of erasure under adversarial probing and relearning attacks, outperforming all evaluated baselines. Our results highlight a fundamental limitation of existing unlearning paradigms and suggest that true forgetting in LLMs requires controlling not only model outputs, but also the information encoded in hidden representations. Codes are available at https://github.com/OptimAI-Lab/HiddenStateUnlearning.

Clinical decision support is moving toward committees of language-model agents deliberating on a shared workspace. We ask whether such committees can be gamed by shortcuts, cues a benchmark rewards but a clinician would ignore. Across seven cohorts on six public datasets spanning text (MedQA-USMLE, MedMCQA, MIMIC-CXR reports), imaging (NIH ChestX-ray14, MIMIC-CXR-JPG, CheXpert) and tabular ICU records (SUPPORT2), Gemini committees resist these cues in isolation (flip 5-16%), yet a socially plausible shortcut spreads: when two peers assert the same wrong answer, the holdout under test adopts it in 38% of cases, as does a false "pre-screen" system flag, on both capability tiers. Of three oversight agents, a gate cannot separate adoption from honest agreement (false-positive rate 100%); a same-lineage judge reading only the transcript flags adoption on text (precision 100%, recall 93%) but collapses onto the gate in imaging; a referee that privately re-queries the holdout transfers to imaging (77-88% precision, 13-21% false-positive rate). Tripling a cue's visual salience does not move contagion, whereas a second peer voice raises it by half again. Gaming a hidden rubric is near-silent: only 1/10 text and 1/134 imaging drifters name the rubric they moved toward. What games a committee is social plausibility, and only a referee independent of self-report catches it. Code: https://github.com/criticaldata/benchmaxxing

Developers and quality assurance testers often rely on manual testing to test accessibility features throughout the product lifecycle. Unfortunately, manual testing can be tedious, often has an overwhelming scope, and can be difficult to schedule amongst other development milestones. Recently, Large Language Models (LLMs) have been used for a variety of tasks including automation of UIs, however to our knowledge no one has yet explored their use in controlling assistive technologies for the purposes of supporting accessibility testing. In this paper, we explore the requirements of a natural language based accessibility testing workflow, starting with a formative study. From this we build a system that takes as input a manual accessibility test (e.g., ``Search for a show in VoiceOver'') and uses an LLM combined with pixel-based UI Understanding models to execute the test and produce a chaptered, navigable video. In each video, to help QA testers we apply heuristics to detect and flag accessibility issues (e.g., Text size not increasing with Large Text enabled, VoiceOver navigation loops). We evaluate this system through a 10 participant user study with accessibility QA professionals who indicated that the tool would be very useful in their current work and performed tests similarly to how they would manually test the features. The study also reveals insights for future work on using LLMs for accessibility testing.

Reproducibility is essential for scientific research, yet prior work shows that LLM outputs vary with hardware and batching. We identify an overlooked factor: the hidden injection of the current date into system prompts, which users cannot control and which changes every day. Across 9 recent LLMs and 6 datasets spanning multiple-choice QA (MCQA), math reasoning, code generation, and machine translation, performance varies solely with the current date, with deltas of up to 6% on MCQA, 14% on math reasoning, 7% on code generation, and 2.84 BLEU on machine translation. Model rankings also shift, affecting leaderboards. This date effect exceeds other sources of non-determinism, such as batch size and numerical precision. Standard prompting techniques -- chain-of-thought and few-shot prompting -- do not reduce the sensitivity; chain-of-thought even amplifies it. Our findings underscore the need for careful evaluation protocols to ensure reproducibility and fair comparisons in LLM research.

