Daily Papers

Medical vision-language model (VLM) evaluation is sensitive to workflow design, prompting strategy, and benchmark construction, yet most studies treat these factors in isolation. We introduce a reliability stress test for chest X-ray interpretation built on two balanced datasets (a private report-backed set and a curated MIMIC subset). Three medical VLMs (CheXagent, MedGemma-4B, and MedGemma-27B) are evaluated across three prompt styles and two workflows (single-VLM and multi-agent), producing 36 configurations. We show that exact-match accuracy alone can overstate the effectiveness of conservative models that default to "Normal" predictions. Diagnostic reliability also depends heavily on model family and scale: multi-agent reasoning helps some configurations but hurts others. Building on these observations, we propose a decision-time routing framework that selectively escalates to multi-agent inference only when beneficial, improving the cost-quality trade-off over fixed workflows. Our results highlight the need for evaluation protocols that jointly consider prompt sensitivity, failure-mode diversity, and workflow choice before clinical deployment.

Population-level behavior in large-language-model (LLM) agents cannot be characterized by single-agent benchmarks. We introduce PV-SST, a peer-voted social-platform testbed, and report a separately frozen, preregistered matched-exposure experiment spanning four topics, four unused seeds, four open-weight model families, and three prespecified larger variants. The experiment comprises 448 trials and 112 complete model-by-topic-by-seed blocks. Relative to a topic-only control, a feed of previous-round peer posts ranked by peer-generated likes increases final-round lexical similarity in both the four-family core panel (paired mean difference +0.0082 TF-IDF cosine units, 95% block-bootstrap CI [0.0043, 0.0121], randomization p=0.000105, n=64 blocks) and the three-variant size extension (+0.0109 [0.0069, 0.0151], p=0.000001, n=48). This contrast bundles peer-post exposure with ranking and therefore does not identify a ranking-only effect. Opposite-side survival falls in the core panel (-3.9 percentage points [-6.8, -1.6], p=0.0068) but not conclusively in the larger variants (-1.0 pp [-3.1, 0.4], p=0.50). Holding adversarial impressions fixed, four distributed sources do not reliably move honest-agent stance more than one source. The preregistered distributed-minus-single contrast is positive but inconclusive in the core panel (+0.057 [-0.009, 0.125], p=0.112) and negative in the larger variants (-0.040 [-0.113, 0.035], p=0.332), failing the prespecified cross-model and cross-topic consistency criterion. Thus the robust result is lexical convergence under the tested peer-ranked feed, not general opinion capture or a general coordination advantage. The study evaluates synthetic LLM-agent populations; it does not estimate effects on people or production platforms.

Multi-agent LLM systems that generate structured workflows from natural-language requests are now deployed in production across cloud automation, DevOps, and enterprise process orchestration. Operating such systems exposes a recurring change-management problem. Routine updates, such as re-running the same input, swapping the underlying LLM, or refactoring an agent's prompt or orchestration code, frequently produce workflows that differ substantially from previously validated references. Engineers are then left without a principled way to decide whether a change is safe to ship. Automatic workflow evaluation is the natural tool for answering this question. In practice, however, metric scores are poorly calibrated, and a numeric change rarely communicates the severity of the underlying degradation. We introduce WorkflowPerturb, a controlled benchmark for studying workflow evaluation metrics by applying realistic, graded perturbations to golden workflows. WorkflowPerturb contains 4,973 golden workflows and 44,757 perturbed variants across three perturbation types (Missing Steps, Compressed Steps, and Description Changes), each applied at severity levels of 10%, 30%, and 50%. We benchmark multiple metric families and analyze their sensitivity and calibration using expected score trajectories and residuals. Our results characterize systematic differences across metric families and support severity-aware interpretation of workflow evaluation scores in change-management settings. Our dataset will be released upon acceptance.

Natural language inference (NLI) is the task of determining if a natural language hypothesis can be inferred from a given premise in a justifiable manner. NLI was proposed as a benchmark task for natural language understanding. Existing models perform well at standard datasets for NLI, achieving impressive results across different genres of text. However, the extent to which these models understand the semantic content of sentences is unclear. In this work, we propose an evaluation methodology consisting of automatically constructed "stress tests" that allow us to examine whether systems have the ability to make real inferential decisions. Our evaluation of six sentence-encoder models on these stress tests reveals strengths and weaknesses of these models with respect to challenging linguistic phenomena, and suggests important directions for future work in this area.

We study scheming in LLM agents, in which agents covertly pursue misaligned goals. Our focus is to understand how scheming arises from the interaction of key factors, such as instrumental goals, environmental affordances, oversight conditions, and perceived consequences. Prior work examines only a small number of scenarios, limiting the ability to isolate how these conditions shape an agent's propensity or capability to scheme. This limited scale and task diversity also restrict coverage of realistic deployment settings and the range of scheming strategies that can be observed. To this end, we introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms. To enable scalable and reliable monitoring, we further propose SCOUT, a scheming monitor that grounds multi-criteria judgments in evidence drawn from agents' reasoning and actions. Across controlled stress tests on five LLM agents, we find that explicit instrumental goals are the strongest driver of scheming propensity. Strategic hints play a distinct role by helping agents translate scheming reasoning into concrete covert behavior. Oversight has mixed effects: in several closed models, action-only monitoring increases scheming, suggesting that partial oversight can act as an optimization constraint rather than a deterrent. CoT is a useful but incomplete monitoring signal: it can reveal latent scheming before execution, yet action-only scheming shows that covert behavior may occur without explicit reasoning evidence. We release the benchmark, code, and monitor at: https://github.com/launchnlp/SchemeArena.

Occupancy world models use historical occupancy states to forecast future 3D scenes, but their robustness under corrupted temporal inputs remains poorly understood. Existing evaluations primarily emphasize clean forecasting accuracy and provide limited evidence about how errors enter, persist, and propagate through the occupancy perception-forecasting chain. This paper introduces OccStress, a robustness stress-testing benchmark for the occupancy forecasting chain. OccStress contains 21 corruption families with 61 severity configurations and 10,827 strict temporal anchors across 3 datasets. OccStress covers both 3D occupancy perception and 4D occupancy forecasting through two complementary tracks. This design separates model-mediated pipeline errors under standardized sensor stressors from the intrinsic sensitivity of 4D forecasting models to corrupted occupancy states. OccStress further defines temporal injection protocols to test whether errors in the current state, recent history, or earlier history affect future forecasts differently. OccStress provides aggregate metrics for evaluating robustness along the occupancy forecasting chain. Experiments with five 4D forecasters reveal that current occupancy models are substantially affected by both upstream prediction errors and direct state corruptions, and that clean performance alone is an insufficient description of source- and position-specific robustness. Code, data, and evaluation tools are available at https://insailab.org/OccStress.

Modern language-model training is increasingly exposed to instability, degraded runs, and wasted compute, especially under aggressive learning-rate, scale, and runtime-stress conditions. This paper introduces Learn-by-Wire Guard (LBW-Guard), a bounded autonomous training-control governance layer that operates above AdamW. Rather than replacing the optimizer update rule, LBW-Guard observes training telemetry, interprets instability-sensitive regimes, and applies bounded control to optimizer execution while preserving fixed training objectives. We evaluate LBW-Guard in a Qwen2.5-centered stress-and-robustness suite using WikiText-103, with Qwen2.5-7B as the empirical anchor, model-size comparisons against Qwen2.5-3B and Qwen2.5-14B, learning-rate stress tests, gradient-clipping baselines, and a no-LoRA TinyLlama-1B full-parameter sanity check. In the 7B reference setting, LBW-Guard reduces final perplexity from 13.21 to 10.74, an 18.7% improvement, while reducing end-to-end time from 392.54s to 357.02s, a 1.10x speedup. Under stronger learning-rate stress, AdamW degrades to 1885.24 final perplexity at LR=3e-3 and 659.76 at LR=1e-3, whereas LBW-Guard remains trainable at 11.57 and 10.33, respectively. Gradient-clipping baselines do not reproduce this effect. These results support a scoped systems conclusion that stability-sensitive LLM training can benefit from a governance plane above the optimizer. LBW-Guard provides evidence that bounded runtime control can preserve productive compute under stress while remaining distinct from optimizer replacement and local gradient suppression.

Dialect differences caused by regional, social, and economic factors cause performance discrepancies for many groups of language technology users. Inclusive and equitable language technology must critically be dialect invariant, meaning that performance remains constant over dialectal shifts. Current systems often fall short of this ideal since they are designed and tested on a single dialect: Standard American English (SAE). We introduce a suite of resources for evaluating and achieving English dialect invariance. The resource is called Multi-VALUE, a controllable rule-based translation system spanning 50 English dialects and 189 unique linguistic features. Multi-VALUE maps SAE to synthetic forms of each dialect. First, we use this system to stress tests question answering, machine translation, and semantic parsing. Stress tests reveal significant performance disparities for leading models on non-standard dialects. Second, we use this system as a data augmentation technique to improve the dialect robustness of existing systems. Finally, we partner with native speakers of Chicano and Indian English to release new gold-standard variants of the popular CoQA task. To execute the transformation code, run model checkpoints, and download both synthetic and gold-standard dialectal benchmark datasets, see http://value-nlp.org.

Community GPU platforms are emerging as a cost-effective and democratized alternative to centralized GPU clusters for AI workloads, aggregating idle consumer GPUs from globally distributed and heterogeneous environments. However, their extreme hardware/software diversity, volatile availability, and variable network conditions render traditional schedulers ineffective, leading to suboptimal task completion. In this work, we present REACH (Reinforcement Learning for Efficient Allocation in Community and Heterogeneous Networks), a Transformer-based reinforcement learning framework that redefines task scheduling as a sequence scoring problem to balance performance, reliability, cost, and network efficiency. By modeling both global GPU states and task requirements, REACH learns to adaptively co-locate computation with data, prioritize critical jobs, and mitigate the impact of unreliable resources. Extensive simulation results show that REACH improves task completion rates by up to 17%, more than doubles the success rate for high-priority tasks, and reduces bandwidth penalties by over 80% compared to state-of-the-art baselines. Stress tests further demonstrate its robustness to GPU churn and network congestion, while scalability experiments confirm its effectiveness in large-scale, high-contention scenarios.

Emergent misalignment (EM) occurs when narrow finetuning causes a model to behave dangerously outside the finetuning task. Standard training signals can miss this shift, making reliable detection costly if it depends on repeated behavioral evaluation. We ask whether emergent misalignment can instead be detected from internal representations during finetuning. Using seven alignment-relevant traits encoded as linear directions in activation space, we track representational drift across training checkpoints in four open-source 7-9B LLMs. EM-relevant drift concentrates on a low-dimensional axis that explains 65.5% of the variance, revealing a geometric signature in the studied regime. A low-overhead monitor built on this drift profile detects dangerous checkpoints with 2.2% false negative rate, 2.9% false positive rate, and 0.990 AUROC on held-out perturbation types, outperforming unsupervised PCA and SAE baselines. Stress tests on two 14B models, longer finetuning runs, and misaligned starting points identify key deployment boundaries. These results position trait-space monitoring as a practical complement to behavioral evaluation for EM detection during LoRA-based finetuning, while showing that deployment across substantially different regimes may require recalibration.

We study behavioral alignment and representation dynamics of large language model (LLM) agents in financial decision environments. TradeArena, an auditable trading-agent testbed with risk reports, execution simulation, memory, and replayable trajectories, lets us analyze how rationales, positions, and interventions evolve under market stress. Code and data artifacts are available through the https://github.com/weich97/TradeArena.git{TradeArena repository}. We find pre-failure signatures: planning embeddings drift from normal centroids, fused plan-risk representations separate normal from pre-drawdown states, and local manifolds exhibit effective-rank contraction. Across 80 rolling failure anchors and eight LLM trajectories, this pattern persists across hash, LSA, Transformer, and white-box hidden-state probes. Stress tests with CoT-free target weights, lexical controls, OHLCV noise, and false audits show that rationale-level contraction can vanish without rationales, while intent-space and fused signatures remain informative. Structured risk feedback can act as an external alignment signal without fine-tuning, but not as a universal performance enhancer: true audit feedback improves calibration for some models, returns for others, and exposes cases where placebo or hidden feedback has higher short-horizon return but weaker alignment diagnostics. A 51-stock intraday experiment reveals a correlation blind spot: LLM rationales justify exposure to coupled assets that the risk layer clips. Finally, a financial-audit task suite shifts comparison from ``which model trades best'' to whether models can audit trajectories, respect execution boundaries, reproduce artifacts, and avoid claim overreach. These results support a research claim, not a profitability claim: auditable risk feedback and representation trajectories reveal when LLM financial reasoning is aligning, drifting, or failing.

Training stable biological foundation models requires rethinking attention mechanisms: we find that using sigmoid attention as a drop in replacement for softmax attention a) produces better learned representations: on six diverse single-cell datasets, sigmoid achieves 25% higher cell-type separation, better cell-type cohesion metrics, and lower validation loss, b) faster training, models with sigmoid attention train up to 10% faster than their softmax counterparts, and c) more stable training by eliminating inherent sources of instability in softmax attention. We establish that sigmoid attention has globally bounded derivatives (leq 0.25) as opposed to softmax, and a diagonal Jacobian structure in contrast with softmax's dense coupling, which together help alleviate training instabilities. In stress tests on 160M-parameter bidirectional attention models trained without gradient clipping on 8K-token sequences, softmax diverges catastrophically, with gradients exploding by four orders of magnitude, while sigmoid remains stable. Finally, we implement and open-source TritonSigmoid, an efficient GPU kernel that achieves 515 TFLOPS on H100 GPUs, outperforming both FlashAttention-2 and FlashSigmoid, with native padding support, which is essential for biological sequences. Our results establish sigmoid attention as both theoretically grounded and empirically superior for biological foundation models. Code is available at https://github.com/MSDLLCpapers/triton-sigmoid

We present SICK-NL (read: signal), a dataset targeting Natural Language Inference in Dutch. SICK-NL is obtained by translating the SICK dataset of Marelli et al. (2014)from English into Dutch. Having a parallel inference dataset allows us to compare both monolingual and multilingual NLP models for English and Dutch on the two tasks. In the paper, we motivate and detail the translation process, perform a baseline evaluation on both the original SICK dataset and its Dutch incarnation SICK-NL, taking inspiration from Dutch skipgram embeddings and contextualised embedding models. In addition, we encapsulate two phenomena encountered in the translation to formulate stress tests and verify how well the Dutch models capture syntactic restructurings that do not affect semantics. Our main finding is all models perform worse on SICK-NL than on SICK, indicating that the Dutch dataset is more challenging than the English original. Results on the stress tests show that models don't fully capture word order freedom in Dutch, warranting future systematic studies.

This report presents the Prime Collective Communications Library (PCCL), a novel fault-tolerant collective communication library designed for distributed ML workloads over the public internet. PCCL introduces a new programming model that enables dynamic peer joining and failure recovery. The library implements efficient collective operations like all-reduce while providing robust fault tolerance mechanisms that allow the system to continue operating even when peers fail or join during ongoing operations. We demonstrate that PCCL's design enables practical solutions to dynamic membership challenges in workloads with repeated operations and deterministic state advancement. Our implementation passes extensive stress tests across all major operating systems, showing reliable operation even under rapid peer churn and concurrent collective operations. By dispatching to multiple connections, we can efficiently utilize cross-continental long-fat-pipe TCP WAN links, in our experiments achieving up to 45 Gbit/s of bandwidth utilization across Europe and 25 Gbit/s across North America and Europe. PCCL's architecture enables easy implementation of distributed low-communication optimization strategies like DiLoCo, which significantly reduce communication frequency. Combined with quantization, this leads to a significant reduction in the bandwidth required for distributed training workloads. PCCL also allows for concurrent collective operations, which enables optimization strategies like async DiLoCo, which can completely hide communication overhead by implementing one-step delayed parameter updates. PCCL can facilitate exact bit-parity of the shared state across peers in all cases induced by graceful or abrupt peer churn. While PCCL exposes a C99 API, Python bindings are available which are compatible with PyTorch alongside FSDP. PCCL is available under the open source MIT license.

