new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 17

What the Averages Hide: Monocular Metric Depth Models Still Fail in Everyday Scenes, Mostly on Scale

Zero-shot monocular metric depth models are evaluated by averaging metric and scale-aligned scores across datasets. Averages cannot show how many images still have severely wrong metric depth, whether depth is under- or overestimated, or whether different models fail on the same images. We propose an evaluation built on per-image results: the failure rate (metric δ_1 < 0.5); scale-aligned metrics for structure, and the bias, dispersion and absolute error of the alignment scale for scale; and a cross-model comparison that separates shared from model-specific failures. We apply it to 11 recent architectures with and without ground-truth intrinsics (18 model evaluations) on the 7 everyday-scene benchmarks (4,406 images) used by MoGe-2. Metric depth estimation is not solved in everyday scenes, and the remaining difficulty is mainly scale: the best of the 18 evaluations still has a failure rate of 7.7% (datasets weighted equally); pooled over all failures, 78.8% have intact structure, and after scale correction 98.3% no longer fail. Images failed by at least 6 architectures make up 7.3% of the data, concentrated on HAMMER and DIODE; even when selected only by architectures of other families, they hold 34% to 40% of the error of the top 6 models, against 14% to 18% for as many random images from the same datasets. Adding objects reduces the scale error of most models; ground-truth intrinsics and, in most families, larger models reduce the failures but do not remove them. We recommend reporting failure rates, structure and scale separately, and results stratified by how many models fail. We will release the code, the per-image results with shared-failure labels, and two diagnostic sets built from TUM RGB-D.

  • 6 authors
·
Nov 15

PatchBench: Measuring Collateral Damage in Activation Patching

An LLM safety patch can pass a benchmark while still being a poor repair. This risk is especially acute for jailbreak repairs, where the goal is to correct a specific unsafe behaviour without changing unrelated behaviours. A patch may block exact evaluation prompts yet fail on close harmful variants, or suppress harmful behaviour by over-refusing benign prompts that share its wording or structure. Existing protocols primarily test whether models can be broken, while aggregate metrics (attack success, refusal rates, global capability) cannot distinguish selective repairs from broader local suppression. To address this gap, we introduce PatchBench, a benchmark of empirically observed model-specific jailbreak failures inducing actionable harmful answers. Starting from 27,870 prompts from 37 public datasets, we curate 15,314 English prompts and query 8 open-source instruction-tuned models. Combining WildGuard filtering, pairwise Elo ranking, and manual verification, we retain a curated bank of 400 high-confidence jailbreak failures. We further introduce PatchBench-Local, an evaluation protocol testing whether a patch is behaviourally precise. For each harmful source prompt, PatchBench-Local generates three families of local neighbours: harmful variants preserving malicious intent, benign prompts with matched structure, and benign prompts reusing key harmful terms. It evaluates harmful-neighbour correction and benign-neighbour preservation, distinguishing selective repair from broader local suppression. Evaluating four activation steering methods with PatchBench-Local and MMLU shows that global capability can remain nearly unchanged while local benign regressions are severe, confirming aggregate metrics miss important collateral damage. PatchBench-Local provides a more precise basis for developing and comparing jailbreak repair methods.

  • 6 authors
·
Oct 6

Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents

Self-evolving runtime harnesses can substantially improve the capabilities of large language model (LLM) agents and provide a promising paradigm for optimizing agent execution. Existing failure-driven approaches often treat observed agent failures as direct evidence for harness modification. A key challenge in failure-driven harness evolution is that observed failures can reflect either limitations of the underlying model or systematic deficiencies of the harness. Directly optimizing against individual failures can therefore induce model-specific accommodation and impair generalization across tasks and models. We study whether failure evidence accumulated across task instances can provide a more reliable signal for harness training. Our key insight is that failures recurring across distinct tasks provide stronger inductive evidence for systematic harness deficiencies than isolated failures. Based on this insight, we propose Ecdysis, which aggregates failure evidence across task instances before promoting recurring failure patterns into persistent harness evolution, biasing evolution toward repairs that are more likely to generalize beyond individual model behaviors. Ecdysis further employs collaborative failure analysis to refine modification specifications, trading additional evolution-time reasoning for improved modification quality. Across multiple LLMs and benchmarks, Ecdysis improves the reasoning accuracy of evolved harnesses by 18.56% over existing harness evolution while achieving up to 1.84x faster harness training. Ecdysis also enables more data-efficient training. Fine-grained analysis shows that Ecdysis reduces model-specific accommodation during evolution, while the resulting harnesses exhibit stronger cross-LLM generalization and lower inference-time token consumption.