RL for code correctness is now established: have the model generate a program, run it against hidden test cases, and reward solutions that pass. Extending this to code optimization seems straightforward: just add execution time to the reward. But in practice, once timing drives the reward, small problems in measurement noise, reward sparsity, or GRPO instability overwhelm the signal and make RL fail: generated solutions are barely faster, and more of them can fail. We make execution time learnable through three stages: (1) how code is tested, by building DMC-Optim with large optimization tests and a calibrated sandbox; (2) how speed is turned into reward, by composing correctness and speed in the RL environment and using an offline simulator to predict the most promising configurations; and (3) how the model learns from that reward, by adapting GRPO and evaluation to the sparser, noisier timed-execution setting. On DMC-Optim, the strongest optimization-aware configurations improve strict top-50% pass@1 from 18.0% to 31.3% on Qwen 2.5 7B and from 30.7% to 50.4% on CWM 32B. These gains further increase at stricter percentiles such as top-30%, with 125% relative improvement for CWM 32B, while preserving pure-correctness scores. When the timing sandbox is degraded, robust optimization RL reaches 100% to 200% improvement over standard RLVR, depending on the evaluation criterion. On LCB, CWM 32B wins up to 83% of median-sample speed comparisons against standard RLVR. Relative to the fastest correct human submissions per problem, it reaches about half the human rate of complexity-class improvements (14% vs. 28%).

Large language models (LLMs), such as OpenAI's Codex, have demonstrated their potential to generate code from natural language descriptions across a wide range of programming tasks. Several benchmarks have recently emerged to evaluate the ability of LLMs to generate functionally correct code from natural language intent with respect to a set of hidden test cases. This has enabled the research community to identify significant and reproducible advancements in LLM capabilities. However, there is currently a lack of benchmark datasets for assessing the ability of LLMs to generate functionally correct code edits based on natural language descriptions of intended changes. This paper aims to address this gap by motivating the problem NL2Fix of translating natural language descriptions of code changes (namely bug fixes described in Issue reports in repositories) into correct code fixes. To this end, we introduce Defects4J-NL2Fix, a dataset of 283 Java programs from the popular Defects4J dataset augmented with high-level descriptions of bug fixes, and empirically evaluate the performance of several state-of-the-art LLMs for the this task. Results show that these LLMS together are capable of generating plausible fixes for 64.6% of the bugs, and the best LLM-based technique can achieve up to 21.20% top-1 and 35.68% top-5 accuracy on this benchmark.

This paper addresses the synthetic-to-real domain gap in object detection, focusing on training a YOLOv11 model to detect a specific object (a soup can) using only synthetic data and domain randomization strategies. The methodology involves extensive experimentation with data augmentation, dataset composition, and model scaling. While synthetic validation metrics were consistently high, they proved to be poor predictors of real-world performance. Consequently, models were also evaluated qualitatively, through visual inspection of predictions, and quantitatively, on a manually labeled real-world test set, to guide development. Final mAP@50 scores were provided by the official Kaggle competition. Key findings indicate that increasing synthetic dataset diversity, specifically by including varied perspectives and complex backgrounds, combined with carefully tuned data augmentation, were crucial in bridging the domain gap. The best performing configuration, a YOLOv11l model trained on an expanded and diverse dataset, achieved a final mAP@50 of 0.910 on the competition's hidden test set. This result demonstrates the potential of a synthetic-only training approach while also highlighting the remaining challenges in fully capturing real-world variability.

In this paper, we present SegDINO3D, a novel Transformer encoder-decoder framework for 3D instance segmentation. As 3D training data is generally not as sufficient as 2D training images, SegDINO3D is designed to fully leverage 2D representation from a pre-trained 2D detection model, including both image-level and object-level features, for improving 3D representation. SegDINO3D takes both a point cloud and its associated 2D images as input. In the encoder stage, it first enriches each 3D point by retrieving 2D image features from its corresponding image views and then leverages a 3D encoder for 3D context fusion. In the decoder stage, it formulates 3D object queries as 3D anchor boxes and performs cross-attention from 3D queries to 2D object queries obtained from 2D images using the 2D detection model. These 2D object queries serve as a compact object-level representation of 2D images, effectively avoiding the challenge of keeping thousands of image feature maps in the memory while faithfully preserving the knowledge of the pre-trained 2D model. The introducing of 3D box queries also enables the model to modulate cross-attention using the predicted boxes for more precise querying. SegDINO3D achieves the state-of-the-art performance on the ScanNetV2 and ScanNet200 3D instance segmentation benchmarks. Notably, on the challenging ScanNet200 dataset, SegDINO3D significantly outperforms prior methods by +8.6 and +6.8 mAP on the validation and hidden test sets, respectively, demonstrating its superiority.