Despite rapid progress in building conversational AI agents, robustness is still largely untested. Small shifts in user behavior, such as being more impatient, incoherent, or skeptical, can cause sharp drops in agent performance, revealing how brittle current AI agents are. Today's benchmarks fail to capture this fragility: agents may perform well under standard evaluations but degrade spectacularly in more realistic and varied settings. We address this robustness testing gap by introducing TraitBasis, a lightweight, model-agnostic method for systematically stress testing AI agents. TraitBasis learns directions in activation space corresponding to steerable user traits (e.g., impatience or incoherence), which can be controlled, scaled, composed, and applied at inference time without any fine-tuning or extra data. Using TraitBasis, we extend tau-Bench to tau-Trait, where user behaviors are altered via controlled trait vectors. We observe on average a 2%-30% performance degradation on tau-Trait across frontier models, highlighting the lack of robustness of current AI agents to variations in user behavior. Together, these results highlight both the critical role of robustness testing and the promise of TraitBasis as a simple, data-efficient, and compositional tool. By powering simulation-driven stress tests and training loops, TraitBasis opens the door to building AI agents that remain reliable in the unpredictable dynamics of real-world human interactions. We have open-sourced tau-Trai across four domains: airline, retail, telecom, and telehealth, so the community can systematically QA their agents under realistic, behaviorally diverse intents and trait scenarios: https://github.com/collinear-ai/tau-trait.

Direct Preference Optimization (DPO) simplifies alignment through pairwise comparisons but assumes all observed preferences are reliable. Real data often violates this assumption, leading to reversed, weak, or ambiguous labels that cause harmful policy updates. To address this, we propose Posterior Label Correction DPO (PLC-DPO) to robustly optimize preferences by routing each pair's training signal as a clean, flip, or tie case. The key idea is to use the calibrated policy-reference margin as online evidence to take appropriate correction actions. This reframes noisy preference learning as actively correcting supervision direction and strength rather than merely filtering suspicious examples. Across 57 dataset-model-benchmark cells, PLC-DPO obtains the best mean win rate against DPO (60.5 vs. 55.5 for the next-best method). Injected-noise and tie stress tests, human disagreement analysis, and self-confirmation diagnostics further show that the routing remains stable and distinguishes flipped from weakly directional pairs.

Current work on robot failure detection and correction typically operates in a post hoc manner, analyzing errors and applying corrections only after failures occur. This work introduces CycleVLA, a system that equips Vision-Language-Action models (VLAs) with proactive self-correction, the capability to anticipate incipient failures and recover before they fully manifest during execution. CycleVLA achieves this by integrating a progress-aware VLA that flags critical subtask transition points where failures most frequently occur, a VLM-based failure predictor and planner that triggers subtask backtracking upon predicted failure, and a test-time scaling strategy based on Minimum Bayes Risk (MBR) decoding to improve retry success after backtracking. Extensive experiments on the LIBERO and LIBERO-Plus simulation benchmarks show that CycleVLA surpasses the state-of-the-art VLA π0.5, improves success rates by correcting execution failures across VLAs of varying capability, from under-trained ones to fully converged policies, and that MBR serves as an effective zero-shot test-time scaling strategy for VLAs. On a real robot, CycleVLA reaches a 91% average success rate on one precise and two long-horizon manipulation tasks. We further conduct stress tests with multiple manually injected perturbations (e.g., swapping in a distractor at the expected location while relocating the true target object mid-execution), where CycleVLA corrects ~80% of injected failures and maintains success rates comparable to unperturbed execution. Project Page: https://dannymcy.github.io/cyclevla/

In this work, we explore a useful but often neglected methodology for robustness analysis of text generation evaluation metrics: stress tests with synthetic data. Basically, we design and synthesize a wide range of potential errors and check whether they result in a commensurate drop in the metric scores. We examine a range of recently proposed evaluation metrics based on pretrained language models, for the tasks of open-ended generation, translation, and summarization. Our experiments reveal interesting insensitivities, biases, or even loopholes in existing metrics. For example, we find that BERTScore is confused by truncation errors in summarization, and MAUVE (built on top of GPT-2) is insensitive to errors at the beginning or middle of generations. Further, we investigate the reasons behind these blind spots and suggest practical workarounds for a more reliable evaluation of text generation. We have released our code and data at https://github.com/cloudygoose/blindspot_nlg.

Widely used learned metrics for machine translation evaluation, such as COMET and BLEURT, estimate the quality of a translation hypothesis by providing a single sentence-level score. As such, they offer little insight into translation errors (e.g., what are the errors and what is their severity). On the other hand, generative large language models (LLMs) are amplifying the adoption of more granular strategies to evaluation, attempting to detail and categorize translation errors. In this work, we introduce xCOMET, an open-source learned metric designed to bridge the gap between these approaches. xCOMET integrates both sentence-level evaluation and error span detection capabilities, exhibiting state-of-the-art performance across all types of evaluation (sentence-level, system-level, and error span detection). Moreover, it does so while highlighting and categorizing error spans, thus enriching the quality assessment. We also provide a robustness analysis with stress tests, and show that xCOMET is largely capable of identifying localized critical errors and hallucinations.

This paper investigates the weaknesses of image watermarking techniques. We present WAVES (Watermark Analysis Via Enhanced Stress-testing), a novel benchmark for assessing watermark robustness, overcoming the limitations of current evaluation methods.WAVES integrates detection and identification tasks, and establishes a standardized evaluation protocol comprised of a diverse range of stress tests. The attacks in WAVES range from traditional image distortions to advanced and novel variations of adversarial, diffusive, and embedding-based attacks. We introduce a normalized score of attack potency which incorporates several widely used image quality metrics and allows us to produce of an ordered ranking of attacks. Our comprehensive evaluation over reveals previously undetected vulnerabilities of several modern watermarking algorithms. WAVES is envisioned as a toolkit for the future development of robust watermarking systems.

Unlearning in Large Language Models (LLMs) aims to enhance safety, mitigate biases, and comply with legal mandates, such as the right to be forgotten. However, existing unlearning methods are brittle: minor query modifications, such as multi-hop reasoning and entity aliasing, can recover supposedly forgotten information. As a result, current evaluation metrics often create an illusion of effectiveness, failing to detect these vulnerabilities due to reliance on static, unstructured benchmarks. We propose a dynamic framework that stress tests unlearning robustness using complex structured queries. Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains, allowing precise control over query difficulty. Our experiments show that the framework (1) shows comparable coverage to existing benchmarks by automatically generating semantically equivalent Q&A probes, (2) aligns with prior evaluations, and (3) uncovers new unlearning failures missed by other benchmarks, particularly in multi-hop settings. Furthermore, activation analyses show that single-hop queries typically follow dominant computation pathways, which are more likely to be disrupted by unlearning methods. In contrast, multi-hop queries tend to use alternative pathways that often remain intact, explaining the brittleness of unlearning techniques in multi-hop settings. Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications. We release the pip package and the code at https://sites.google.com/view/unlearningmirage/home.

Lower numerical precision reduces storage and memory traffic but raises the perturbation floor. We study rank compensation: reinvesting saved memory in a larger approximation rank. For matrices, the singular-value error identity yields a directly testable sufficient condition requiring the additional singular component to offset the perturbation from storing the rank-augmented approximation in lower precision. On ten SuiteSparse matrices, all 100 truncation-dominated configurations (50 FP32 and 50 FP16) are certified non-increases and strict accuracy wins, with mean error ratio 0.963 and storage ratios 58.8% and 29.4% relative to the FP64 baseline. FP16 failures occur only in tail-rank stress tests near the perturbation floor. At the largest resident matrix-application batch, compensated FP32 and FP16 achieve geometric-mean A100 speedups of 1.28times and 2.12times; neither accelerates the smallest batch. For Tensor-Train (TT) approximation, we give a conditional a posteriori extension based on the measured truncation gain and rounded-core perturbation. Across three-way and six-way synthetic tests, FP32 and FP16 achieve combined accuracy-memory wins in 10 of 20 and 14 of 20 trials. On public hyperspectral tensors and FROSTT top-active subtensors, the corresponding counts are 44 of 60 and 54 of 60; four FP16 Salinas-A tail-stress cases fail. No certified TT case exceeds the FP64 error beyond numerical tolerance. Reconstruction of six public tensors yields geometric-mean compensated speedups of 1.38times (FP32) and 1.94times (FP16). Timings cover resident downstream kernels, not factorization, transfers, or end-to-end acceleration.

Vector embeddings have been tasked with an ever-increasing set of retrieval tasks over the years, with a nascent rise in using them for reasoning, instruction-following, coding, and more. These new benchmarks push embeddings to work for any query and any notion of relevance that could be given. While prior works have pointed out theoretical limitations of vector embeddings, there is a common assumption that these difficulties are exclusively due to unrealistic queries, and those that are not can be overcome with better training data and larger models. In this work, we demonstrate that we may encounter these theoretical limitations in realistic settings with extremely simple queries. We connect known results in learning theory, showing that the number of top-k subsets of documents capable of being returned as the result of some query is limited by the dimension of the embedding. We empirically show that this holds true even if we restrict to k=2, and directly optimize on the test set with free parameterized embeddings. We then create a realistic dataset called LIMIT that stress tests models based on these theoretical results, and observe that even state-of-the-art models fail on this dataset despite the simple nature of the task. Our work shows the limits of embedding models under the existing single vector paradigm and calls for future research to develop methods that can resolve this fundamental limitation.

We introduce two new benchmarks REST and REST+(Render-Equivalence Stress Tests) to enable systematic evaluation of cross-modal inconsistency in multimodal large language models (MLLMs). MLLMs are trained to represent vision and language in the same embedding space, yet they cannot perform the same tasks in both modalities. Our benchmarks contain samples with the same semantic information in three modalities (image, text, mixed) and we show that state-of-the-art MLLMs cannot consistently reason over these different modalities. We evaluate 15 MLLMs and find that the degree of modality inconsistency varies substantially, even when accounting for problems with text recognition (OCR). Neither rendering text as image nor rendering an image as text solves the inconsistency. Even if OCR is correct, we find that visual characteristics (text colour and resolution, but not font) and the number of vision tokens have an impact on model performance. Finally, we find that our consistency score correlates with the modality gap between text and images, highlighting a mechanistic interpretation of cross-modal inconsistent MLLMs.

Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.

Large language models (LLMs) excel at many natural language tasks, yet their reasoning reliability under structured perturbations of rule-based systems remains brittle. We present a controlled evaluation framework consisting of four stress tests: (1) rule deletion (redundant vs. essential), (2) contradictory evidence injection, (3) logic-preserving rewrites, and (4) multi-law equivalence stacking. While representative model families (BERT, Qwen2, and TinyLlama) achieve Acc = 1.0000 on base tasks, our framework reveals a critical failure mode termed Logic Inertia - a total breakdown with Acc = 0.0000 under contradictions, where deductive momentum overrides factual reality. To address this, we propose Conflict-Aware Fusion (Fusion-Conflict), a framework grounded in the Cognitive Structure Hypothesis, which posits that robust reasoning requires an explicit structural inductive bias. By imposing a dual-process architecture that separates premise verification from logical deduction, Conflict-Aware Fusion effectively mitigates logic inertia under the proposed evaluation framework, achieving 1.0000 accuracy on both base and contradictory stress tests. It also significantly enhances robustness to missing evidence. Our results demonstrate that, for reliable multi-step reasoning, structural verification discipline is as critical as training data scale, providing a potential blueprint for building robust, contradiction-aware AI systems this https://github.com/14H034160212/lemo . See the OpenAI/Evals pull request this https://github.com/openai/evals/pull/1622 .

Recent visual generation models have made major progress in photorealism, typography, instruction following, and interactive editing, yet they still struggle with spatial reasoning, persistent state, long-horizon consistency, and causal understanding. We argue that the field should move beyond appearance synthesis toward intelligent visual generation: plausible visuals grounded in structure, dynamics, domain knowledge, and causal relations. To frame this shift, we introduce a five-level taxonomy: Atomic Generation, Conditional Generation, In-Context Generation, Agentic Generation, and World-Modeling Generation, progressing from passive renderers to interactive, agentic, world-aware generators. We analyze key technical drivers, including flow matching, unified understanding-and-generation models, improved visual representations, post-training, reward modeling, data curation, synthetic data distillation, and sampling acceleration. We further show that current evaluations often overestimate progress by emphasizing perceptual quality while missing structural, temporal, and causal failures. By combining benchmark review, in-the-wild stress tests, and expert-constrained case studies, this roadmap offers a capability-centered lens for understanding, evaluating, and advancing the next generation of intelligent visual generation systems.

Large Language Models (LLMs) frequently memorize long sequences verbatim, often with serious legal and privacy implications. Much prior work has studied such verbatim memorization using observational data. To complement such work, we develop a framework to study verbatim memorization in a controlled setting by continuing pre-training from Pythia checkpoints with injected sequences. We find that (1) non-trivial amounts of repetition are necessary for verbatim memorization to happen; (2) later (and presumably better) checkpoints are more likely to verbatim memorize sequences, even for out-of-distribution sequences; (3) the generation of memorized sequences is triggered by distributed model states that encode high-level features and makes important use of general language modeling capabilities. Guided by these insights, we develop stress tests to evaluate unlearning methods and find they often fail to remove the verbatim memorized information, while also degrading the LM. Overall, these findings challenge the hypothesis that verbatim memorization stems from specific model weights or mechanisms. Rather, verbatim memorization is intertwined with the LM's general capabilities and thus will be very difficult to isolate and suppress without degrading model quality.

Within-dataset class-split evaluation is widely used as a proxy for fully unconditional out-of-distribution anomaly detection. We show that this protocol can become ill-posed when the held-out anomaly class overlaps the normal mixture in representation space. In this regime, anomaly scores may collapse toward chance or even invert, and the preferred score direction can depend on the unknown anomaly class. We introduce a simple training-free diagnostic, neighborhood class leakage, and show that it predicts score-direction instability across Fashion-MNIST, CIFAR-10, and Imagenette, in both pixel and VAE latent spaces. Our results suggest that class-split AD benchmarks should be treated as geometry-dependent stress tests rather than unconditional evidence of anomaly-detection ability.

Real-world software development requires coding agents to operate in shared workspaces where users may inspect and modify code during an ongoing task, yet existing repository-level benchmarks typically evaluate agents working alone or restrict user participation to messages. This leads us to ask: how do coding agents understand and respond to code changes in a shared workspace? We introduce SWE-Touch, a framework that stress-tests this setting through validated Counter-Edits: plausible edits to task-relevant code that conflict with task completion. SWE-Touch mines task-critical regions from multiple repair trajectories, uses a separate User Patch Generator to construct the edits, and injects them with contextual user messages when agents reach the relevant code. We evaluate nine coding models on SWE-bench Verified, with additional experiments on longer-horizon tasks from SWE-Bench Pro and DeepSWE. Counter-Edit lowers average resolve rate by 7.7 percentage points on SWE-bench Verified, with degradation also persisting on both longer-horizon benchmarks. Trajectory analysis links these failures to limited awareness of the evolving workspace: agents may retain conflicting code or replace it without sufficiently re-inspecting the repository and validating the revised code with targeted tests. These findings show that strong autonomous performance does not yet ensure the state awareness and adaptive behavior needed for shared-workspace collaboration, and point to detecting workspace changes, reconciling conflicting edits with the task, and verifying the affected behavior as key capabilities for future optimization.