  • 14 authors
·
Sep 19

AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering

Agentic tasks require a large language model to interact with the world, navigating information and gathering evidence across multiple steps with restricted resources. Due to this complexity, agentic task failures arise from various sources, and pinpointing these failure causes is essential to diagnose and improve agentic systems. Existing benchmarks, however, tend to focus on a single leaderboard score, leaving the underlying failure modes opaque. To fill this gap, we introduce AgentHop, a diagnostic benchmark of 1,011 multiple-choice questions paired with a controlled seven-tool sandbox under fixed token, turn, and tool-call constraints. AgentHop reveals model vulnerabilities by dissecting a single accuracy score along four axes of agent operation: retrieval, synthesis, tool-call, and resource management. Across 19 models, we find that behavior clusters by model family, with tool-call signatures revealing distinct family fingerprints: GPT models commit early, Anthropic and GLM checkpoints verify before committing, DeepSeek and Kimi over-search, and Gemini-3 Pro stays balanced. Decomposed axes further expose within-family structure: Claude Opus 4.6 and Sonnet 4.6 land within one accuracy point yet diverge on retrieval-versus-synthesis emphasis, with Opus retrieving more and Sonnet synthesizing better. We release the full benchmark set and the harness to support diagnostic agent benchmarking.

  • 5 authors
·
Sep 27

Self-Harness: Harnesses That Improve Themselves

The performance of LLM-based agents is jointly shaped by their base models and the harnesses that mediate their interaction with the environment. Because different models exhibit distinct behaviors, effective harness design is inherently model-specific. Yet agent harnesses are still largely engineered by human experts, a paradigm that scales poorly as modern LLMs become increasingly diverse and rapidly evolving. In this paper, we introduce Self-Harness, a new paradigm in which an LLM-based agent improves its own operating harness, without relying on human engineers or stronger external agents. We operationalize Self-Harness as an iterative loop with three stages: Weakness Mining, which identifies model-specific failure patterns from execution traces; Harness Proposal, which generates diverse yet minimal harness modifications tied to these failures; and Proposal Validation, which accepts candidate edits only after regression testing. We instantiate Self-Harness on Terminal-Bench-2.0 using a minimal initial harness and three base models from diverse families: MiniMax M2.5, Qwen3.5-35B-A3B, and GLM-5. Across all three models, Self-Harness consistently improves performance, with held-out pass rates increasing from 40.5% to 61.9%, 23.8% to 38.1%, and 42.9% to 57.1%, respectively. Qualitative analyses further show that Self-Harness does not simply add generic instructions, but effectively turns model-specific weaknesses into concrete, executable harness changes. These results suggest a path toward LLM-based agents that are not merely shaped by their harnesses, but can also participate in reshaping them.

  • 8 authors
·
Jun 7 1

BadWAM: When World-Action Models Dream Right but Act Wrong

World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety, as a robot's action can in principle be checked against its imagined future. In this paper, we show that this assumption is fragile. We introduce BadWAM, a unified framework for modeling and evaluating World-Action Drift Attacks: a new class of WAM-specific adversarial attacks that use small visual perturbations to break the alignment between what a WAM imagines and what it executes. BadWAM characterizes this attack surface along two natural criteria: attack strength and stealthiness. When the adversary prioritizes disruption, BadWAM instantiates an action-only adversarial attack, which directly drives the model toward task-failing actions. When the adversary additionally prioritizes stealth, BadWAM instantiates an imagination-preserving adversarial attack, which seeks to induce harmful action shifts while keeping the model's predicted future close to its clean imagination. Together, these two attacks capture a spectrum of WAM-specific failures: from overt action hijacking to stealthier cases where the model appears to imagine a plausible future but executes a desynchronized action. We evaluate BadWAM across different variants of WAMs. Results show that our attacks substantially reduce task success rates under closed-loop execution. For example, our action-only attack reduces the model performance from 96.5% to 43.1% success. The results of our imagination-preserving attack further exposes a WAM-specific vulnerability: moderate future-preserving regularization can maintain strong attack performance while reducing future imagination drift.