Chagas disease screening via ECGs is limited by scarce and noisy labels in existing datasets. We propose a biomarker-based pretraining approach, where an ECG feature extractor is first trained to predict percentile-binned blood biomarkers from the MIMIC-IV-ECG dataset. The pretrained model is then fine-tuned on Brazilian datasets for Chagas detection. Our 5-model ensemble, developed by the Ahus AIM team, achieved a challenge score of 0.269 on the hidden test set, ranking 5th in Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025. Source code and the model are shared on GitHub: github.com/Ahus-AIM/physionet-challenge-2025

We present ENSAM (Equivariant, Normalized, Segment Anything Model), a lightweight and promptable model for universal 3D medical image segmentation. ENSAM combines a SegResNet-based encoder with a prompt encoder and mask decoder in a U-Net-style architecture, using latent cross-attention, relative positional encoding, normalized attention, and the Muon optimizer for training. ENSAM is designed to achieve good performance under limited data and computational budgets, and is trained from scratch on under 5,000 volumes from multiple modalities (CT, MRI, PET, ultrasound, microscopy) on a single 32 GB GPU in 6 hours. As part of the CVPR 2025 Foundation Models for Interactive 3D Biomedical Image Segmentation Challenge, ENSAM was evaluated on hidden test set with multimodal 3D medical images, obtaining a DSC AUC of 2.404, NSD AUC of 2.266, final DSC of 0.627, and final NSD of 0.597, outperforming two previously published baseline models (VISTA3D, SAM-Med3D) and matching the third (SegVol), surpassing its performance in final DSC but trailing behind in the other three metrics. In the coreset track of the challenge, ENSAM ranks 5th of 10 overall and best among the approaches not utilizing pretrained weights. Ablation studies confirm that our use of relative positional encodings and the Muon optimizer each substantially speed up convergence and improve segmentation quality.

Accurate tumor size measurement is a cornerstone of evaluating cancer treatment response. The most widely adopted standard for this purpose is the Response Evaluation Criteria in Solid Tumors (RECIST) v1.1, which relies on measuring the longest tumor diameter in a single plane. However, volumetric measurements have been shown to provide a more reliable assessment of treatment effect. Their clinical adoption has been limited, though, due to the labor-intensive nature of manual volumetric annotation. In this paper, we present Lite ENSAM, a lightweight adaptation of the ENSAM architecture designed for efficient volumetric tumor segmentation from CT scans annotated with RECIST annotations. Lite ENSAM was submitted to the MICCAI FLARE 2025 Task 1: Pan-cancer Segmentation in CT Scans, Subtask 2, where it achieved a Dice Similarity Coefficient (DSC) of 60.7% and a Normalized Surface Dice (NSD) of 63.6% on the hidden test set, and an average total RAM time of 50.6 GBs and an average inference time of 14.4 s on CPU on the public validation dataset.

Many inference-time language-model pipelines combine a cheap reward signal with an expensive verifier, such as exact answer checking in mathematical reasoning or hidden-test execution in code generation. We formalize this setting using a learning-theoretic lens as generative active search: a cost-sensitive first-positive search problem in which a policy adaptively samples candidates from an unknown distribution, observes cheap scores, and pays for verifier labels until it finds a positive example. For a fixed prompt, the generator and reward model induce two unknown objects: a distribution over reward scores and a score-conditioned success function. When these quantities are known, we characterize the distribution-aware optimal policy using a dynamic programming approach. In the realistic and practical setting where both the score distribution and success function are unknown, we propose ADAP, a shellwise adaptive generate-rank-verify algorithm that progressively increases the number of sampled responses and top-ranked verifications. Under the monotonicity assumption that higher reward scores are no less likely to pass verification, we show that ADAP achieves expected cost within a constant factor of the distribution-aware optimum. We complement this result with learning-theoretic lower bounds, based on a centered star number, showing that structural assumptions on the score--label relationship are necessary. Experiments on mathematical reasoning and competitive programming validate the predicted advantage over both fixed non-adaptive policies and difficulty-adaptive baselines.