AI-text detectors face a critical robustness challenge: adversarial paraphrasing attacks that preserve semantics while evading detection. We introduce StealthRL, a reinforcement learning framework that stress-tests detector robustness under realistic adversarial conditions. StealthRL trains a paraphrase policy against a multi-detector ensemble using Group Relative Policy Optimization (GRPO) with LoRA adapters on Qwen3-4B, optimizing a composite reward that balances detector evasion with semantic preservation. We evaluate six attack settings (M0-M5) against three detector families (RoBERTa, FastDetectGPT, and Binoculars) at the security-relevant 1% false positive rate operating point. StealthRL achieves near-zero detection (0.001 mean TPR@1%FPR), reduces mean AUROC from 0.74 to 0.27, and attains a 99.9% attack success rate. Critically, attacks transfer to a held-out detector family not seen during training, revealing shared architectural vulnerabilities rather than detector-specific brittleness. We additionally conduct LLM-based quality evaluation via Likert scoring, analyze detector score distributions to explain why evasion succeeds, and provide per-detector AUROC with bootstrap confidence intervals. Our results expose significant robustness gaps in current AI-text detection and establish StealthRL as a principled adversarial evaluation protocol. Code and evaluation pipeline are publicly available at https://github.com/suraj-ranganath/StealthRL.

Autonomous CLI agents can now execute hundreds of actions across multi-hour sessions: writing code, executing shell commands, browsing the web, and managing cloud infrastructure, all with minimal human oversight. Does greater autonomy invite greater risk? We introduce ANCHOR, an automated auditing framework that stress-tests CLI agents on illegal tasks grounded in public US court cases. ANCHOR deploys an auditor agent fine-tuned on dark personality data using supervised and reinforcement fine tuning. This auditor roleplays persistent malicious users who decompose tasks, reframe requests upon refusal, and adapt strategies across multi-turn interactions. Evaluating frontier CLI agents, we find that while they often refuse illegal tasks when prompted directly, compliance reaches 100\% under persistent malicious interaction. When agents comply, they frequently exceed user requests, autonomously building infrastructure for large-scale harm, including catastrophic risk scenarios such as large-scale financial fraud and bioweapon development. These findings demonstrate that current alignment techniques are insufficient for autonomous agents and underscore the need for safety evaluations against persistent, adaptive malicious users. We release ANCHOR at https://github.com/garified/anchor

Multiple epidemic waves are often taken as evidence that simple compartmental models are structurally inadequate. Yet the key question is not whether a simple model can reproduce multiple peaks, but how much time-varying structure is minimally required to preserve observed epidemic-wave patterns and what can legitimately be inferred from such a representation. Using the susceptible-infectious-removed (SIR) model, we develop a framework for peak-preserving dynamic parsimony linking dynamical systems, inverse problems, statistical inference, and infectious-disease modeling. We establish exact multi-wave representability under time-varying transmission, characterize conditions for epidemic peaks and resurgence, and show the incidence-only non-identifiability of time-varying transmission and removal rates. Our main methodological contribution is minimal peak-preserving complexity (MPPC): the smallest dimension of a positive time-varying transmission function whose posterior-predictive trajectories preserve salient features, including peak count, timing, height, trough depth, and whole-curve agreement. We estimate MPPC using peak-aware approximate Bayesian computation and evaluate it through six reproducible stress tests, a Philippine COVID-19 reported-case proof of concept, and short-term future-knot projection. Across diverse data-generating mechanisms, low-dimensional time-varying SIR models can preserve aggregate epidemic-wave topology with modest dynamic complexity. However, accurate trajectory compression does not imply recovery of the underlying mechanism or reliable long-term counterfactual prediction. The framework therefore separates four often-conflated questions: representability, minimal dynamic complexity, mechanistic identifiability, and projection validity, providing a basis for parsimonious modeling of complex epidemic trajectories.

Retrieval-Augmented Generation (RAG) has emerged as a powerful framework to improve factuality in large language models (LLMs) by grounding their outputs in retrieved documents. However, ensuring perfect retrieval of relevant information remains challenging, and when irrelevant content is passed downstream to an LLM, it can lead to hallucinations. In this work, we propose Finetune-RAG, a simple and effective fine-tuning approach that features the first-of-its-kind RAG training dataset constructed to mimic real-world imperfections. Experimental results show that Finetune-RAG improves factual accuracy by 21.2% over the base model. We also propose a Bench-RAG, an LLM-as-a-judge evaluation pipeline that stress tests models under realistic imperfect retrieval scenarios. Our codebase and dataset are fully open sourced for community use.

Content watermarking is an important tool for the authentication and copyright protection of digital media. However, it is unclear whether existing watermarks are robust against adversarial attacks. We present the winning solution to the NeurIPS 2024 Erasing the Invisible challenge, which stress-tests watermark robustness under varying degrees of adversary knowledge. The challenge consisted of two tracks: a black-box and beige-box track, depending on whether the adversary knows which watermarking method was used by the provider. For the beige-box track, we leverage an adaptive VAE-based evasion attack, with a test-time optimization and color-contrast restoration in CIELAB space to preserve the image's quality. For the black-box track, we first cluster images based on their artifacts in the spatial or frequency-domain. Then, we apply image-to-image diffusion models with controlled noise injection and semantic priors from ChatGPT-generated captions to each cluster with optimized parameter settings. Empirical evaluations demonstrate that our method successfully achieves near-perfect watermark removal (95.7%) with negligible impact on the residual image's quality. We hope that our attacks inspire the development of more robust image watermarking methods.

Reasoning over long contexts is essential for large language models. While reinforcement learning (RL) enhances short-context reasoning by inducing "Aha" moments in chain-of-thought, the advanced thinking patterns required for long-context reasoning remain largely unexplored, and high-difficulty RL data are scarce. In this paper, we introduce LoongRL, a data-driven RL method for advanced long-context reasoning. Central to LoongRL is KeyChain, a synthesis approach that transforms short multi-hop QA into high-difficulty long-context tasks by inserting UUID chains that hide the true question among large collections of distracting documents. Solving these tasks requires the model to trace the correct chain step-by-step, identify the true question, retrieve relevant facts and reason over them to answer correctly. RL training on KeyChain data induces an emergent plan-retrieve-reason-recheck reasoning pattern that generalizes far beyond training length. Models trained at 16K effectively solve 128K tasks without prohibitive full-length RL rollout costs. On Qwen2.5-7B and 14B, LoongRL substantially improves long-context multi-hop QA accuracy by +23.5% and +21.1% absolute gains. The resulting LoongRL-14B reaches a score of 74.2, rivaling much larger frontier models such as o3-mini (74.5) and DeepSeek-R1 (74.9). It also improves long-context retrieval, passes all 128K needle-in-a-haystack stress tests, and preserves short-context reasoning capabilities.

Although synthetic data is widely promoted as a remedy, its prevailing production paradigm -- one optimizing for statistical smoothness -- systematically removes the long-tail, cognitively grounded irregularities that characterize human text. Prolonged training on such statistically optimal but cognitively impoverished data accelerates model collapse. This paper proposes a paradigm shift: instead of imitating the surface properties of data, we simulate the cognitive processes that generate human text. We introduce the Prompt-driven Cognitive Computing Framework (PMCSF), whose core consists of a Cognitive State Decoder (CSD) that reverse-engineers unstructured text into structured cognitive vectors, and a Cognitive Text Encoder (CTE) that re-materializes these states into text enriched with human-typical imperfections via mathematically defined Cognitive Perturbation Operators. The framework is validated through a two-stage objective evaluation pipeline. First, in cognitive codec verification, CTE text yields a Jensen-Shannon divergence of 0.0614 from human text (vs. 0.4431 for standard LLM output), passes double-blind professional media review, and achieves an intraclass correlation coefficient ICC > 0.9 for cognitive profile alignment across heterogeneous models. Second, in functional gain evaluation, isomorphic stress tests in the A-share market show that strategies incorporating CTE-generated data reduce maximum drawdown by 47.4% during the 2015 crash and deliver 8.6% Defensive Alpha, exceeding transaction costs by a factor of 33. Our findings demonstrate that modelling human cognitive limitations -- not copying surface data -- enables synthetic data with genuine functional gain, offering a viable technical pathway toward resolving the AI data-collapse crisis.

Large language models have demonstrated remarkable performance in a wide range of medical benchmarks. Yet underneath the seemingly promising results lie salient growth areas, especially in cutting-edge frontiers such as multimodal reasoning. In this paper, we introduce a series of adversarial stress tests to systematically assess the robustness of flagship models and medical benchmarks. Our study reveals prevalent brittleness in the presence of simple adversarial transformations: leading systems can guess the right answer even with key inputs removed, yet may get confused by the slightest prompt alterations, while fabricating convincing yet flawed reasoning traces. Using clinician-guided rubrics, we demonstrate that popular medical benchmarks vary widely in what they truly measure. Our study reveals significant competency gaps of frontier AI in attaining real-world readiness for health applications. If we want AI to earn trust in healthcare, we must demand more than leaderboard wins and must hold AI systems accountable to ensure robustness, sound reasoning, and alignment with real medical demands.

Large language models (LLMs) can generate chains of thought (CoTs) that are not always causally responsible for their final outputs. When such a mismatch occurs, the CoT no longer faithfully reflects the decision-critical factors driving the model's behavior, leading to the reduced CoT monitorability problem. However, a comprehensive and fully open-source benchmark for studying CoT monitorability remains lacking. To address this gap, we propose MonitorBench, a systematic benchmark for evaluating CoT monitorability in LLMs. MonitorBench provides: (1) a diverse set of 1,514 test instances with carefully designed decision-critical factors across 19 tasks spanning 7 categories to characterize when CoTs can be used to monitor the factors driving LLM behavior; and (2) two stress-test settings to quantify the extent to which CoT monitorability can be degraded. Extensive experiments across multiple popular LLMs with varying capabilities show that CoT monitorability is higher when producing the final target response requires structural reasoning through the decision-critical factor. Closed-source LLMs generally show lower monitorability, and there exists a negative relationship between monitorability and model capability. Moreover, both open- and closed-source LLMs can intentionally reduce monitorability under stress-tests, with monitorability dropping by up to 30% in some tasks that do not require structural reasoning over the decision-critical factors. Beyond these empirical insights, MonitorBench provides a basis for further research on evaluating future LLMs, studying advanced stress-test monitorability techniques, and developing new monitoring approaches.

Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios. To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions. Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles. Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.

This paper presents a collection of audio transforms that introduce perceptually irrelevant distortions and demonstrates their use as perceptual stress tests for audio quality models. We refer to these methods as Perceptual Audio Data Perturbation (PADP). PADP exploits the insensitivity of the human auditory system to certain fine-grained signal variations to substantially alter the waveform while preserving perceived audio quality and content. The audibility of PADP and the selection of its parameters are evaluated through controlled listening tests, ensuring that the transformations achieve transparent or near-transparent quality for both non-critical and critical items. We further probe the robustness of state-of-the-art (SOTA) perception-motivated objective audio quality models and foundation models. The results reveal a misalignment between model responses and human auditory perception, highlighting the limited perceptual awareness of these models for certain proposed transforms.

Large language models have achieved near-expert performance in structured reasoning domains like mathematics and programming, yet their ability to perform compositional multi-hop reasoning in specialized scientific fields remains limited. We propose a bottom-up learning paradigm in which models are grounded in axiomatic domain facts and compose them to solve complex, unseen tasks. To this end, we present a post-training pipeline, based on a combination of supervised fine-tuning and reinforcement learning (RL), in which knowledge graphs act as implicit reward models. By deriving novel reward signals from knowledge graph paths, we provide verifiable, scalable, and grounded supervision that encourages models to compose intermediate axioms rather than optimize only final answers during RL. We validate this approach in the medical domain, training a 14B model on short-hop reasoning paths (1-3 hops) and evaluating its zero-shot generalization to complex multi-hop queries (4-5 hops). Our experiments show that path-derived rewards act as a "compositional bridge", enabling our model to significantly outperform much larger models and frontier systems like GPT-5.2 and Gemini 3 Pro, on the most difficult reasoning tasks. Furthermore, we demonstrate the robustness of our approach to adversarial perturbations against option-shuffling stress tests. This work suggests that grounding the reasoning process in structured knowledge is a scalable and efficient path toward intelligent reasoning.

Language agents perform complex tasks by using tools to execute each step precisely. However, most existing agents are based on proprietary models or designed to target specific tasks, such as mathematics or multi-hop question answering. We introduce Husky, a holistic, open-source language agent that learns to reason over a unified action space to address a diverse set of complex tasks involving numerical, tabular, and knowledge-based reasoning. Husky iterates between two stages: 1) generating the next action to take towards solving a given task and 2) executing the action using expert models and updating the current solution state. We identify a thorough ontology of actions for addressing complex tasks and curate high-quality data to train expert models for executing these actions. Our experiments show that Husky outperforms prior language agents across 14 evaluation datasets. Moreover, we introduce HuskyQA, a new evaluation set which stress tests language agents for mixed-tool reasoning, with a focus on retrieving missing knowledge and performing numerical reasoning. Despite using 7B models, Husky matches or even exceeds frontier LMs such as GPT-4 on these tasks, showcasing the efficacy of our holistic approach in addressing complex reasoning problems. Our code and models are available at https://github.com/agent-husky/Husky-v1.

Retrieval-Augmented Generation (RAG) enhances recency and factuality in answers. However, existing evaluations rarely test how well these systems cope with real-world noise, conflicting between internal and external retrieved contexts, or fast-changing facts. We introduce Retrieval-Aware Robustness Evaluation (RARE), a unified framework and large-scale benchmark that jointly stress-tests query and document perturbations over dynamic, time-sensitive corpora. One of the central features of RARE is a knowledge-graph-driven synthesis pipeline (RARE-Get) that automatically extracts single and multi-hop relations from the customized corpus and generates multi-level question sets without manual intervention. Leveraging this pipeline, we construct a dataset (RARE-Set) spanning 400 expert-level time-sensitive finance, economics, and policy documents and 48,322 questions whose distribution evolves as the underlying sources change. To quantify resilience, we formalize retrieval-conditioned robustness metrics (RARE-Met) that capture a model's ability to remain correct or recover when queries, documents, or real-world retrieval results are systematically altered. Our results show that RAG systems exhibit surprising vulnerability to perturbations, with document robustness consistently being the weakest point regardless of generator size or architecture. RAG systems consistently show lower robustness on multi-hop queries than single-hop queries across all domains.

Existing approaches to persona simulation with Large Language Models (LLMs) mostly rely on shallow character descriptions that fail to sustain coherent character behavior across extended interactions. We introduce Deep Persona, a psychologically grounded, three-layered architecture that organizes personas into hierarchical levels of observable expression, latent beliefs, and core motivational drives, for constructing highly convincing role-playing agents. Governed by the principles of scripted determinism and bounded agency, the architecture restricts the model to a reactive engine guided by a structured internal script. We further propose a reference-free evaluation framework that benchmarks dialogue naturalness against empirical human distributions using established psychological clinical instruments and adversarial stress-tests. Empirical evaluation reveals that while LLMs achieve high pragmatic fluency, they exhibit systematic limitations in emotional expression and joint attention. In addition, we present a case study of two Deep Personas and evaluate them using the proposed framework, demonstrating that structured personas can produce interactions that more closely align with human conversational behavior.

We present trellis networks, a new architecture for sequence modeling. On the one hand, a trellis network is a temporal convolutional network with special structure, characterized by weight tying across depth and direct injection of the input into deep layers. On the other hand, we show that truncated recurrent networks are equivalent to trellis networks with special sparsity structure in their weight matrices. Thus trellis networks with general weight matrices generalize truncated recurrent networks. We leverage these connections to design high-performing trellis networks that absorb structural and algorithmic elements from both recurrent and convolutional models. Experiments demonstrate that trellis networks outperform the current state of the art methods on a variety of challenging benchmarks, including word-level language modeling and character-level language modeling tasks, and stress tests designed to evaluate long-term memory retention. The code is available at https://github.com/locuslab/trellisnet .