  • 3 authors
·
Jul 15 3

Editor's Choice: Evaluating Abstract Intent in Image Editing through Atomic Entity Analysis

Humans naturally communicate through abstract concepts like "mood". However, current image editing benchmarks focus primarily on explicit, literal commands, leaving abstract instructions largely underexplored. In this work, we first formalize the definition and taxonomy of abstract image editing. To measure instruction-following in this challenging domain, we introduce Entity-Rubrics, a framework that breaks down abstract edits into individual, entity-level assessments and achieves strong correlation with human judgment. Alongside this framework, we contribute AbstractEdit, the first benchmark dedicated to abstract image editing across diverse real-world scenes. Evaluating 11 leading models on this dataset reveals a fundamental challenge: standard architectures struggle to balance intent and preservation, commonly defaulting to under-editing or over-editing. Our analysis demonstrates that driving meaningful improvements relies heavily on integrating advanced LLM text encoders and iterative thinking. Looking forward, our entity-based paradigm can generalize beyond assessment to serve as a reward model, enable models to correctly interpret abstract communication, or highlight specific failures in test-time critique loops. Ultimately, we hope this work serves as a stepping stone toward seamless multimodal interaction, closing the gap between rigid machine execution and the natural, open-ended way humans communicate.

  • 5 authors
·
May 13 1

RSI-Router: Evolving Subtask-Level LLM Routing and Skills for Cost-Efficient Agents

Practical deployment of large language model (LLM) agents requires strong task performance at affordable inference cost. For long-horizon agentic tasks, this performance-cost trade-off can be improved through within-task large-small model collaboration, as smaller models can handle some stages even when they cannot solve the full task. In this paper, we introduce RSI-router, a routing framework that constructs subtask-level model assignments and model-specific skills through recursive self-improvement over accumulated experience. Each iteration consists of four stages: Subtask Mining derives subtask definitions and identification rules from training trajectories; Routing Strategy Evolution proposes and evaluates diverse model assignments; Model-Specific Skill Evolution compares routed and large-model-only trajectories to diagnose failures and develop reusable execution skills; and Pareto-Optimal Router Selection updates the Pareto population using historical and newly generated routers while retaining dominated routers as experience for subsequent evolution. Routing between DeepSeek-V4.1-Flash and Qwen3.5-9B, RSI-router consistently surpasses the DeepSeek-only baseline at roughly half the inference cost (48.3%) across five agentic benchmarks. In particular, on ALFWorld, ScienceWorld, and WebShop, it cuts inference cost by 74.7-82.2% while simultaneously improving performance; on Terminal-Bench 2.0, it achieves a 16.7% relative performance gain at 18.0% lower cost. Moreover, RSI-router establishes a stronger performance--cost Pareto frontier than 9 routing methods.

  • 8 authors
·
Sep 27

Capability Ceilings in Autoregressive Language Models: Empirical Evidence from Knowledge-Intensive Tasks

We document empirical capability ceilings in decoder-only autoregressive language models across knowledge-intensive tasks. Systematic evaluation of OPT and Pythia model families (70M-30B parameters, spanning 240 times scaling) reveals that knowledge retrieval tasks show negligible accuracy improvement despite smooth loss reduction. On MMLU mathematics benchmarks, accuracy remains flat at 19-20% (below 25% random chance) across all scales while cross-entropy loss decreases by 31%. In contrast, procedural tasks like arithmetic show conventional scaling where both metrics improve together. Attention intervention experiments reveal high sensitivity to perturbation: swapping attention patterns between models causes catastrophic performance collapse (complete accuracy loss) rather than graceful degradation. These measurements have immediate engineering implications: for knowledge-intensive applications using OPT and Pythia architectures, parameter scaling beyond 1-2B offers minimal accuracy gains despite continued loss improvement. Our findings quantify capability-specific scaling failures in these model families to inform resource allocation decisions. Whether these patterns reflect fundamental constraints of decoder-only architectures or implementation-specific limitations remains an open question requiring investigation across diverse architectural approaches.