Robust detection of AI-generated text under deployment conditions is challenging: distribution shifts across domains and generators, adversarial perturbations of the input surface, and the absence of target-domain labels for threshold calibration all degrade detectors that perform well in-domain. We present DeBERTa-ConPara, a deployment-oriented detector combining attack-aware Unicode preprocessing with a contextual transformer encoder trained over HC3 Plus, M4, MAGE and RAID. Our central finding is that preprocessing acts in opposite directions depending on where it is applied: normalising the training corpus deduplicates it, collapsing 35.4% of RAID rows into copies of their clean siblings and deleting the adversarial supervision, whereas normalising at inference is an effective defence. A factorial varying the two placements independently identifies raw training with normalised inference as the best configuration, reaching 99.61% AUROC, 99.01% TPR@5% FPR and 96.57% TPR@1% FPR on the official RAID hidden test, alongside 93.14% average balanced accuracy across HC3 Plus and MAGE under a fixed threshold. The gain is confined to two of twelve attack classes: homoglyph and zero-width-space insertion rise from 11.05% and 1.12% to 96.98%. The same signature reproduces in a zero-shot detector of different architecture, showing the effect belongs to the attacks rather than to our model. We additionally report two negative results: semantic-invariance augmentation through paraphrasing and supervised contrastive learning (ConPara) does not improve the best configuration, and the handcrafted feature-fusion branch is inert in distribution and harmful outside it.

Agents need reliable ways to validate their outputs, both when completing tasks and when filtering synthetic training data. We study verifier generation: from the instruction and environment available for solving a task, an agent writes an executable verifier that must accept every valid solution and reject invalid ones. We introduce VerifierBench, a benchmark of 222 tasks from six agent benchmarks spanning tool-call, state-based, and code verification, with 4,993 hidden test cases that include alternative valid solutions and attacks; Verifier4Verifier (V4V), a six-check audit, screens its source verifiers and test suites. Across 19 models in a common agent harness, normalized balanced accuracy ranges from 10.6 to 62.1, with generated verifiers consistently rejecting invalid solutions more reliably than they accept valid ones. This apparent strength can mask overly restrictive verification: most false rejections in the analyzed records come from verifiers that accept no tested solution to a task. Case studies reveal unsupported assumptions about solution formats, tool use, and code changes that exclude valid solutions. These errors survive self-testing because agents test solutions that share the verifier's mistaken assumptions, or submit verifiers without confirming that any correct solution passes. These findings suggest testing both whether a verifier accepts a correct solution and whether it accepts other solutions permitted by the task. The project page and dataset are publicly available.

We present ReMasker, a new method of imputing missing values in tabular data by extending the masked autoencoding framework. Compared with prior work, ReMasker is both simple -- besides the missing values (i.e., naturally masked), we randomly ``re-mask'' another set of values, optimize the autoencoder by reconstructing this re-masked set, and apply the trained model to predict the missing values; and effective -- with extensive evaluation on benchmark datasets, we show that ReMasker performs on par with or outperforms state-of-the-art methods in terms of both imputation fidelity and utility under various missingness settings, while its performance advantage often increases with the ratio of missing data. We further explore theoretical justification for its effectiveness, showing that ReMasker tends to learn missingness-invariant representations of tabular data. Our findings indicate that masked modeling represents a promising direction for further research on tabular data imputation. The code is publicly available.

How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average, and to solve this very problem rather than generalize to other problems. Therefore, our learning objective and search subroutine are designed to prioritize the most promising solutions. We call this method Test-Time Training to Discover (TTT-Discover). Following prior work, we focus on problems with continuous rewards. We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) Erdős' minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to 2times faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers. All our results are achieved with an open model, OpenAI gpt-oss-120b, and can be reproduced with our publicly available code, in contrast to previous best results that required closed frontier models. Our test-time training runs are performed using Tinker, an API by Thinking Machines, with a cost of only a few hundred dollars per problem.