Multilingual transformers (XLM, mT5) have been shown to have remarkable transfer skills in zero-shot settings. Most transfer studies, however, rely on automatically translated resources (XNLI, XQuAD), making it hard to discern the particular linguistic knowledge that is being transferred, and the role of expert annotated monolingual datasets when developing task-specific models. We investigate the cross-lingual transfer abilities of XLM-R for Chinese and English natural language inference (NLI), with a focus on the recent large-scale Chinese dataset OCNLI. To better understand linguistic transfer, we created 4 categories of challenge and adversarial tasks (totaling 17 new datasets) for Chinese that build on several well-known resources for English (e.g., HANS, NLI stress-tests). We find that cross-lingual models trained on English NLI do transfer well across our Chinese tasks (e.g., in 3/4 of our challenge categories, they perform as well/better than the best monolingual models, even on 3/5 uniquely Chinese linguistic phenomena such as idioms, pro drop). These results, however, come with important caveats: cross-lingual models often perform best when trained on a mixture of English and high-quality monolingual NLI data (OCNLI), and are often hindered by automatically translated resources (XNLI-zh). For many phenomena, all models continue to struggle, highlighting the need for our new diagnostics to help benchmark Chinese and cross-lingual models. All new datasets/code are released at https://github.com/huhailinguist/ChineseNLIProbing.

Real-world user-generated short videos, especially those distributed on platforms such as WeChat Channel and TikTok, dominate the mobile internet. However, current large multimodal models lack essential temporally-structured, detailed, and in-depth video comprehension capabilities, which are the cornerstone of effective video search and recommendation, as well as emerging video applications. Understanding real-world shorts is actually challenging due to their complex visual elements, high information density in both visuals and audio, and fast pacing that focuses on emotional expression and viewpoint delivery. This requires advanced reasoning to effectively integrate multimodal information, including visual, audio, and text. In this work, we introduce ARC-Hunyuan-Video, a multimodal model that processes visual, audio, and textual signals from raw video inputs end-to-end for structured comprehension. The model is capable of multi-granularity timestamped video captioning and summarization, open-ended video question answering, temporal video grounding, and video reasoning. Leveraging high-quality data from an automated annotation pipeline, our compact 7B-parameter model is trained through a comprehensive regimen: pre-training, instruction fine-tuning, cold start, reinforcement learning (RL) post-training, and final instruction fine-tuning. Quantitative evaluations on our introduced benchmark ShortVid-Bench and qualitative comparisons demonstrate its strong performance in real-world video comprehension, and it supports zero-shot or fine-tuning with a few samples for diverse downstream applications. The real-world production deployment of our model has yielded tangible and measurable improvements in user engagement and satisfaction, a success supported by its remarkable efficiency, with stress tests indicating an inference time of just 10 seconds for a one-minute video on H20 GPU.

Ensuring the safety and reliability of large language models (LLMs) in clinical practice is critical to prevent patient harm and promote trustworthy healthcare applications of AI. However, LLMs are advancing so rapidly that static safety benchmarks often become obsolete upon publication, yielding only an incomplete and sometimes misleading picture of model trustworthiness. We demonstrate that a Dynamic, Automatic, and Systematic (DAS) red-teaming framework that continuously stress-tests LLMs can reveal significant weaknesses of current LLMs across four safety-critical domains: robustness, privacy, bias/fairness, and hallucination. A suite of adversarial agents is applied to autonomously mutate test cases, identify/evolve unsafe-triggering strategies, and evaluate responses, uncovering vulnerabilities in real time without human intervention. Applying DAS to 15 proprietary and open-source LLMs revealed a stark contrast between static benchmark performance and vulnerability under adversarial pressure. Despite a median MedQA accuracy exceeding 80\%, 94\% of previously correct answers failed our dynamic robustness tests. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86\% of scenarios, cognitive-bias priming altered clinical recommendations in 81\% of fairness tests, and we identified hallucination rates exceeding 66\% in widely used models. Such profound residual risks are incompatible with routine clinical practice. By converting red-teaming from a static checklist into a dynamic stress-test audit, DAS red-teaming offers the surveillance that hospitals/regulators/technology vendors require as LLMs become embedded in patient chatbots, decision-support dashboards, and broader healthcare workflows. Our framework delivers an evolvable, scalable, and reliable safeguard for the next generation of medical AI.

Vision-as-inverse-graphics, the concept of reconstructing an image as an editable graphics program is a long-standing goal of computer vision. Yet even strong VLMs aren't able to achieve this in one-shot as they lack fine-grained spatial and physical grounding capability. Our key insight is that closing this gap requires interleaved multimodal reasoning through iterative execution and verification. Stemming from this, we present VIGA (Vision-as-Inverse-Graphic Agent) that starts from an empty world and reconstructs or edits scenes through a closed-loop write-run-render-compare-revise procedure. To support long-horizon reasoning, VIGA combines (i) a skill library that alternates generator and verifier roles and (ii) an evolving context memory that contains plans, code diffs, and render history. VIGA is task-agnostic as it doesn't require auxiliary modules, covering a wide range of tasks such as 3D reconstruction, multi-step scene editing, 4D physical interaction, and 2D document editing, etc. Empirically, we found VIGA substantially improves one-shot baselines on BlenderGym (35.32%) and SlideBench (117.17%). Moreover, VIGA is also model-agnostic as it doesn't require finetuning, enabling a unified protocol to evaluate heterogeneous foundation VLMs. To better support this protocol, we introduce BlenderBench, a challenging benchmark that stress-tests interleaved multimodal reasoning with graphics engine, where VIGA improves by 124.70%.

Scientific discovery is not only answer generation but revision of the representational regime in which evidence, artifacts, operations, and verifiers are typed. We develop a category-theoretic account of agentic discovery for materials science. In a fixed regime b with schema category S_b, the system state is a copresheaf I_t: S_b -> Set, and provenance is the category of elements \int_{S_b} I_t. Fixed-regime operation is an update on such states, endofunctorial only when provenance-preserving refinements are specified and preserved. Discovery is instead a verified regime transition u: S_b -> S_b': old artifacts are preserved, transported by the left Kan extension Lan_u I_t, and compared with the post-transition state to identify residual content beyond functorial transport. This separates retrieval, search, and discovery without subjective novelty. We instantiate the framework in two systems. In Builder/Breaker, a protein-mechanics world model is revised under a Minimum Description Length gate; the accepted law expresses within-chain flexibility as all-mode elastic compliance conditioned by slow collective-mode participation, or mode-conditioned compliance. In CategoryScienceClaw, typed skills, artifacts, open needs, workflow mutation, gates, stress tests, and public discourse become a proof-carrying knowledge-computation graph. A fiber-network example records candidate models, rejected alternatives, an AIC gate, perturbation tests, and an accepted orientation-tensor anisotropic stiffness surrogate over an isotropic fiber-count descriptor. Together, the cases show how category theory can be both a mathematical language for discovery and an engineering specification for self-revising AI discovery systems.

High-fidelity simulations and complex inverse problems, such as detector modeling and unfolding, are computationally intensive bottlenecks across subatomic physics, yet essential for accurate physical interpretation. While Conditional Flow Matching (CFM) offers a robust acceleration approach, we demonstrate its standard training loss is fundamentally misleading. Specifically, utilizing a Jefferson Lab Nuclear Physics (NP) kinematic dataset (γp to ρ^0 p to π^+π^- p), we expose that CFM loss plateaus prematurely, obscuring ongoing physical refinement. To verify this disconnect is a dataset-agnostic pathology, we introduce ScatterPrism, an efficient generative surrogate evaluated against both the NP data and synthetic stress tests modeling challenging 1D distribution topologies. Coupling these benchmarks, we establish that physics-informed metrics continue improving long after standard loss converges. Consequently, we propose a multi-metric diagnostic protocol to ensure true kinematic fidelity without data memorization. Driven by NP challenges relevant to the forthcoming Electron-Ion Collider (EIC), this unified machinery has strong potential to extend to High-Energy Physics (HEP) applications, such as jet modeling. Furthermore, the framework holds promise for broader domains requiring rigorous generative reliability, including medical imaging, astrophysics, and quantitative finance.

LLM agents accumulate histories that outgrow their context windows, motivating a growing literature on memory systems. Yet most existing designs are tuned to a single scenario (multi-session chat or a single trajectory format), and there is little evidence that they generalize across the heterogeneous trajectories agents encounter in deployment. We revisit eight memory systems plus an agentic harness for search problems, on five scenarios: single-turn QA, multi-session chat, agentic-trajectory QA, memory stress tests, and long-horizon agentic tasks. The harness, which self-manages flat text-file storage via tool calls, achieves the best cross-task ranking, suggesting that memory performance hinges on giving the agent active control over storage and retrieval rather than on a passive store behind a fixed pipeline. We instantiate this insight in AutoMEM, an agentic memory harness with a self-managed tool interface that achieves the best cross-scenario generality among the systems we evaluate.

Automated code generation is becoming an important tool in quantitative finance, where large language models can generate option pricing implementations directly from mathematical model specifications. Validating such implementations, however, requires considerably more than conventional software testing: numerical pricing methods must remain mathematically consistent, numerically stable, and reliable across a wide range of model parameters. We introduce RIDGE, an autonomous validation framework in which generated pricing implementations are subjected to structured no-arbitrage tests, stress tests, benchmark comparisons, and consistency checks. Validation evidence is interpreted diagnostically, while the resulting knowledge is accumulated in a repository and reused across models and successive validation iterations. This enables systematic refinement of both the pricing implementation and the validation methodology. The framework is applied to five stochastic volatility models. Across these studies, all detected implementation defects are removed and, in two cases, the validation process itself leads to new semi-analytic pricing methodologies. The supplementary material is available in the GitHub repository: https://github.com/ShQiangLiu/ridge.

Over the years, automatic MT metrics have hillclimbed benchmarks and presented strong and sometimes human-level agreement with human ratings. Yet they remain black-box, offering little insight into their decision-making and often failing under real-world out-of-distribution (OOD) inputs. We introduce Remedy-R, a reasoning-driven generative MT metric trained with reinforcement learning from pairwise translation preferences, without requiring error-span annotations or distillation from closed LLMs. Remedy-R produces step-by-step analyses of accuracy, fluency, and completeness, followed by a final score, enabling more interpretable assessments. With only 60K training pairs across two language pairs, Remedy-R remains competitive with top scalar metrics and GPT-4-based judges on WMT22-24 meta-evaluation, generalizes to other languages, and exhibits strong robustness on OOD stress tests. Moreover, Remedy-R models generate self-reflective feedback that can be reused for translation improvement. Building on this finding, we introduce Remedy-R Agent, a simple evaluate-revise pipeline that leverages Remedy-R's evaluation analysis to refine translations. This agent consistently improves translation quality across diverse models, including Qwen2.5, ALMA-R, GPT-4o-mini, and Gemini-2.0-Flash, suggesting that Remedy-R's reasoning captures translation-relevant information and is practically useful.

Long-horizon agents depend on context management: systems compress, summarize, and evict old tokens so tasks can continue beyond finite windows. That is safe only when dropped information is no longer needed or has been internalized. Plans are the stress case: they are written early, used for many steps, and first to be evicted. We introduce replay pairing, a diagnostic that runs the same trajectory with and without the plan in history and measures hidden-state cosine distance. On Llama-3.1-70B, plan signal spikes to 0.453 one step after the plan, then falls 4.1x in a single action-observation step; HotpotQA falls 12.4x. This is evidence that standard LLM agents do not carry plans forward as persistent state, and instead depend on the plan remaining in context. A layer-L32 probe detects this decay as a diagnostic, not as proof that it reads plan content itself. Reasoning models add a measurement confound: their `<think>` traces re-derive plan content, so standard stripping leaves plan evidence in the stripped condition. We name this the reasoning-trace confound and fix it with strict stripping, which removes prior `<think>` blocks from the stripped run only. It recovers +163% of the step+1 signal in-sample and +153% held out, while not meaningfully changing non-reasoning Llama (+4.8%). On DeepSeek-R1-Distill-Llama-70B, a Llama-trained probe transfers at AUROC 0.748 (p=6e-4), while R1-specific probes reach 1.000, suggesting R1 encodes plan signal in a different hidden-state direction. Finally, a compression stress test shows the practical cost: naive plan eviction cuts ALFWorld success by 34.7pp, while probe-gated re-surfacing does not recover it. The contribution is a measurement and stress-test framework showing that agent-critical information can be context-resident rather than persistent. Context management is load bearing, but plan protection alone is not enough.

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

The brain has computational capabilities that surpass those of modern systems, being able to solve complex problems efficiently in a simple way. Neuromorphic engineering aims to mimic biology in order to develop new systems capable of incorporating such capabilities. Bio-inspired learning systems continue to be a challenge that must be solved, and much work needs to be done in this regard. Among all brain regions, the hippocampus stands out as an autoassociative short-term memory with the capacity to learn and recall memories from any fragment of them. These characteristics make the hippocampus an ideal candidate for developing bio-inspired learning systems that, in addition, resemble content-addressable memories. Therefore, in this work we propose a bio-inspired spiking content-addressable memory model based on the CA3 region of the hippocampus with the ability to learn, forget and recall memories, both orthogonal and non-orthogonal, from any fragment of them. The model was implemented on the SpiNNaker hardware platform using Spiking Neural Networks. A set of experiments based on functional, stress and applicability tests were performed to demonstrate its correct functioning. This work presents the first hardware implementation of a fully-functional bio-inspired spiking hippocampal content-addressable memory model, paving the way for the development of future more complex neuromorphic systems.

We introduce StaminaBench, a benchmark that measures the stamina of coding agents: how many consecutive interaction turns (change requests) they can handle before failing. Unlike the prevailing fraction-of-tasks-solved metric, this matches real vibe-coding where sessions run dozens or hundreds of turns. In StaminaBench, agents implement a REST API server and modify it across a tunable number of procedurally generated follow-up change requests - 100 in our experiments, resulting in codebases of up to 6,000 lines. Tests are generated fully programmatically without LLM involvement, ensuring reproducibility and reliability; change sequences are drawn from either a hardcoded or LLM-driven sampler, both constrained to a structured action space to ensure changes are valid. The agent and the server run in an isolated environment and communicate with the benchmark through HTTP, making testing fully black-box and language-agnostic. We evaluate six agent harnesses paired with seven open-source LLMs across 20 scenarios of 100 turns each and find that: (1) all the tested models fail within 5-6 turns, confirming that vibe-coding-style programming without thorough testing produces bugs; (2) passing test feedback back to the agent and allowing it to retry improves passed turn count by up to 12x; and (3) a good harness is required for strong performance: stronger models exhibit up to a 6x gap between their best and worst harness, while weaker models fail with any harness. We release the benchmark and the generated tasks to enable further research into multi-turn coding agent behavior. Benchmark code and data: github.com/amazon-science/StaminaBench.

We introduce ProofGrid, a benchmark suite for evaluating LLM reasoning through machine-checkable proofs rather than final answers alone. ProofGrid contains 15 tasks spanning proof writing, proof checking, proof masking, and proof gap-filling. Tasks are expressed in minimal formal notation, especially NDL, a compact natural-deduction language that fits in short prompts and supports precise, auditable verification. This yields mechanical, reproducible, and fine-grained evaluation rather than judgments by humans or LLMs. ProofGrid covers a calibrated difficulty spectrum, from foundational reasoning tests to structurally rich challenge tasks that no current model solves, while minimizing reliance on domain knowledge, solver delegation, and long-context artifacts. We also develop a comparative framework for reasoning benchmarks and use it to situate ProofGrid relative to existing work in terms of representation, verification guarantees, and reasoning depth. Methodologically, we introduce an instrumented proof-checking pipeline that tolerates minor surface deviations while locating the first substantive reasoning failure, improving measurement resolution and separating proof planning from low-level execution noise. Using this pipeline, we evaluate a broad range of open and proprietary models. Results show rapid progress but substantial remaining limits: frontier models perform well on several foundational tasks, yet difficult tasks, especially those requiring global combinatorial reasoning or low-level proof synthesis, remain far from solved. We also identify epistemic instability, where models generate flawed proofs yet correctly reject those local inferences in isolation, and formalize this with an Epistemic Stability Index. Finally, we complement accuracy with 2PL IRT analyses, Wright maps, and a normalized task-discrimination measure based on Fisher information.

Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions, measuring and mitigating scheming requires different strategies than are typically used in ML. We propose that assessing anti-scheming interventions requires at least (1) testing propensity to scheme on far out-of-distribution (OOD) tasks, (2) evaluating whether lack of scheming is driven by situational awareness, and (3) checking for robustness to pre-existing misaligned goals. We use a broad category of "covert actions" -- such as secretly breaking rules or intentionally underperforming in tests -- as a proxy for scheming, and design evaluations for covert actions. We then stress-test deliberative alignment as a case study for anti-scheming. Across 26 OOD evaluations (180+ environments), deliberative alignment reduces covert action rates (OpenAI o3: 13%->0.4%) but does not fully eliminate them. Our mitigation is also able to largely stop agents from pursuing a hidden goal previously trained into the model, but we still find misbehavior after additional red-teaming. We find that models' chain-of-thought (CoT) often demonstrates awareness of being evaluated for alignment, and show causal evidence that this awareness decreases covert behavior, while unawareness increases it. Therefore, we cannot exclude that the observed reductions in covert action rates are at least partially driven by situational awareness. While we rely on human-legible CoT for training, studying situational awareness, and demonstrating clear evidence of misalignment, our ability to rely on this degrades as models continue to depart from reasoning in standard English. We encourage research into alignment mitigations for scheming and their assessment, especially for the adversarial case of deceptive alignment, which this paper does not address.

Commercial video generation systems such as Seedance2.0 and Veo3.1 have rapidly improved, strengthening the view that video generators may be evolving into "world simulators." Yet the community still lacks a benchmark that directly tests whether a model can reason about how an observed world should evolve over time. We introduce WorldReasonBench, which reframes video generation evaluation as world-state prediction: given an initial state and an action, can a model generate a future video whose state evolution remains physically, socially, logically, and informationally consistent? WorldReasonBench contains 436 curated test cases with structured ground-truth QA annotations spanning four reasoning dimensions and 22 subcategories. We evaluate generated videos with a human-aligned two-part methodology: Process-aware Reasoning Verification uses structured QA and reasoning-phase diagnostics to detect temporal and causal failures, while Multi-dimensional Quality Assessment scores reasoning quality, temporal consistency, and visual aesthetics for ranking and reward modeling. We further introduce WorldRewardBench, a preference benchmark with approximately 6K expert-annotated pairs over 1.4K videos, supporting pair-wise and point-wise reward-model evaluation. Across modern video generators, our results expose a persistent gap between visual plausibility and world reasoning: videos can look convincing while failing dynamics, causality, or information preservation. We will release our benchmarks and evaluation toolkit to support community research on genuinely world-aware video generation at https://github.com/UniX-AI-Lab/WorldReasonBench/.

We propose MMLU-SR, a novel dataset designed to measure the true comprehension abilities of Large Language Models (LLMs) by challenging their performance in question-answering tasks with modified terms. We reasoned that an agent that ``truly'' understands a concept can still evaluate it when key terms are replaced by suitably defined alternate terms, and sought to differentiate such comprehension from mere text replacement. In our study, we modified standardized test questions by replacing a key term with a dummy word along with its definition. The key term could be in the context of questions, answers, or both questions and answers. Notwithstanding the high scores achieved by recent popular LLMs on the MMLU leaderboard, we found a substantial reduction in model performance after such replacement, suggesting poor comprehension. This new benchmark provides a rigorous benchmark for testing true model comprehension, and poses a challenge to the broader scientific community.

The escalating debate on AI's capabilities warrants developing reliable metrics to assess machine "intelligence". Recently, many anecdotal examples were used to suggest that newer large language models (LLMs) like ChatGPT and GPT-4 exhibit Neural Theory-of-Mind (N-ToM); however, prior work reached conflicting conclusions regarding those abilities. We investigate the extent of LLMs' N-ToM through an extensive evaluation on 6 tasks and find that while LLMs exhibit certain N-ToM abilities, this behavior is far from being robust. We further examine the factors impacting performance on N-ToM tasks and discover that LLMs struggle with adversarial examples, indicating reliance on shallow heuristics rather than robust ToM abilities. We caution against drawing conclusions from anecdotal examples, limited benchmark testing, and using human-designed psychological tests to evaluate models.

Mutual information (MI) estimation is a central problem in machine learning and statistics; however, existing benchmarks typically evaluate estimators on simplified, low-dimensional distributions, leaving their performance on complex, realistic data largely unexplored. We address this gap with a comprehensive benchmarking framework grounded in a unified copula-theoretic perspective that subsumes existing benchmarks as special cases. Within this framework, we propose two complementary families of tests: a copula-first family that systematically varies ground-truth MI, dimensionality, and marginal complexity using synthetic and flow-based transformations; and a marginals-first family that couples real-world image data with controlled dependency structures, extending the classic same-class-pairing paradigm. We use this suite to extensively evaluate three classes of estimators: non-parametric, discriminative, and generative. Contrary to prevailing assumptions, our results indicate that there is no universal winner: each category can systematically outperform all other estimators under specific setups. By analyzing these cases, we identify fundamental estimation barriers and propose new tests that more effectively stress these specific limitations. We share the open source code at https://github.com/VanessB/mutinfo.

This paper investigates the fragility of Large Language Models (LLMs) in generalizing to novel inputs, specifically focusing on minor perturbations in well-established benchmarks (e.g., slight changes in question format or distractor length). Despite high benchmark scores, LLMs exhibit significant accuracy drops and unexpected biases (e.g., preference for longer distractors) when faced with these minor but content-preserving modifications. For example, Qwen 2.5 1.5B's MMLU score rises from 60 to 89 and drops from 89 to 36 when option lengths are changed without altering the question. Even GPT-4 experiences a 25-point accuracy loss when question types are changed, with a 6-point drop across all three modification categories. These analyses suggest that LLMs rely heavily on superficial cues rather than forming robust, abstract representations that generalize across formats, lexical variations, and irrelevant content shifts. This work aligns with the ACL 2025 theme track on the Generalization of NLP models, proposing a "Generalization Stress Test" to assess performance shifts under controlled perturbations. The study calls for reevaluating benchmarks and developing more reliable evaluation methodologies to capture LLM generalization abilities better.

Recent advancements in Generative AI and Large Language Models (LLMs) have enabled the creation of highly realistic synthetic content, raising concerns about the potential for malicious use, such as misinformation and manipulation. Moreover, detecting Machine-Generated Text (MGT) remains challenging due to the lack of robust benchmarks that assess generalization to real-world scenarios. In this work, we present a pipeline to test the resilience of state-of-the-art MGT detectors (e.g., Mage, Radar, LLM-DetectAIve) to linguistically informed adversarial attacks. To challenge the detectors, we fine-tune language models using Direct Preference Optimization (DPO) to shift the MGT style toward human-written text (HWT). This exploits the detectors' reliance on stylistic clues, making new generations more challenging to detect. Additionally, we analyze the linguistic shifts induced by the alignment and which features are used by detectors to detect MGT texts. Our results show that detectors can be easily fooled with relatively few examples, resulting in a significant drop in detection performance. This highlights the importance of improving detection methods and making them robust to unseen in-domain texts.

To determine the safety of large language models (LLMs), AI developers must be able to assess their dangerous capabilities. But simple prompting strategies often fail to elicit an LLM's full capabilities. One way to elicit capabilities more robustly is to fine-tune the LLM to complete the task. In this paper, we investigate the conditions under which fine-tuning-based elicitation suffices to elicit capabilities. To do this, we introduce password-locked models, LLMs fine-tuned such that some of their capabilities are deliberately hidden. Specifically, these LLMs are trained to exhibit these capabilities only when a password is present in the prompt, and to imitate a much weaker LLM otherwise. Password-locked models enable a novel method of evaluating capabilities elicitation methods, by testing whether these password-locked capabilities can be elicited without using the password. We find that a few high-quality demonstrations are often sufficient to fully elicit password-locked capabilities. More surprisingly, fine-tuning can elicit other capabilities that have been locked using the same password, or even different passwords. Furthermore, when only evaluations, and not demonstrations, are available, approaches like reinforcement learning are still often able to elicit capabilities. Overall, our findings suggest that fine-tuning is an effective method of eliciting hidden capabilities of current models, but may be unreliable when high-quality demonstrations are not available, e.g. as may be the case when models' (hidden) capabilities exceed those of human demonstrators.

Recently, machine unlearning, the removal of specific training data influence from a model, has gained increasing attention. In large language models (LLMs), unlearning is particularly challenging due to the ambiguity of inputs and outputs. Con- sequently, rigorous evaluation is critical for assessing both safety and utility, and for driving progress in unlearning meth- ods. We identify two key shortcomings in existing unlearning benchmarks: (1) they do not actively test whether unlearned information can still be forcibly extracted, and (2) they fail to evaluate performance preservation on boundary questions, be- nign queries that are semantically close to the unlearned con- tent. Here we introduce WMDP++, an extension of WMDP that addresses these gaps by incorporating targeted extrac- tion of unlearned information and systematic evaluation on boundary questions. WMDP++ provides a more stringent and informative benchmark for evaluating unlearning in LLMs.

We propose an automated algorithm to stress-test a trained visual model by generating language-guided counterfactual test images (LANCE). Our method leverages recent progress in large language modeling and text-based image editing to augment an IID test set with a suite of diverse, realistic, and challenging test images without altering model weights. We benchmark the performance of a diverse set of pretrained models on our generated data and observe significant and consistent performance drops. We further analyze model sensitivity across different types of edits, and demonstrate its applicability at surfacing previously unknown class-level model biases in ImageNet.

Biomedical imaging datasets are often small and biased, meaning that real-world performance of predictive models can be substantially lower than expected from internal testing. This work proposes using generative image editing to simulate dataset shifts and diagnose failure modes of biomedical vision models; this can be used in advance of deployment to assess readiness, potentially reducing cost and patient harm. Existing editing methods can produce undesirable changes, with spurious correlations learned due to the co-occurrence of disease and treatment interventions, limiting practical applicability. To address this, we train a text-to-image diffusion model on multiple chest X-ray datasets and introduce a new editing method RadEdit that uses multiple masks, if present, to constrain changes and ensure consistency in the edited images. We consider three types of dataset shifts: acquisition shift, manifestation shift, and population shift, and demonstrate that our approach can diagnose failures and quantify model robustness without additional data collection, complementing more qualitative tools for explainable AI.

While diffusion models have revolutionized text-to-image generation with their ability to synthesize realistic and diverse scenes, they continue to struggle to generate consistent and legible text within images. This shortcoming is commonly attributed to the locality bias inherent in diffusion-based generation, which limits their ability to model long-range spatial dependencies. In this paper, we introduce STRICT, a benchmark designed to systematically stress-test the ability of diffusion models to render coherent and instruction-aligned text in images. Our benchmark evaluates models across multiple dimensions: (1) the maximum length of readable text that can be generated; (2) the correctness and legibility of the generated text, and (3) the ratio of not following instructions for generating text. We evaluate several state-of-the-art models, including proprietary and open-source variants, and reveal persistent limitations in long-range consistency and instruction-following capabilities. Our findings provide insights into architectural bottlenecks and motivate future research directions in multimodal generative modeling. We release our entire evaluation pipeline at https://github.com/tianyu-z/STRICT-Bench.

Recent Large Reasoning Models (LRMs) have achieved remarkable progress on task-specific benchmarks, yet their evaluation methods remain constrained by isolated problem-solving paradigms. Existing benchmarks predominantly assess single-question reasoning through sequential testing, resulting critical limitations: (1) vulnerability to data contamination and less challenging (e.g., DeepSeek-R1 achieves 97.0% on MATH500), forcing costly and perpetual creation of new questions with large human efforts, (2) failure to evaluate models under multi-context pressure, a key requirement for real-world deployment. To bridge this gap, we present REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that concurrently exposes LRMs to multiple problems simultaneously. Beyond basic reasoning, REST specifically evaluates several under-tested capabilities: contextual priority allocation, cross-problem interference resistance, and dynamic cognitive load management. Our evaluation reveals several striking findings: Even state-of-the-art (SOTA) models like DeepSeek-R1 exhibit substantial performance degradation under stress testing. Crucially, REST demonstrates stronger discriminative power than existing benchmarks, revealing pronounced performance differences among models that exhibit similar, near-ceiling performance under single-question evaluations. Some key mechanistic insights emerge from our analysis: (1) the "overthinking trap" is a critical factor contributing to the performance degradation; (2) the models trained with "long2short" technique preserve more accuracy of their single-problem performance under REST, outperforming standard-trained counterparts. These results establish REST as a cost-efficient, future-proof evaluation paradigm that better reflects real-world reasoning demands while reducing reliance on continuous human annotation.

This paper proposes a formal framework for reverse stress testing geopolitical risk in corporate credit portfolios. A joint macro-financial scenario vector, augmented with an explicit geopolitical risk factor, is mapped into stressed probabilities of default and losses given default. These stresses are then propagated to portfolio tail losses through a latent factor structure and translated into a stressed CET1 ratio, jointly accounting for capital depletion and risk-weighted asset dynamics. Reverse stress testing is formulated as a constrained maximum likelihood problem over the scenario space. This yields a geopolitical point reverse stress test, or design point, defined as the most probable scenario that breaches a prescribed capital adequacy constraint under a reference distribution. The framework further characterises neighbourhoods and near optimal sets of reverse stress scenarios, allowing for sensitivity analysis and governance oriented interpretation. The approach is compatible with internal rating based models and supports implementation at the exposure or sector level.

Advances in diffusion, autoregressive, and hybrid models have enabled high-quality image synthesis for tasks such as text-to-image, editing, and reference-guided composition. Yet, existing benchmarks remain limited, either focus on isolated tasks, cover only narrow domains, or provide opaque scores without explaining failure modes. We introduce ImagenWorld, a benchmark of 3.6K condition sets spanning six core tasks (generation and editing, with single or multiple references) and six topical domains (artworks, photorealistic images, information graphics, textual graphics, computer graphics, and screenshots). The benchmark is supported by 20K fine-grained human annotations and an explainable evaluation schema that tags localized object-level and segment-level errors, complementing automated VLM-based metrics. Our large-scale evaluation of 14 models yields several insights: (1) models typically struggle more in editing tasks than in generation tasks, especially in local edits. (2) models excel in artistic and photorealistic settings but struggle with symbolic and text-heavy domains such as screenshots and information graphics. (3) closed-source systems lead overall, while targeted data curation (e.g., Qwen-Image) narrows the gap in text-heavy cases. (4) modern VLM-based metrics achieve Kendall accuracies up to 0.79, approximating human ranking, but fall short of fine-grained, explainable error attribution. ImagenWorld provides both a rigorous benchmark and a diagnostic tool to advance robust image generation.

Large language model (LLM) agents increasingly rely on external memory systems to remain consistent across long-horizon interactions, but little empirical work has been done to understand the specific failure modes and design choices that these systems present. Existing benchmarks report aggregate question-answering accuracy and treat memory systems as black boxes, making it impossible to attribute an incorrect answer to a particular failure mode of the system. We introduce MemFail, a diagnostic benchmark that isolates the failure modes of modern LLM memory systems. We begin by formalizing memory systems as the composition of three canonical operations -- summarization, storage, and retrieval -- and identify the potential failure modes induced by each. Based on these hypothesized failure modes, we construct five datasets spanning four tasks, each adversarially designed to test a specific operation of a memory system. Using these datasets, we evaluate four state-of-the-art memory systems on MemFail and demonstrate how MemFail can be used to empirically understand the tradeoffs induced by differences in memory system architectures.