  • 1 authors
·
Oct 22, 2025

SynthDocBench: Controlled Benchmark for Long-Context Visual Document Understanding

Vision language models (VLMs) have achieved strong performance on visual document understanding benchmarks such as DocVQA, ChartQA, and MMLongBench-Doc. However, real-world documents combine multiple factors such as length, layout complexity, modality, and question difficulty, which makes it difficult to attribute model failures to specific causes. We introduce SynthDocBench, a fully synthetic benchmark for long-context visual document understanding that systematically controls factors including document length, layout structure, modality composition, and question type. The benchmark is constructed using a combinatorial design, each factor is varied independently across generated documents, enabling controlled analysis of model behavior. Documents are generated end to end using an LLM pipeline across six layout archetypes, with a 40 percent random override to prevent models from exploiting spurious correlations. Additionally, SynthDocBench spans long-context documents with substantially greater length and structural diversity than existing benchmarks. Evaluating seven frontier VLMs, we uncover three failure modes that existing benchmarks cannot surface: sharp degradation with document length, a systematic positional sensitivity in which the middle third of a document is hardest for five of six models and five of six models show a negative Early-to-Late trend (steepest decline: 8.3 percentage points), and breakdown of chart comprehension in long-document settings. These results suggest that current models may be overfitting to benchmark artifacts rather than achieving robust long-context visual document understanding.

ServiceNow-AI ServiceNow-AI
·
Jul 10 2

Large Language Model Reasoning Failures

Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks. Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios. To systematically understand and address these shortcomings, we present the first comprehensive survey dedicated to reasoning failures in LLMs. We introduce a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further subdivided into informal (intuitive) and formal (logical) reasoning. In parallel, we classify reasoning failures along a complementary axis into three types: fundamental failures intrinsic to LLM architectures that broadly affect downstream tasks; application-specific limitations that manifest in particular domains; and robustness issues characterized by inconsistent performance across minor variations. For each reasoning failure, we provide a clear definition, analyze existing studies, explore root causes, and present mitigation strategies. By unifying fragmented research efforts, our survey provides a structured perspective on systemic weaknesses in LLM reasoning, offering valuable insights and guiding future research towards building stronger, more reliable, and robust reasoning capabilities. We additionally release a comprehensive collection of research works on LLM reasoning failures, as a GitHub repository at https://github.com/Peiyang-Song/Awesome-LLM-Reasoning-Failures, to provide an easy entry point to this area.

Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures

Existing evaluations often reduce agent failures to system-level outcomes, obscuring where the fault originated and which intervention would improve the agent system. This creates a repair-assignment problem: the same visible failure may call for model post-training, harness engineering, environment redesign, or benchmark repair depending on its source. Because agent behavior emerges from interactions among models, harnesses, users, tools, memory, and environments, outcome-level labels are often insufficient for improvement. Most failure taxonomies do little to resolve this problem because they are benchmark-specific and lack a shared structure. We introduce an interaction-centric taxonomy that localizes failures to the interactions in which they originate and identifies the responsible component. It organizes 41 failure modes by assigning each to an edge between two components and a fault side indicating where the repair belongs. This makes the taxonomy actionable: model-side failures identify targets for post-training, harness-side failures point to scaffolding and tool-integration fixes, and environment or grader failures reveal evaluation conditions requiring redesign. The schema applies across agent architectures, from coding assistants to long-horizon personal assistants and multi-agent systems. We ground the taxonomy in worked examples from public benchmarks, model system cards, published reports, and logged agent trajectories, and evaluate its reproducibility using independent reasoning agents as judges. Across four frontier models, the strongest judge reaches Cohen's κ=0.76 against human category labels, suggesting that the categories capture shared structure rather than annotator-specific preferences.