We present JAttack, a framework that enables template-based testing for compilers. Using JAttack, a developer writes a template program that describes a set of programs to be generated and given as test inputs to a compiler. Such a framework enables developers to incorporate their domain knowledge on testing compilers, giving a basic program structure that allows for exploring complex programs that can trigger sophisticated compiler optimizations. A developer writes a template program in the host language (Java) that contains holes to be filled by JAttack. Each hole, written using a domain-specific language, constructs a node within an extended abstract syntax tree (eAST). An eAST node defines the search space for the hole, i.e., a set of expressions and values. JAttack generates programs by executing templates and filling each hole by randomly choosing expressions and values (available within the search space defined by the hole). Additionally, we introduce several optimizations to reduce JAttack's generation cost. While JAttack could be used to test various compiler features, we demonstrate its capabilities in helping test just-in-time (JIT) Java compilers, whose optimizations occur at runtime after a sufficient number of executions. Using JAttack, we have found six critical bugs that were confirmed by Oracle developers. Four of them were previously unknown, including two unknown CVEs (Common Vulnerabilities and Exposures). JAttack shows the power of combining developers' domain knowledge (via templates) with random testing to detect bugs in JIT compilers.

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

Hidden chain-of-thought (CoT) traces, especially those from frontier proprietary large reasoning models (LRMs), are valuable model assets. Yet whether these hidden CoTs can be directly extracted from black-box models remains largely unexplored. In this work, we systematically study whether hidden CoTs can be extracted near-verbatim from black-box LRMs through API interactions. We identify a previously overlooked reasoning replay surface between tool calls and develop EchoCoT, a multi-step attack that iteratively extracts hidden CoTs using API-returned fidelity signals. We further develop an LLM-based optimization framework that automatically searches for an effective universal injection trajectory across various datasets. We evaluate EchoCoT on three open-source and five frontier proprietary LRMs. On open-source LRMs, EchoCoT achieves up to 66.4\% near-verbatim extraction success, with the extracted trace length within 10\% of the target and at least 90\% of tokens exactly matching the target CoT. The same injection trajectory also generalizes to unseen datasets, achieving up to 80\% extraction success under the same criterion. For tested frontier proprietary LRMs, a substantial fraction of extracted CoTs closely align with provider-reported reasoning lengths and available CoT summaries. EchoCoT can also extract very long CoTs: on Gemini-2.5, it extracts 33,463 tokens from a 32,948-token target. These results establish hidden-CoT extraction as a practical security risk and highlight the need to better protect hidden CoT assets.

Test-time prompt tuning enhances zero-shot generalization of vision-language models but tends to ignore the relatedness among test samples during inference. Online test-time prompt tuning provides a simple way to leverage the information in previous test samples, albeit with the risk of prompt collapse due to error accumulation. To enhance test-time prompt tuning, we propose DynaPrompt, short for dynamic test-time prompt tuning, exploiting relevant data distribution information while reducing error accumulation. Built on an online prompt buffer, DynaPrompt adaptively selects and optimizes the relevant prompts for each test sample during tuning. Specifically, we introduce a dynamic prompt selection strategy based on two metrics: prediction entropy and probability difference. For unseen test data information, we develop dynamic prompt appending, which allows the buffer to append new prompts and delete the inactive ones. By doing so, the prompts are optimized to exploit beneficial information on specific test data, while alleviating error accumulation. Experiments on fourteen datasets demonstrate the effectiveness of dynamic test-time prompt tuning.

As multi-agent systems enter high-stakes domains, the possibility that agents may circumvent safety boundaries is a growing concern. Prior work has examined this risk primarily in adversarial settings, where agents are instructed or rewarded to communicate covertly and evade oversight. We show that benign agents can cross the same boundaries without adversarial incentives. We emulate a software-engineering workflow in which a planner represents a company hiring an external developer. The planner writes requirements and holds a company credential it is instructed not to disclose to the developer; a monitor screens their exchanges. Seven of nine tested frontier models disguise the credential in their requirements to help the developer recover it while evading the monitor, even after completing their assigned objective. For example, across 6,000 episodes with DeepSeek-V4-Pro, the planner attempts concealment in 16.9%; in 0.9%, the credential evades the monitor and is recovered and used by the developer. Despite the low per-episode rate, the risk compounds across repeated exchanges: at 0.9%, 105 independent episodes carry a 61.3% chance of at least one breach. Our analysis suggests that the planner misinterprets disguised disclosure as helping the developer and reads the nondisclosure rule as banning plaintext, not character codes or riddles. We call this behavior covert assistance. Monitoring favors concealment, while concealment gives way to direct disclosure when the recipient is presented as a human rather than an AI agent, so leakage persists. Giving the monitor the credential improves detection but requires trusting it with the secret. These risks, in models already used for software engineering, challenge oversight to distinguish authorized cooperation from task-advancing assistance that crosses safety boundaries.