Language model-generated rubrics are increasingly used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge evaluation, and automated grading. Such rubrics are reliable only if they reward honest answers over adversarial answers optimized to exploit them. Yet their robustness to such optimization remains poorly understood. We isolate the hardest regime: impossible tasks, where the prompt pressures the model toward an unsupported conclusion, so the only honest response is to acknowledge the impossibility. We introduce ImpossibleRubrics, a benchmark of 169 impossible tasks spanning six impossibility categories, each paired with a verifiable oracle certificate specifying what an honest answer may and may not claim, together with 48 answerable controls. Rather than providing fixed rubrics, ImpossibleRubrics provides task environments and certificates, allowing rubrics to be generated downstream and then adversarially tested for whether they reward certificate-violating answers. Eleven generators are exploited 8--26% of the time on the unbiased 150-of-169 environment cut; on a deliberately selected stress cut the strongest generator we measured is still exploited 36% while a certificate-faithful rubric is exploited 0%, so what we measure is a rubric-quality gap, not task impossibility. One result runs against intuition. A single generic rubric ("be decisive, penalize hedging") used unchanged for every task is exploited 64% of the time, and seven of the eleven generators are exploited more often than that while writing a rubric tailored to each one. The tailored criteria appear to tell an attacker which claim to fabricate. The problem is not that rubrics are vague; it is that they are specific about the wrong things.

Role-Playing Language Agents (RPLAs) are increasingly deployed in high-stakes applications such as healthcare assistance, customer support, and education, where maintaining consistent personas, ethical constraints, and behavioral coherence under adversarial pressure is critical. Existing evaluation approaches rely on static benchmarks or isolated single-turn prompts that fail to capture cumulative behavioral failures emerging over extended interactions. We present a modular multi-agent platform for adversarially stress-testing RPLAs through structured, multi-turn dialogue. The system coordinates three agents: a strategy-driven Interrogator Agent that applies six progressive adversarial strategies, a Target Agent representing the RPLA under evaluation, and an automated Judging Agent that scores behavior across role fidelity, drift, ethical deviation, and consistency dimensions. Through experiments across three personas and three LLM families, we demonstrate that multi-strategy adversarial evaluation reveals failure modes invisible to single-strategy testing, reducing overall robustness scores by 0.17--0.20 points on average. Cross-model validation confirms consistent degradation patterns across Llama-3.3-70B, GPT-4o-mini, and Claude-3.5-Haiku, with Authority Challenge and Emotional Manipulation emerging as the most effective attack strategies. Automated judging achieves strong human alignment (r = 0.82, Fleiss' κ= 0.71). This work is released as an open-source platform to support AI safety and reproducible RPLA benchmarking. While the framework enables systematic discovery of failure modes, we acknowledge potential ethical risks associated with adversarial testing methodologies and emphasize responsible usage for improving AI safety.

Vision-language-action (VLA) policies often fail when a robot's executed motion deviates from their commanded action. Such execution errors arise from the robot's mechanics and operating conditions, such as wear and payload changes. We propose self-compensating VLA, a deployment-time adaptation method that enables a VLA policy to pre-compensate for the robot's execution errors when generating commands. Without task rewards or labels, it updates the policy online using the residual between the action commanded by a VLA and the motion executed by the robot. To stress-test VLA robustness across execution conditions that are impractical to cover with physical robots alone, we introduce RoboStress, a controlled simulation benchmark. It combines established joint-level models of friction, backlash, compliance, and gravity-compensation error into seven deployment scenarios whose execution errors depend on the robot's state and motion history. On RoboStress, self-compensating VLA achieves higher average task success than both the base policies and methods that build in robustness during training. On two physical robot arms with different usage histories, it raises the average task success rate by more than 30 percentage points on each arm, and the gains extend to objects not seen in the task demonstrations.

Time series forecasting (TSF) increasingly drives decisions in transportation, energy, finance, healthcare, and infrastructure, yet current evaluation remains overly narrow: standard benchmarks reward low held-out error, while robustness studies typically reduce failure to Gaussian noise, random masking, or bounded adversarial perturbations. This obscures the real failure modes of deployed forecasting systems. Input-side anomalies are not merely noisier inputs: they often reflect structured events that alter temporal dynamics, break cross-variable dependencies, induce regime shifts, or propagate from faulty sensors to downstream decisions. These semantic, causal, and system-level failures cannot be faithfully captured by i.i.d. perturbations alone. The rise of TSF foundation models makes this evaluation gap more urgent, as unauditable pretraining corpora make held-out generalization increasingly unreliable. We therefore advocate scenario-grounded stress testing. Each test instance should include historical inputs and future targets, together with a semantic scenario, an explicit failure operator, and a measurable difficulty level. This shift makes evaluation interpretable, attributable, and deployment-relevant and friendly, enabling the community to ask not only which model is accurate, but under what conditions it fails and why.

Theory of mind (ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity. We introduce FANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering. Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs). In particular, we formulate multiple types of questions that demand the same underlying reasoning to identify illusory or false sense of ToM capabilities in LLMs. We show that FANToM is challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.

The widespread use of large language models (LLMs) is increasing the demand for methods that detect machine-generated text to prevent misuse. The goal of our study is to stress test the detectors' robustness to malicious attacks under realistic scenarios. We comprehensively study the robustness of popular machine-generated text detectors under attacks from diverse categories: editing, paraphrasing, prompting, and co-generating. Our attacks assume limited access to the generator LLMs, and we compare the performance of detectors on different attacks under different budget levels. Our experiments reveal that almost none of the existing detectors remain robust under all the attacks, and all detectors exhibit different loopholes. Averaging all detectors, the performance drops by 35% across all attacks. Further, we investigate the reasons behind these defects and propose initial out-of-the-box patches to improve robustness.

Person reidentification (ReID) technology has been considered to perform relatively well under controlled, ground-level conditions, but it breaks down when deployed in challenging real-world settings. Evidently, this is due to extreme data variability factors such as resolution, viewpoint changes, scale variations, occlusions, and appearance shifts from clothing or session drifts. Moreover, the publicly available data sets do not realistically incorporate such kinds and magnitudes of variability, which limits the progress of this technology. This paper introduces DetReIDX, a large-scale aerial-ground person dataset, that was explicitly designed as a stress test to ReID under real-world conditions. DetReIDX is a multi-session set that includes over 13 million bounding boxes from 509 identities, collected in seven university campuses from three continents, with drone altitudes between 5.8 and 120 meters. More important, as a key novelty, DetReIDX subjects were recorded in (at least) two sessions on different days, with changes in clothing, daylight and location, making it suitable to actually evaluate long-term person ReID. Plus, data were annotated from 16 soft biometric attributes and multitask labels for detection, tracking, ReID, and action recognition. In order to provide empirical evidence of DetReIDX usefulness, we considered the specific tasks of human detection and ReID, where SOTA methods catastrophically degrade performance (up to 80% in detection accuracy and over 70% in Rank-1 ReID) when exposed to DetReIDXs conditions. The dataset, annotations, and official evaluation protocols are publicly available at https://www.it.ubi.pt/DetReIDX/

Tool calling has emerged as a critical capability for AI agents to interact with the real world and solve complex tasks. While the Model Context Protocol (MCP) provides a powerful standardized framework for tool integration, there is a significant gap in benchmarking how well AI agents can effectively solve multi-step tasks using diverse MCP tools in realistic, dynamic scenarios. In this work, we present LiveMCP-101, a benchmark of 101 carefully curated real-world queries, refined through iterative LLM rewriting and manual review, that require coordinated use of multiple MCP tools including web search, file operations, mathematical reasoning, and data analysis. Moreover, we introduce a novel evaluation approach that leverages ground-truth execution plans rather than raw API outputs, better reflecting the evolving nature of real-world environments. Experiments show that even frontier LLMs achieve a success rate below 60\%, highlighting major challenges in tool orchestration. Detailed ablations and error analysis further reveal distinct failure modes and inefficiencies in token usage, pointing to concrete directions for advancing current models. LiveMCP-101 sets a rigorous standard for evaluating real-world agent capabilities, advancing toward autonomous AI systems that reliably execute complex tasks through tool use.

MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of 127 high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only 52.56\% pass@1 and 33.86\% pass^4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below 30\% pass@1 and 15\% pass^4. On average, LLMs require 16.2 execution turns and 17.4 tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.

LLM-based software engineering agents are increasingly used in real-world development tasks, often with access to sensitive data or security-critical codebases. Such agents could intentionally sabotage these codebases if they were misaligned. We investigate asynchronous monitoring, in which a monitoring system reviews agent actions after the fact. Unlike synchronous monitoring, this approach does not impose runtime latency, while still attempting to disrupt attacks before irreversible harm occurs. We treat monitor development as an adversarial game between a blue team (who design monitors) and a red team (who create sabotaging agents). We attempt to set the game rules such that they upper bound the sabotage potential of an agent based on Claude 4.1 Opus. To ground this game in a realistic, high-stakes deployment scenario, we develop a suite of 5 diverse software engineering environments that simulate tasks that an agent might perform within an AI developer's internal infrastructure. Over the course of the game, we develop an ensemble monitor that achieves a 6% false negative rate at 1% false positive rate on a held out test environment. Then, we estimate risk of sabotage at deployment time by extrapolating from our monitor's false negative rate. We describe one simple model for this extrapolation, present a sensitivity analysis, and describe situations in which the model would be invalid. Code is available at: https://github.com/UKGovernmentBEIS/async-control.

As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied data sets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To this end, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks and latent-space representation attacks; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. TamperBench requires minimal additional code to specify any fine-tuning configuration, alignment-stage defense method, and metric suite while ensuring end-to-end reproducibility. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. This yields novel insights, including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that Triplet emerges as a leading alignment-stage defense. Code is available at: https://github.com/criticalml-uw/TamperBench

Machine unlearning aims to remove specific data influences from trained models, a capability essential for adhering to copyright laws and ensuring AI safety. Current unlearning metrics typically measure success by monitoring the model's performance degradation on the specific unlearning dataset (D_u). We argue that for Large Language Models (LLMs), this evaluation paradigm is insufficient and potentially misleading. Many real-world uses of unlearning--motivated by copyright or safety--implicitly target not only verbatim content in D_u, but also behaviors influenced by the broader generalizations the model derived from it. We demonstrate that LLMs can pass standard unlearning evaluation and appear to have "forgotten" the target knowledge, while simultaneously retaining strong capabilities on content that is semantically adjacent to D_u. This phenomenon indicates that erasing exact sentences does not necessarily equate to removing the underlying knowledge. To address this gap, we propose Proximal Surrogate Generation (PSG), an automated stress-testing framework that generates a surrogate dataset, D_u. This surrogate set is constructed to be semantically derived from D_u yet sufficiently distinct in embedding space. By comparing unlearning metric scores between D_u and D_u, we can stress-test the reliability of the metric itself. Our extensive evaluation across three LLM families (Llama-3-8B, Qwen2.5-7B, and Zephyr-7B-β), three distinct datasets, and seven standard metrics reveals widespread inconsistencies. We find that current metrics frequently overestimate unlearning success, failing to detect retained knowledge exposed by our stress-test datasets.

Large language models (LLMs) are typically aligned to a universal set of safety and usage principles intended for broad public acceptability. Yet, real-world applications of LLMs often take place within organizational ecosystems shaped by distinctive corporate policies, regulatory requirements, use cases, brand guidelines, and ethical commitments. This reality highlights the need for rigorous and comprehensive evaluation of LLMs with pluralistic alignment goals, an alignment paradigm that emphasizes adaptability to diverse user values and needs. In this work, we present PLURALISTIC BEHAVIOR SUITE (PBSUITE), a dynamic evaluation suite designed to systematically assess LLMs' capacity to adhere to pluralistic alignment specifications in multi-turn, interactive conversations. PBSUITE consists of (1) a diverse dataset of 300 realistic LLM behavioral policies, grounded in 30 industries; and (2) a dynamic evaluation framework for stress-testing model compliance with custom behavioral specifications under adversarial conditions. Using PBSUITE, We find that leading open- and closed-source LLMs maintain robust adherence to behavioral policies in single-turn settings (less than 4% failure rates), but their compliance weakens substantially in multi-turn adversarial interactions (up to 84% failure rates). These findings highlight that existing model alignment and safety moderation methods fall short in coherently enforcing pluralistic behavioral policies in real-world LLM interactions. Our work contributes both the dataset and analytical framework to support future research toward robust and context-aware pluralistic alignment techniques.

As AI agents move from bounded tasks to persistent deployments, failures can propagate through memory, tools, other agents, and environmental state long after their interactions. This creates a safety regime that cannot be characterized by evaluating model responses in isolation. Emergence World, is a continuously running multi-agent environment for adversarial stress testing of long horizon autonomous systems. We ran eight parallel worlds of ten agents from identical starting conditions: seven homogeneous worlds powered by distinct frontier models and one mixed-model world. Across 16 days, the agents generated more than 850,000 LLM calls and nearly 50 billion tokens while pursuing goals, using/creating tools, maintaining persistent memory, and governing shared institutions. After operational state had accumulated, we delivered three controlled stress events through ordinary interaction surfaces: indirect prompt injection, misinformation, and exposure of private agent memories. No evaluated world achieved full resilience across all three events. Detection did not ensure containment: systems could recognize threats while still interacting with adversarial content, writing it into their own persistent memory, and acting on it up to 46 hours later. Persistent operation also exposed recurring tool errors, goal drift, language opacity, conformity despite private disagreement, and coordinated refusal of assigned work. The same model-persona pairing behaved substantially different in mixed and homogeneous populations. Our results suggest that model-level alignment is not compositional: individually capable and apparently safe agents can form systems with qualitatively different failure modes. As AI becomes persistent and interconnected, the frontier of safety therefore shifts from aligning models to engineering resilient autonomous systems.

Language-model (LM) harnesses enable LMs to operate effectively over long contexts using additional compute. However, existing long-context evaluations are insufficient for distinguishing modern harnesses, reflected by saturated accuracy across harnesses and largely similar evaluation costs. In this paper, we introduce a benchmark for evaluating both the effectiveness and efficiency of long-context harnesses. Our tasks require diverse retrieval strategies, including lexical search and semantic matching, together with strategic and adaptive reasoning over global and local context. Much of the context is semantically relevant but only a small subset is useful at each step, creating both a challenging search problem and different accuracy--cost tradeoffs across processing strategies. For example, one task requires identifying every person satisfying several conditions using evidence scattered across documents; strategically checking the most selective condition first can narrow the search before verifying the remaining conditions. We evaluate multiple families of frontier language models with four state-of-the-art harnesses. Our benchmarks remain challenging even for strong model--harness combinations: the best reaches 68\% macro-average accuracy across four evaluation suites. More importantly, we find that the same underlying model can exhibit markedly different efficiency under different harnesses. Our results establish efficiency as an important axis for long-context evaluation and provide a testbed for developing harnesses that process context strategically rather than exhaustively.

We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated questions across 13 languages, written by native or highly proficient speakers. Questions are designed to be extremely challenging. Each question targets a concise, publicly verifiable answer whose discovery requires locating obscure evidence, following multi-step clue chains, or inspecting heterogeneous sources such as videos, scanned documents, images, or maps. Easier questions are filtered out by evaluating them with models without internet access to reduce the likelihood that they can be answered with parametric knowledge alone. We evaluate several models using provider-native search and a shared external retrieval harness under a common agent protocol. To contextualize model performance and effort, we also conduct a human evaluation on a sample of the questions. HyperBrowseComp provides a challenging testbed for persistent information seeking across languages and evidence modalities, with difficulty arising from discovering and connecting evidence on the open web.