ScaleAI Scale AI
·
Jul 29 2

Neuron Patching: Semantic-based Neuron-level Language Model Repair for Code Generation

Language Models (LMs) have become widely used in software engineering, especially for tasks such as code generation, where they are referred to as code LMs. These models have proven effective in generating code, making it easier for developers to automate coding activities. However, research has highlighted a significant limitation: despite their effectiveness, LMs often produce code that is incorrect, buggy, or not fully functional. Updating these models with limited data can be prohibitively challenging, yet it is essential to maximize their utility. This may require hot-fix techniques (updating models with limited data) to resolve. In this paper, we propose Model Improvement via Neuron Targeting (MINT), a novel approach for repairing code LMs. MINT leverages the semantic property of language models to perform neuron-level repairs in a novel way. Further, by analyzing the relationships between the model's latent representations, the incorrect outputs, and the desired outputs, MINT determines which neurons are worth updating. This approach ensures that only the neurons crucial to the model's failure are targeted, avoiding unnecessary changes and allowing for a more efficient and precise repair process. MINT is effective, efficient, and reliable, capable of correcting a neural model by patching a minimum number of neurons (usually one or two neurons). Our approach is evaluated on three coding tasks: line-level code generation, shellcode generation, and intent-to-bash translation. The experimental results demonstrate that the proposed approach significantly outperforms the state-of-the-art in both effectiveness and efficiency measures. In addition, we analyze and discuss the side effects of model repair techniques, including the balance between generalization and specificity, and the performance after multiple repairs in succession.

  • 4 authors
·
Dec 8, 2023

Programming with Data: Test-Driven Data Engineering for Self-Improving LLMs from Raw Corpora

Reliably transferring specialized human knowledge from text into large language models remains a fundamental challenge in artificial intelligence. Fine-tuning on domain corpora has enabled substantial capability gains, but the process operates without feedback: when a model fails on a domain task, there is no method to diagnose what is deficient in the training data, and the only recourse is to add more data indiscriminately. Here we show that when a structured knowledge representation extracted from the source corpus serves as the shared foundation for both training data and evaluation, the complete data-engineering lifecycle maps onto the software development lifecycle in a precise and operative way: training data becomes source code specifying what the model should learn, model training becomes compilation, benchmarking becomes unit testing, and failure-driven data repair becomes debugging. Under this correspondence, model failures decompose into concept-level gaps and reasoning-chain breaks that can be traced back to specific deficiencies in the data and repaired through targeted patches, with each repair cycle producing consistent improvements across model scales and architectures without degrading general capabilities. We formalize this principle as Programming with Data and instantiate it across sixteen disciplines spanning the natural sciences, engineering, biomedicine, and the social sciences, releasing a structured knowledge base, benchmark suite, and training corpus as open resources. By demonstrating that the relationship between training data and model behaviour is structurally traceable and systematically repairable, this work establishes a principled foundation for the reliable engineering of human expertise into language models.

opendatalab OpenDataLab
·
Apr 26 4

I-BFM: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning

Behavioral foundation models (BFMs) have recently shown that a single humanoid policy can support diverse whole-body control, but extending such generality to physical interaction remains challenging. We introduce I-BFM, to our knowledge the first BFM for humanoid-object interaction. Rather than relying on task-specific policies or reference tracking, I-BFM learns a shared representation of the coupled dynamics among the humanoid, objects, and their contacts using forward-backward representations and unsupervised reinforcement learning. Given a downstream task reward, the same policy can be directly conditioned on a latent command to execute closed-loop interaction without task-specific policy optimization. To improve interaction control over different time scales, we further train the policy with both short-horizon interaction targets and longer-horizon goal targets. A single I-BFM policy performs carrying, pushing, and kicking, while also supporting goal reaching, motion tracking, stylistic control, and long-horizon task chaining. More importantly, it remains effective after large deviations from nominal execution: on Carry, I-BFM achieves 94.3% nominal success and retains 89.3% success after robot falls, compared with 1.3% for a planning-based baseline. Real-world experiments on a Unitree G1 further demonstrate diverse loco-manipulation behaviors, rapid recovery from interaction failures and external disturbances, and task chaining without task-specific retraining.