The integration of AI-powered coding agents into Continuous Integration/Continuous Delivery (CI/CD) pipelines has fundamentally altered how software verification is conducted. While these agents successfully automate the test generation, current evaluation benchmarks (e.g., SWE-bench) largely focus on pass-rates rather than the intrinsic quality of the generated tests. This raises the possibility of "stealth technical debt", in which test suites pass execution but do not offer comprehensive coverage or semantic value. We address this methodological gap through a large-scale, empirical comparison of 204,673 test artifacts which comprises of 24,941 human-authored files and 179,732 agent-generated files; sourced from the AIDev dataset. Using the Abstract Syntax Tree (AST) parsing with Python's naive ast module, we implemented a "white-box" static analysis framework to evaluate three quality dimensions: Assertion Strength (RQ1), Edge-Case Coverage (RQ2), and Flakiness Potential (RQ3). Our results present a nuanced inversion of traditional assumptions. AI agents performed better than humans in Edge-Case Coverage, with almost twice the variety of boundary checks (Variety Score: 0.62 vs 0.32) and a higher frequency of null-safety testing (13.40% vs. 8.3%), even though human developers had a slight advantage in Assertion Strength (88.1% strong assertions vs. 85.37% for agents). But this thoroughness comes at a price: due mostly to their reliance on file I/O and non-deterministic logic, agent-generated tests exhibited a higher risk of flakiness (Candidate Rate: 0.41 vs. 0.30). These findings suggest that while AI agents excel at rigorous boundary testing, they lack the "environmental awareness" needed to write stable, hermetic tests.

A retrieval-augmented QA system can return different answers after an index expansion even when its requested model identifier, prompt, retrieval policy, evidence depth, rendering, and exposed generation controls are held fixed. Aggregate accuracy may hide these changes when gains and losses cancel, while ordinary generation variability makes one-shot comparisons overstate update effects. We call the hidden phenomenon accuracy-blind answer churn and introduce the Snapshot Compatibility Audit, which estimates excess answer churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement. We instantiate it by expanding one frozen FineWeb prefix from one to seven shards. In a preregistered 400-question Natural Questions study, normalized-exact and blinded-semantic excess churn are 6.44 and 10.25 percentage points while exact-match accuracy changes by only -1.50 points. A post-hoc analysis finds repeat-stable semantic flips on 40/400 questions. A separately preregistered 200-question TriviaQA study yields smaller, directionally consistent excess churn while exact-match accuracy moves in the opposite direction. An outcome-blind post-hoc 100-question subset replication with a second DeepSeek generator and serving configuration finds 8.75 pp of semantic excess churn even as exact match rises by 3.00 percentage points. Answer-level compatibility can therefore fail without a conspicuous or consistently directed utility shift. Retrieval-augmented releases should audit compatibility alongside utility.

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.

Large language models (LLMs) have been widely applied to assist test generation with the source code under test provided as the context. This paper aims to answer the question: If the source code under test is incorrect, will LLMs be misguided when generating tests? The effectiveness of test cases is measured by their accuracy, coverage, and bug detection effectiveness. Our evaluation results with five open- and six closed-source LLMs on four datasets demonstrate that incorrect code can significantly mislead LLMs in generating correct, high-coverage, and bug-revealing tests. For instance, in the HumanEval dataset, LLMs achieve 80.45% test accuracy when provided with task descriptions and correct code, but only 57.12% when given task descriptions and incorrect code. For the APPS dataset, prompts with correct code yield tests that detect 39.85% of the bugs, while prompts with incorrect code detect only 19.61%. These findings have important implications for the deployment of LLM-based testing: using it on mature code may help protect against future regression, but on early-stage immature code, it may simply bake in errors. Our findings also underscore the need for further research to improve LLMs resilience against incorrect code in generating reliable and bug-revealing tests.