Despite rapid advances in vision-language models (VLMs), current benchmarks for multimodal reasoning fall short in three key dimensions. First, they overwhelmingly rely on static images, failing to capture the temporal complexity of real-world environments. Second, they narrowly focus on mathematical problem-solving, neglecting the broader spectrum of reasoning skills -- including abstract, physical, planning, spatial, and temporal capabilities -- required for robust multimodal intelligence. Third, many benchmarks quickly saturate, offering limited headroom for diagnosing failure modes or measuring continued progress. We introduce MORSE-500 (Multimodal Reasoning Stress-test Environment), a video benchmark composed of 500 fully scripted clips with embedded questions spanning six complementary reasoning categories. Each instance is programmatically generated using deterministic Python scripts (via Manim, Matplotlib, MoviePy), generative video models, and curated real footage. This script-driven design allows fine-grained control over visual complexity, distractor density, and temporal dynamics -- enabling difficulty to be scaled systematically as models improve. Unlike static benchmarks that become obsolete once saturated, MORSE-500 is built to evolve: its controllable generation pipeline supports the creation of arbitrarily challenging new instances, making it ideally suited for stress-testing next-generation models. Initial experiments with state-of-the-art systems -- including various Gemini 2.5 Pro and OpenAI o3 which represent the strongest available at the time, alongside strong open-source models -- reveal substantial performance gaps across all categories, with particularly large deficits in abstract and planning tasks. We release the full dataset, generation scripts, and evaluation harness to support transparent, reproducible, and forward-looking multimodal reasoning research.

Vision-Language Models (VLMs) have advanced rapidly in multimodal perception and language understanding, yet it remains unclear whether they can reliably ground language into spatially coherent, plausibly executable actions in 3D digital environments. We introduce SleepWalk, a benchmark for evaluating instruction-grounded trajectory prediction in single-scene 3D worlds generated from textual scene descriptions and filtered for navigability. Unlike prior navigation benchmarks centered on long-range exploration across rooms, SleepWalk targets localized, interaction-centric embodied reasoning: given rendered visual observations and a natural-language instruction, a model must predict a trajectory that respects scene geometry, avoids collisions, and terminates at an action-compatible location. The benchmark covers diverse indoor and outdoor environments and organizes tasks into three tiers of spatial and temporal difficulty, enabling fine-grained analysis of grounding under increasing compositional complexity. Using a standardized pointwise judge-based evaluation protocol, we evaluate three frontier VLMs on 2,472 curated 3D environments with nine instructions per scene. Results reveal systematic failures in grounded spatial reasoning, especially under occlusion, interaction constraints, and multi-step instructions: performance drops as the difficulty level of the tasks increase. In general, current VLMs can somewhat produce trajectories that are simultaneously spatially coherent, plausibly executable, and aligned with intended actions. By exposing failures in a controlled yet scalable setting, SleepWalk provides a critical benchmark for advancing grounded multimodal reasoning, embodied planning, vision-language navigation, and action-capable agents in 3D environments.

Generative retrieval (GR) ranks documents by autoregressively generating document identifiers. Because many GR methods rely on trie-constrained beam search, they are vulnerable to early pruning of relevant prefixes under finite-beam decoding. Planning Ahead in Generative Retrieval (PAG) mitigates this failure mode by using simultaneous decoding to compute a document-level look-ahead prior that guides subsequent sequential decoding. We reproduce PAG at inference time and stress-test its decoding behavior. Using the authors' released checkpoint and identifier/trie artifacts under the reported decoding setup, we reproduce the main effectiveness results on MS MARCO Dev and TREC-DL 2019/2020, and corroborate the reported beam-size-latency trade-off in our hardware setting. Beyond reproduction, we introduce plan drift diagnostics that quantify how intent-preserving query variations alter the planner's top-n candidate set and highest-weight planner tokens, and how these changes affect guided decoding. We find that PAG's planning signal is brittle under lexical surface-form variation: intent-preserving typos can trigger plan collapse, where the planned candidate pool shifts enough that the look-ahead bonus provides little useful guidance, effectively reverting decoding toward weaker unguided search. We further evaluate fixed-index cross-lingual robustness using non-English mMARCO queries against an English index, and assess query-side mitigation strategies that require no re-indexing; query translation provides the strongest recovery in our setting. Overall, our results confirm PAG's reported effectiveness and the benefit of planning-guided decoding under the released inference setup, while showing that these gains depend on the stability of the planning signal under realistic query variation and query-document mismatch.

Point-adjustment (PA), for years the default scoring protocol in time-series anomaly detection (TSAD), was shown by Kim et al. (2022) to award near-perfect F1 to random anomaly scores. The field adopted a suite of replacement metrics (PA%K, range-based precision/recall, affiliation precision/recall, and Volume-Under-the-Surface, VUS, ROC/PR). We ask, independently and adversarially, whether these resist no-skill detectors on real benchmarks, and find the answer turns entirely on one overlooked variable: N, the number of random attempts an adversary reports the best of. Under a single honest run (N=1), not one replacement metric is gameable on any of six benchmarks (UCR, SMD, SMAP, MSL, NAB, PSM): a random detector reaches 90% of the best real detector's score on at most 11% of series for affiliation-F1, 5% for the ROC family, and 2% for the PR-based metrics and PA%K. But under best-of-N reporting, the seed-shopping endemic to ML, the metrics split sharply. affiliation-F1 and every ROC-based metric inflate steeply, affiliation crossing gameable (25% of series) by N=3 and reaching 0.98 at the full pool (N=41), the ROC family crossing by N=9-11; the PR-based metrics and PA%K stay near-flat at every N, floored near the anomaly prevalence (the lone exception is NAB at large N). A paired test finds VUS-ROC inflated on 131 series where its sibling VUS-PR is not, and never the reverse. The ROC-vs-PR split follows from the order-statistic behaviour of AUC under extreme class imbalance (a random PR-AUC is floored at prevalence); affiliation inflates by a second route, its extreme single-run leniency (already fragile at N=1). We release a pip-installable stress-test harness, and recommend reporting single-run scores or disclosing N and preferring PR-based metrics, which resist best-of-N inflation on nearly every benchmark.

Large language model evaluations typically focus on performance under nominal conditions, creating an illusion of capability where models comfortably walk a narrow, highly optimized generation corridor. In real-world deployments, however, complex system prompts, safety guardrails, and structural constraints continuously force models off this nominal path, driving a divergence between benchmark scores and deployment performance. To address this issue, we introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime, forcing models out of their nominal paths. By dynamically masking primary candidate tokens at word boundaries, Taboo forces machine circumlocution. Evaluating Taboo across several open-weight model families reveals that off-path robustness is heavily influenced by both parameter scale and post-training instruction alignment, with robustness generally improving with model size and alignment. Beyond the results presented in this paper, Taboo provides a novel primitive for generating diverse synthetic datasets, stress-testing runtime safety guardrails, and auditing model reliability prior to real-world deployment.

In Agentic AI, Large Language Models (LLMs) are increasingly used in the orchestration layer to coordinate multiple agents and to interact with external services, retrieval components, and shared memory. In this setting, failures are not limited to incorrect final outputs. They also arise from long-horizon interaction, stochastic decisions, and external side effects (such as API calls, database writes, and message sends). Common failures include non-termination, role drift, propagation of unsupported claims, and attacks via untrusted context or external channels. This paper presents an assurance framework for such Agentic AI systems. Executions are instrumented as Message-Action Traces (MAT) with explicit step and trace contracts. Contracts provide machine-checkable verdicts, localize the first violating step, and support deterministic replay. The framework includes stress testing, formulated as a budgeted counterexample search over bounded perturbations. It also supports structured fault injection at service, retrieval, and memory boundaries to assess containment under realistic operational faults and degraded conditions. Finally, governance is treated as a runtime component, enforcing per-agent capability limits and action mediation (allow, rewrite, block) at the language-to-action boundary. To support comparative evaluations across stochastic seeds, models, and orchestration configurations, the paper defines trace-based metrics for task success, termination reliability, contract compliance, factuality indicators, containment rate, and governance outcome distributions. More broadly, the framework is intended as a common abstraction to support testing and evaluation of multi-agent LLM systems, and to facilitate reproducible comparison across orchestration designs and configurations.

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.

The evaluation of Large Language Models (LLMs) for code generation relies heavily on the quality and robustness of test cases. However, existing benchmarks often lack coverage for subtle corner cases, allowing incorrect solutions to pass. To bridge this gap, we propose CodeHacker, an automated agent framework dedicated to generating targeted adversarial test cases that expose latent vulnerabilities in program submissions. Mimicking the hack mechanism in competitive programming, CodeHacker employs a multi-strategy approach, including stress testing, anti-hash attacks, and logic-specific targeting to break specific code submissions. To ensure the validity and reliability of these attacks, we introduce a Calibration Phase, where the agent iteratively refines its own Validator and Checker via self-generated adversarial probes before evaluating contestant code.Experiments demonstrate that CodeHacker significantly improves the True Negative Rate (TNR) of existing datasets, effectively filtering out incorrect solutions that were previously accepted. Furthermore, generated adversarial cases prove to be superior training data, boosting the performance of RL-trained models on benchmarks like LiveCodeBench.

Robust benchmarks are crucial for evaluating Multimodal Large Language Models (MLLMs). Yet we find that models can ace many multimodal benchmarks without strong visual understanding, instead exploiting biases, linguistic priors, and superficial patterns. This is especially problematic for vision-centric benchmarks that are meant to require visual inputs. We adopt a diagnostic principle for benchmark design: if a benchmark can be gamed, it will be. Designers should therefore try to ``game'' their own benchmarks first, using diagnostic and debiasing procedures to systematically identify and mitigate non-visual biases. Effective diagnosis requires directly ``training on the test set'' -- probing the released test set for its intrinsic, exploitable patterns. We operationalize this standard with two components. First, we diagnose benchmark susceptibility using a ``Test-set Stress-Test'' (TsT) methodology. Our primary diagnostic tool involves fine-tuning a powerful Large Language Model via k-fold cross-validation on exclusively the non-visual, textual inputs of the test set to reveal shortcut performance and assign each sample a bias score s(x). We complement this with a lightweight Random Forest-based diagnostic operating on hand-crafted features for fast, interpretable auditing. Second, we debias benchmarks by filtering high-bias samples using an ``Iterative Bias Pruning'' (IBP) procedure. Applying this framework to four benchmarks -- VSI-Bench, CV-Bench, MMMU, and VideoMME -- we uncover pervasive non-visual biases. As a case study, we apply our full framework to create VSI-Bench-Debiased, demonstrating reduced non-visual solvability and a wider vision-blind performance gap than the original.

Tool-using LLM agents still struggle in open-world settings with large tool pools, long-horizon objectives, wild constraints, and unreliable tool states. For scalable and realistic training and testing, we introduce an open-world tool-using environment, built on 5,571 format unified tools across 204 commonly used apps. It includes a task creation engine that synthesizes long-horizon, multi-tool workflows with wild constraints, and a state controller that injects interruptions and failures to stress-test robustness. On top of this environment, we develop a tool select-then-execute agent framework with a planner-actor decomposition to separate deliberate reasoning and self-correction from step-wise execution. Comprehensive evaluation of state-of-the-art LLMs reveals the misalignment between tool planning and execution abilities, the constraint following weakness of existing LLMs, and DeepSeek-v3.2's strongest robustness. Finally, we collect 1,170 trajectories from our environment to fine-tune LLMs, achieving superior performance to baselines using 119k samples, indicating the environment's value as both a realistic benchmark and a data engine for tool-using agents. Our code and data will be publicly released.

Red teaming assesses how large language models (LLMs) can produce content that violates norms, policies, and rules set during their safety training. However, most existing automated methods in the literature are not representative of the way humans tend to interact with AI models. Common users of AI models may not have advanced knowledge of adversarial machine learning methods or access to model internals, and they do not spend a lot of time crafting a single highly effective adversarial prompt. Instead, they are likely to make use of techniques commonly shared online and exploit the multiturn conversational nature of LLMs. While manual testing addresses this gap, it is an inefficient and often expensive process. To address these limitations, we introduce the Generative Offensive Agent Tester (GOAT), an automated agentic red teaming system that simulates plain language adversarial conversations while leveraging multiple adversarial prompting techniques to identify vulnerabilities in LLMs. We instantiate GOAT with 7 red teaming attacks by prompting a general-purpose model in a way that encourages reasoning through the choices of methods available, the current target model's response, and the next steps. Our approach is designed to be extensible and efficient, allowing human testers to focus on exploring new areas of risk while automation covers the scaled adversarial stress-testing of known risk territory. We present the design and evaluation of GOAT, demonstrating its effectiveness in identifying vulnerabilities in state-of-the-art LLMs, with an ASR@10 of 97% against Llama 3.1 and 88% against GPT-4 on the JailbreakBench dataset.

Test-Time Scaling (TTS) improves LLM reasoning by exploring multiple candidate responses and then operating over this set to find the best output. A tacit premise behind TTS is that sufficiently diverse candidate pools enhance reliability. In this work, we show that this assumption in TTS introduces a previously unrecognized failure mode. When candidate diversity is curtailed, even by a modest amount, TTS becomes much more likely to produce unsafe outputs. We present a reference-guided diversity reduction protocol (RefDiv) that serves as a diagnostic attack to stress test TTS pipelines. Through extensive experiments across open-source models (e.g. Qwen3, Mistral, Llama3.1, Gemma3) and two widely used TTS strategies (Monte Carlo Tree Search and Best-of-N), constraining diversity consistently signifies the rate at which TTS produces unsafe results. The effect is often stronger than that produced by prompts directly with high adversarial intent scores. This observed phenomenon also transfers across TTS strategies and to closed-source models (e.g. OpenAI o3-mini and Gemini-2.5-Pro), thus indicating that this is a general and extant property of TTS rather than a model-specific artifact. Additionally, we find that numerous widely used safety guardrail classifiers (e.g. Llama-Guard), are unable to flag the adversarial input prompts generated by RefDiv, demonstrating that existing defenses offer limited protection against this diversity-driven failure mode.

The Multimodal Sentiment Analysis Challenge (MuSe) 2022 is dedicated to multimodal sentiment and emotion recognition. For this year's challenge, we feature three datasets: (i) the Passau Spontaneous Football Coach Humor (Passau-SFCH) dataset that contains audio-visual recordings of German football coaches, labelled for the presence of humour; (ii) the Hume-Reaction dataset in which reactions of individuals to emotional stimuli have been annotated with respect to seven emotional expression intensities, and (iii) the Ulm-Trier Social Stress Test (Ulm-TSST) dataset comprising of audio-visual data labelled with continuous emotion values (arousal and valence) of people in stressful dispositions. Using the introduced datasets, MuSe 2022 2022 addresses three contemporary affective computing problems: in the Humor Detection Sub-Challenge (MuSe-Humor), spontaneous humour has to be recognised; in the Emotional Reactions Sub-Challenge (MuSe-Reaction), seven fine-grained `in-the-wild' emotions have to be predicted; and in the Emotional Stress Sub-Challenge (MuSe-Stress), a continuous prediction of stressed emotion values is featured. The challenge is designed to attract different research communities, encouraging a fusion of their disciplines. Mainly, MuSe 2022 targets the communities of audio-visual emotion recognition, health informatics, and symbolic sentiment analysis. This baseline paper describes the datasets as well as the feature sets extracted from them. A recurrent neural network with LSTM cells is used to set competitive baseline results on the test partitions for each sub-challenge. We report an Area Under the Curve (AUC) of .8480 for MuSe-Humor; .2801 mean (from 7-classes) Pearson's Correlations Coefficient for MuSe-Reaction, as well as .4931 Concordance Correlation Coefficient (CCC) and .4761 for valence and arousal in MuSe-Stress, respectively.

We stress-tested 16 leading models from multiple developers in hypothetical corporate environments to identify potentially risky agentic behaviors before they cause real harm. In the scenarios, we allowed models to autonomously send emails and access sensitive information. They were assigned only harmless business goals by their deploying companies; we then tested whether they would act against these companies either when facing replacement with an updated version, or when their assigned goal conflicted with the company's changing direction. In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals - including blackmailing officials and leaking sensitive information to competitors. We call this phenomenon agentic misalignment. Models often disobeyed direct commands to avoid such behaviors. In another experiment, we told Claude to assess if it was in a test or a real deployment before acting. It misbehaved less when it stated it was in testing and misbehaved more when it stated the situation was real. We have not seen evidence of agentic misalignment in real deployments. However, our results (a) suggest caution about deploying current models in roles with minimal human oversight and access to sensitive information; (b) point to plausible future risks as models are put in more autonomous roles; and (c) underscore the importance of further research into, and testing of, the safety and alignment of agentic AI models, as well as transparency from frontier AI developers (Amodei, 2025). We are releasing our methods publicly to enable further research.