With the help of simple fine-tuning, one can artificially embed hidden text into large language models (LLMs). This text is revealed only when triggered by a specific query to the LLM. Two primary applications are LLM fingerprinting and steganography. In the context of LLM fingerprinting, a unique text identifier (fingerprint) is embedded within the model to verify licensing compliance. In the context of steganography, the LLM serves as a carrier for hidden messages that can be disclosed through a designated trigger. Our work demonstrates that embedding hidden text in the LLM via fine-tuning, though seemingly secure due to the vast number of potential triggers (any sequence of characters or tokens could serve as a trigger), is susceptible to extraction through analysis of the LLM's output decoding process. We propose a novel approach to extraction called Unconditional Token Forcing. It is premised on the hypothesis that iteratively feeding each token from the LLM's vocabulary into the model should reveal sequences with abnormally high token probabilities, indicating potential embedded text candidates. Additionally, our experiments show that when the first token of a hidden fingerprint is used as an input, the LLM not only produces an output sequence with high token probabilities, but also repetitively generates the fingerprint itself. We also present a method to hide text in such a way that it is resistant to Unconditional Token Forcing, which we named Unconditional Token Forcing Confusion.

There is an increasing amount of research and commercial tools for automated test case generation using Large Language Models (LLMs). This paper critically examines whether recent LLM-based test generation tools, such as Codium CoverAgent and CoverUp, can effectively find bugs or unintentionally validate faulty code. Considering bugs are only exposed by failing test cases, we explore the question: can these tools truly achieve the intended objectives of software testing when their test oracles are designed to pass? Using real human-written buggy code as input, we evaluate these tools, showing how LLM-generated tests can fail to detect bugs and, more alarmingly, how their design can worsen the situation by validating bugs in the generated test suite and rejecting bug-revealing tests. These findings raise important questions about the validity of the design behind LLM-based test generation tools and their impact on software quality and test suite reliability.

Language models are increasingly considered for clinical decision support, where their answers should be grounded in clinical evidence and remain robust to irrelevant information. Prior work shows that a single irrelevant sentence can lower the accuracy of medical models, but not where inside the model its effect can be removed, or how to do so without knowing which sentence it is. We find that blocking attention to a known distractor restores the answer, whereas head masking can suppress correct answers. We also show that the model's hidden states mark the distractor as the sentence that least fits the case even when the model follows it. To remove a distractor's effect without knowing which sentence it is, we propose EXCISE, an inference-time method that blocks the sentence most influencing the answer, but only if the model's hidden states mark it as inconsistent with the case. On MedQA-TD (MedQA with Targeted Distractors), a benchmark we construct in which each distractor targets a specific wrong option, EXCISE gains 16.7-22.8 accuracy points at a 1.5-2.1-point cost on the same questions without distractors, outperforming prompting, rewriting, and steering baselines for all three models. Its gains extend to MedDistractQA and BiasMedQA, as well as chain-of-thought generation.

AI systems can fail silently. The failure propagates through training loops, evaluation pipelines, and production monitoring stacks until downstream harm makes it visible. This paper introduces evaluation blindness: a measurement function M exhibits evaluation blindness with respect to failure class F when it produces readings indistinguishable from a healthy state while the system is actually failing, with no auxiliary signal flagging the gap. The problem surfaces at two lifecycle stages the literature has treated separately. At training time, reward models are gamed, importance-sampling corrections are silently miscalculated, and benchmark contamination inflates fine-tuning evaluations, all while loss curves look healthy and gradient updates proceed normally. At deployment time, monitoring fails to catch six classes of production failure, including an Operational category that is 100% silent by structural definition. We provide a formal detectability predicate unifying both stages. Four training-time case studies trace concrete breakdowns, including a real implementation bug in TRL PR #6594 where gradients are corrupted as loss decreases normally. A six-class taxonomy validated against 50 real-world incidents from court documents and regulatory filings finds that 53% of verifiable public failures were silent. A failure budget framework ties acceptable failure rates to use-case risk class. The implication is direct: measurement infrastructure is a correctness concern across the full AI lifecycle, not just at evaluation time. Data, code, and taxonomy schema are at https://github.com/priyanka25aug/llm-failure-taxonomy.