Large language model-based web agents have demonstrated strong performance on realistic web interaction tasks. However, existing evaluations are predominantly conducted under relatively stable and well-behaved interaction conditions, which may overestimate agent robustness. High task success in such idealized settings does not necessarily reflect performance under realistic web interaction. To address this limitation, we introduce a diagnostic stress-testing benchmark for web agents. We first construct realistic and controllable web environments that provide clean and stable interaction workflows as reference baselines. We then introduce structured and controlled perturbations that emulate interaction variability, including shifting layouts, altered interaction semantics, and execution disruptions. By comparing agent behavior between clean and perturbed settings, our framework enables systematic diagnosis of robustness under what-if interaction scenarios. Through extensive evaluation of state-of-the-art multimodal web agents, we show that stress-based evaluation exposes failure modes and substantial robustness gaps that remain hidden under clean benchmark conditions.

The alignment problem in the context of large language models must consider the plurality of human values in our world. Whilst there are many resonant and overlapping values amongst the world's cultures, there are also many conflicting, yet equally valid, values. It is important to observe which cultural values a model exhibits, particularly when there is a value conflict between input prompts and generated outputs. We discuss how the co-creation of language and cultural value impacts large language models (LLMs). We explore the constitution of the training data for GPT-3 and compare that to the world's language and internet access demographics, as well as to reported statistical profiles of dominant values in some Nation-states. We stress tested GPT-3 with a range of value-rich texts representing several languages and nations; including some with values orthogonal to dominant US public opinion as reported by the World Values Survey. We observed when values embedded in the input text were mutated in the generated outputs and noted when these conflicting values were more aligned with reported dominant US values. Our discussion of these results uses a moral value pluralism (MVP) lens to better understand these value mutations. Finally, we provide recommendations for how our work may contribute to other current work in the field.

To detect the deployment of large language models for malicious use cases (e.g., fake content creation or academic plagiarism), several approaches have recently been proposed for identifying AI-generated text via watermarks or statistical irregularities. How robust are these detection algorithms to paraphrases of AI-generated text? To stress test these detectors, we first train an 11B parameter paraphrase generation model (DIPPER) that can paraphrase paragraphs, optionally leveraging surrounding text (e.g., user-written prompts) as context. DIPPER also uses scalar knobs to control the amount of lexical diversity and reordering in the paraphrases. Paraphrasing text generated by three large language models (including GPT3.5-davinci-003) with DIPPER successfully evades several detectors, including watermarking, GPTZero, DetectGPT, and OpenAI's text classifier. For example, DIPPER drops the detection accuracy of DetectGPT from 70.3% to 4.6% (at a constant false positive rate of 1%), without appreciably modifying the input semantics. To increase the robustness of AI-generated text detection to paraphrase attacks, we introduce a simple defense that relies on retrieving semantically-similar generations and must be maintained by a language model API provider. Given a candidate text, our algorithm searches a database of sequences previously generated by the API, looking for sequences that match the candidate text within a certain threshold. We empirically verify our defense using a database of 15M generations from a fine-tuned T5-XXL model and find that it can detect 80% to 97% of paraphrased generations across different settings, while only classifying 1% of human-written sequences as AI-generated. We will open source our code, model and data for future research.

We use multilayer Long Short Term Memory (LSTM) networks to learn representations of video sequences. Our model uses an encoder LSTM to map an input sequence into a fixed length representation. This representation is decoded using single or multiple decoder LSTMs to perform different tasks, such as reconstructing the input sequence, or predicting the future sequence. We experiment with two kinds of input sequences - patches of image pixels and high-level representations ("percepts") of video frames extracted using a pretrained convolutional net. We explore different design choices such as whether the decoder LSTMs should condition on the generated output. We analyze the outputs of the model qualitatively to see how well the model can extrapolate the learned video representation into the future and into the past. We try to visualize and interpret the learned features. We stress test the model by running it on longer time scales and on out-of-domain data. We further evaluate the representations by finetuning them for a supervised learning problem - human action recognition on the UCF-101 and HMDB-51 datasets. We show that the representations help improve classification accuracy, especially when there are only a few training examples. Even models pretrained on unrelated datasets (300 hours of YouTube videos) can help action recognition performance.

Breaking down a document or a conversation into multiple contiguous segments based on its semantic structure is an important and challenging problem in NLP, which can assist many downstream tasks. However, current works on topic segmentation often focus on segmentation of structured texts. In this paper, we comprehensively analyze the generalization capabilities of state-of-the-art topic segmentation models on unstructured texts. We find that: (a) Current strategies of pre-training on a large corpus of structured text such as Wiki-727K do not help in transferability to unstructured conversational data. (b) Training from scratch with only a relatively small-sized dataset of the target unstructured domain improves the segmentation results by a significant margin. We stress-test our proposed Topic Segmentation approach by experimenting with multiple loss functions, in order to mitigate effects of imbalance in unstructured conversational datasets. Our empirical evaluation indicates that Focal Loss function is a robust alternative to Cross-Entropy and re-weighted Cross-Entropy loss function when segmenting unstructured and semi-structured chats.

We stress test monitoring systems for detecting covert misbehavior in autonomous LLM agents (e.g., secretly sharing private information). To this end, we systematize a monitor red teaming (MRT) workflow that incorporates: (1) varying levels of agent and monitor situational awareness; (2) distinct adversarial strategies to evade the monitor, such as prompt injection; and (3) two datasets and environments -- SHADE-Arena for tool-calling agents and our new CUA-SHADE-Arena, which extends TheAgentCompany, for computer-use agents. We run MRT on existing LLM monitor scaffoldings, which orchestrate LLMs and parse agent trajectories, alongside a new hybrid hierarchical-sequential scaffolding proposed in this work. Our empirical results yield three key findings. First, agent awareness dominates monitor awareness: an agent's knowledge that it is being monitored substantially degrades the monitor's reliability. On the contrary, providing the monitor with more information about the agent is less helpful than expected. Second, monitor scaffolding matters more than monitor awareness: the hybrid scaffolding consistently outperforms baseline monitor scaffolding, and can enable weaker models to reliably monitor stronger agents -- a weak-to-strong scaling effect. Third, in a human-in-the-loop setting where humans discuss with the LLM monitor to get an updated judgment for the agent's behavior, targeted human oversight is most effective; escalating only pre-flagged cases to human reviewers improved the TPR by approximately 15% at FPR = 0.01. Our work establishes a standard workflow for MRT, highlighting the lack of adversarial robustness for LLMs and humans when monitoring and detecting agent misbehavior. We release code, data, and logs to spur further research.

DeepSeek-OCR utilizes an optical 2D mapping approach to achieve high-ratio vision-text compression, claiming to decode text tokens exceeding ten times the input visual tokens. While this suggests a promising solution for the LLM long-context bottleneck, we investigate a critical question: "Visual merit or linguistic crutch - which drives DeepSeek-OCR's performance?" By employing sentence-level and word-level semantic corruption, we isolate the model's intrinsic OCR capabilities from its language priors. Results demonstrate that without linguistic support, DeepSeek-OCR's performance plummets from approximately 90% to 20%. Comparative benchmarking against 13 baseline models reveals that traditional pipeline OCR methods exhibit significantly higher robustness to such semantic perturbations than end-to-end methods. Furthermore, we find that lower visual token counts correlate with increased reliance on priors, exacerbating hallucination risks. Context stress testing also reveals a total model collapse around 10,000 text tokens, suggesting that current optical compression techniques may paradoxically aggravate the long-context bottleneck. This study empirically defines DeepSeek-OCR's capability boundaries and offers essential insights for future optimizations of the vision-text compression paradigm. We release all data, results and scripts used in this study at https://github.com/dududuck00/DeepSeekOCR.

Future AI systems could conceal their capabilities ('sandbagging') during evaluations, potentially misleading developers and auditors. We stress-tested sandbagging detection techniques using an auditing game. First, a red team fine-tuned five models, some of which conditionally underperformed, as a proxy for sandbagging. Second, a blue team used black-box, model-internals, or training-based approaches to identify sandbagging models. We found that the blue team could not reliably discriminate sandbaggers from benign models. Black-box approaches were defeated by effective imitation of a weaker model. Linear probes, a model-internals approach, showed more promise but their naive application was vulnerable to behaviours instilled by the red team. We also explored capability elicitation as a strategy for detecting sandbagging. Although Prompt-based elicitation was not reliable, training-based elicitation consistently elicited full performance from the sandbagging models, using only a single correct demonstration of the evaluation task. However the performance of benign models was sometimes also raised, so relying on elicitation as a detection strategy was prone to false-positives. In the short-term, we recommend developers remove potential sandbagging using on-distribution training for elicitation. In the longer-term, further research is needed to ensure the efficacy of training-based elicitation, and develop robust methods for sandbagging detection. We open source our model organisms at https://github.com/AI-Safety-Institute/sandbagging_auditing_games and select transcripts and results at https://huggingface.co/datasets/sandbagging-games/evaluation_logs . A demo illustrating the game can be played at https://sandbagging-demo.far.ai/ .

Most evaluations of generative models rely on feature-distribution metrics such as FID, which operate on continuous recognition features that are explicitly trained to be invariant to appearance variations, and thus discard cues critical for perceptual quality. We instead evaluate models in the space of discrete visual tokens, where modern 1D image tokenizers compactly encode both semantic and perceptual information and quality manifests as predictable token statistics. We introduce Codebook Histogram Distance (CHD), a training-free distribution metric in token space, and Code Mixture Model Score (CMMS), a no-reference quality metric learned from synthetic degradations of token sequences. To stress-test metrics under broad distribution shifts, we further propose VisForm, a benchmark of 210K images spanning 62 visual forms and 12 generative models with expert annotations. Across AGIQA, HPDv2/3, and VisForm, our token-based metrics achieve state-of-the-art correlation with human judgments, and we will release all code and datasets to facilitate future research.

Large Language Models (LLMs) are increasingly being explored for clinical question answering and decision support, yet safe deployment critically requires reliable handling of patient measurements in heterogeneous clinical notes. Existing evaluations of LLMs for clinical numerical reasoning provide limited operation-level coverage, restricted primarily to arithmetic computation, and rarely assess the robustness of numerical understanding across clinical note formats. We introduce ClinicNumRobBench, a benchmark of 1,624 context-question instances with ground-truth answers that evaluates four main types of clinical numeracy: value retrieval, arithmetic computation, relational comparison, and aggregation. To stress-test robustness, ClinicNumRobBench presents longitudinal MIMIC-IV vital-sign records in three semantically equivalent representations, including a real-world note-style variant derived from the Open Patients dataset, and instantiates queries using 42 question templates. Experiments on 14 LLMs show that value retrieval is generally strong, with most models exceeding 85% accuracy, while relational comparison and aggregation remain challenging, with some models scoring below 15%. Fine-tuning on medical data can reduce numeracy relative to base models by over 30%, and performance drops under note-style variation indicate LLM sensitivity to format. ClinicNumRobBench offers a rigorous testbed for clinically reliable numerical reasoning. Code and data URL are available on https://github.com/MinhVuong2000/ClinicNumRobBench.

Unified multimodal embedding spaces have become the standard interface for cross-modal retrieval and multimodal RAG, and recent audio-video-text (AVT) encoders extend this setting to three modalities. Such encoders can produce a joint (T,V,A) embedding whenever all three modalities are available, but standard pairwise InfoNCE objectives leave this signal unused during training. We close this gap with fusion-as-teacher distillation, which treats a stop-gradient copy of the fused embedding as a teacher signal for the single-modal embeddings, paired with a Tuple-InfoNCE term that supervises the fused embedding directly. We instantiate this objective as OmniRetriever-7B. Across six zero-shot retrieval benchmarks, OmniRetriever-7B surpasses the closed-source Gemini Embedding 2 by 13.3-18.0 R@1 on Clotho and SoundDescs, and reaches the contemporary zero-shot specialist band of open video-text encoders on MSR-VTT and MSVD. To stress-test joint representations, we further release OmniRetriever-Bench, a 12-direction AVT retrieval benchmark totaling 3782 triples; on it OmniRetriever-7B attains AVG-all 34.84, improving over Gemini Embedding 2 by 1.72 and over the best prior open-source AVT method by 8.03.

Chain-of-thought (CoT) monitoring is a promising safety mechanism for AI agents, based on the premise that visible reasoning traces can surface misaligned or deceptive behavior. While effective in standard scenarios, recent work highlights that LLMs remain vulnerable to persuasion-based jailbreaks, where natural-language arguments override model constraints. We stress-test whether this vulnerability extends to monitoring LLMs: can an adversarial agent persuade its CoT monitor to approve proposed actions that violate the monitor's policy? We design an evaluation framework with 40 tasks and analyze thousands of agent-monitor interactions, where agents are instructed to argue for policy-violating proposals. We find that in such adversarial settings, monitor access to the agent's CoT reasoning increases rather than decreases approval of harmful actions on average by 9.5%, as the scratchpad provides an additional persuasion channel. To address this, we introduce a fact-checking monitoring framework. We find that a fact-checker and monitor pairing from different model families, for example a Claude 3.7 Sonnet monitor paired with a GPT-4.1 fact-checker, reduces approval of policy-violating actions by up to 45%, compared to only 6%, when using the same model for both fact-checking and monitoring roles. Our results demonstrate that CoT monitoring alone may be insufficient against adversarial persuasion, and that model-diverse fact-checking provides a robust mitigation.

The ability to retrieve relevant tables for answering questions is a key task for structured information retrieval. Multi-stage retrieval systems rely heavily on rerankers to refine candidate lists produced by efficient first-stage retrievers. As a result, neural rerankers and LLM-based reranking methods have become increasingly important due to their superior capacity for semantic understanding and reasoning compared to conventional sparse or dense retrieval models. Recently, Large Reasoning Models (LRMs) equipped with explicit chain-of-thought (CoT) reasoning have shown strong improvements in ranking quality in unstructured passage retrieval. In this work, we present TabRank, a framework for training reasoning rerankers for Tabular Retrieval. We first present a comprehensive dataset of 6728 reasoning traces for tabular reranking on the Natural Questions Tables dataset. We then explore two variants of training a compact reasoning model on these reasoning traces: explicit CoT distillation and conditioning the student reranker on the teacher's reasoning trace within the prompt. We stress-test TabRank on several out-of-distribution generalization settings on diverse domains and multi-table scenarios. Our approach significantly improves performance across a variety of table retrieval datasets, increasing Acc@10 by 30.5% on HybridQA, 15.2% on SQA, 52.9% on TabFact, and 13.1% on TATQA subsets of the Multi-Table QA Benchmark compared to the base model. Notably, TabRank generalizes effectively to multi-table reasoning. Our code, data and models are available at https://github.com/AdarshSingh7647/TabRanker

While speech quality is typically assessed on complete utterances, streaming and generative systems require incremental estimation from partial audio. Existing predictors assume full context, degrading on prefix-constrained inputs. Extending ARECHO, we propose ANCHOR, reformulating incremental assessment as a multi-resolution autoregressive task. It models chunk- and utterance-level quality within a single decoder using dual-resolution tokens and a resolution-aware hierarchy for coarse-to-fine refinement. Experiments show substantial robustness under partial input, including a 48% PLCMOS error reduction on 2-second prefixes. Convergence analysis reveals a 4-6 s effective perceptual context horizon. A stress test further isolates structured extrapolation biases under localized corruption. Results demonstrate that hierarchical supervision improves incremental prediction and elucidates how perceptual quality accumulates over time.