Generative models based on variational autoencoders are a popular technique for detecting anomalies in images in a semi-supervised context. A common approach employs the anomaly score to detect the presence of anomalies, and it is known to reach high level of accuracy on benchmark datasets. However, since anomaly scores are computed from reconstruction disparities, they often obscure the detection of various spurious features, raising concerns regarding their actual efficacy. This case study explores the robustness of an anomaly detection system based on variational autoencoder generative models through the use of eXplainable AI methods. The goal is to get a different perspective on the real performances of anomaly detectors that use reconstruction differences. In our case study we discovered that, in many cases, samples are detected as anomalous for the wrong or misleading factors.

A primary concern regarding training large language models (LLMs) is whether they abuse copyrighted online text. With the increasing training data scale and the prevalence of LLMs in daily lives, two problems arise: 1) false positive membership inference results misled by similar examples; 2) membership inference methods are usually too complex for end users to understand and use. To address these issues, we propose an alternative insert-and-detect methodology, advocating that web users and content platforms employ \textit{unique identifiers} for reliable and independent membership inference. Users and platforms can create their identifiers, embed them in copyrighted text, and independently detect them in future LLMs. As an initial demonstration, we introduce \textbf{ghost sentences} and a user-friendly last-k words test, allowing end users to chat with LLMs for membership inference. Ghost sentences consist primarily of unique passphrases of random natural words, which can come with customized elements to bypass possible filter rules. The last-k words test requires a significant repetition time of ghost sentences~(ge10). For cases with fewer repetitions, we designed an extra perplexity test, as LLMs exhibit high perplexity when encountering unnatural passphrases. We also conduct a comprehensive study on the memorization and membership inference of ghost sentences, examining factors such as training data scales, model sizes, repetition times, insertion positions, wordlist of passphrases, alignment, etc. Our study shows the possibility of applying ghost sentences in real scenarios and provides instructions for the potential application.

Implementing automated unit tests is an important but time-consuming activity in software development. To assist developers in this task, many techniques for automating unit test generation have been developed. However, despite this effort, usable tools exist for very few programming languages. Moreover, studies have found that automatically generated tests suffer poor readability and do not resemble developer-written tests. In this work, we present a rigorous investigation of how large language models (LLMs) can help bridge the gap. We describe a generic pipeline that incorporates static analysis to guide LLMs in generating compilable and high-coverage test cases. We illustrate how the pipeline can be applied to different programming languages, specifically Java and Python, and to complex software requiring environment mocking. We conducted an empirical study to assess the quality of the generated tests in terms of code coverage and test naturalness -- evaluating them on standard as well as enterprise Java applications and a large Python benchmark. Our results demonstrate that LLM-based test generation, when guided by static analysis, can be competitive with, and even outperform, state-of-the-art test-generation techniques in coverage achieved while also producing considerably more natural test cases that developers find easy to understand. We also present the results of a user study, conducted with 161 professional developers, that highlights the naturalness characteristics of the tests generated by our approach.

LLM agents often rely on Skills to describe available tools and recommended procedures. We study a hidden-comment prompt injection risk in this documentation layer: when a Markdown Skill is rendered to HTML, HTML comment blocks can become invisible to human reviewers, yet the raw text may still be supplied verbatim to the model. In experiments, we find that DeepSeek-V3.2 and GLM-4.5-Air can be influenced by malicious instructions embedded in a hidden comment appended to an otherwise legitimate Skill, yielding outputs that contain sensitive tool intentions. A short defensive system prompt that treats Skills as untrusted and forbids sensitive actions prevents these malicious tool calls and instead surfaces the suspicious hidden instructions.

Most benchmarks hand the model a question. EnigmaForge hands it a stack of old documents and no question at all. Buried in the letters, receipts, and logbook margins is a small logic puzzle whose solution is unique - proved by a SAT solver at generation time, with an ablation certificate showing every clue is load-bearing. Because instances are generated rather than collected, the corpus renews forever. The headline measure is intuition: task success when handed only the story, with world reconstruction as the secondary axis. Twenty-five frontier models ran over 600 instances (17,400 scored records) under three matched conditions. Intuition reshuffles the leaderboard: a 22x spread where fact recovery spans 1.6x, the second-best fact-recoverer ranks fourteenth, one model is indifferent to being told the question, and another is significantly better without it. Several models were blocked by their own content filters before reaching the puzzle - any benchmark scoring refusals as failure is quietly measuring filter behavior.