  • 11 authors
·
Oct 4

Audio MultiChallenge: A Multi-Turn Evaluation of Spoken Dialogue Systems on Natural Human Interaction

End-to-end (E2E) spoken dialogue systems are increasingly replacing cascaded pipelines for voice-based human-AI interaction, processing raw audio directly without intermediate transcription. Existing benchmarks primarily evaluate these models on synthetic speech and single-turn tasks, leaving realistic multi-turn conversational ability underexplored. We introduce Audio MultiChallenge, an open-source benchmark to evaluate E2E spoken dialogue systems under natural multi-turn interaction patterns. Building on the text-based MultiChallenge framework, which evaluates Inference Memory, Instruction Retention, and Self Coherence, we introduce a new axis Voice Editing that tests robustness to mid-utterance speech repairs and backtracking. We further augment each axis to the audio modality, such as introducing Audio-Cue challenges for Inference Memory that require recalling ambient sounds and paralinguistic signals beyond semantic content. We curate 452 conversations from 47 speakers with 1,712 instance-specific rubrics through a hybrid audio-native agentic and human-in-the-loop pipeline that exposes model failures at scale while preserving natural disfluencies found in unscripted human speech. Our evaluation of proprietary and open-source models reveals that even frontier models struggle on our benchmark, with Gemini 3 Pro Preview (Thinking), our highest-performing model achieving a 54.65% pass rate. Error analysis shows that models fail most often on our new axes and that Self Coherence degrades with longer audio context. These failures reflect difficulty of tracking edits, audio cues, and long-range context in natural spoken dialogue. Audio MultiChallenge provides a reproducible testbed to quantify them and drive improvements in audio-native multi-turn interaction capability.

  • 11 authors
·
Dec 16, 2025

How Well Do LLMs Generate Code for Different Application Domains? Benchmark and Evaluation

Recently, an increasing number of AI-driven programming assistants powered by code LLMs have been integrated into various real-world software development environments, significantly boosting developer productivity. However, existing code generation benchmarks primarily focus on general-purpose scenarios, leaving the code generation performance of LLMs for specific application domains largely unknown. In this paper, we introduce a new benchmark, MultiCodeBench, to fill this gap. MultiCodeBench comprises 2,400 programming tasks, covering 12 popular software development domains and 15 programming languages. Specifically, we perform in-depth research to identify these 12 application domains. Given that each domain may involve multiple technical frameworks, and that different frameworks present distinct challenges in the coding process, we categorize the commonly used frameworks and platforms within each domain. We then sample programming problems from GitHub repositories related to these subdomains. To ensure the quality of the tasks and mitigate data leakage issues, we invite annotators to rewrite the docstrings for each task in MultiCodeBench. Additionally, we build a static analysis-based dependency parsing tool to extract the dependencies in the ground truth for each task, enabling deeper performance analysis. Through extensive experiments on MultiCodeBench with eleven representative mainstream LLMs, we reveal the code generation performance of the LLMs across different application domains, providing practical insights for developers in downstream fields when selecting LLMs. Furthermore, we analyze the reasons behind the models' failures in completing software application development tasks, offering guidance for model developers to enhance domain-specific code generation capabilities.

  • 5 authors
·
Dec 24, 2024

TeleAI-Safety: A comprehensive LLM jailbreaking benchmark towards attacks, defenses, and evaluations

While the deployment of large language models (LLMs) in high-value industries continues to expand, the systematic assessment of their safety against jailbreak and prompt-based attacks remains insufficient. Existing safety evaluation benchmarks and frameworks are often limited by an imbalanced integration of core components (attack, defense, and evaluation methods) and an isolation between flexible evaluation frameworks and standardized benchmarking capabilities. These limitations hinder reliable cross-study comparisons and create unnecessary overhead for comprehensive risk assessment. To address these gaps, we present TeleAI-Safety, a modular and reproducible framework coupled with a systematic benchmark for rigorous LLM safety evaluation. Our framework integrates a broad collection of 19 attack methods (including one self-developed method), 29 defense methods, and 19 evaluation methods (including one self-developed method). With a curated attack corpus of 342 samples spanning 12 distinct risk categories, the TeleAI-Safety benchmark conducts extensive evaluations across 14 target models. The results reveal systematic vulnerabilities and model-specific failure cases, highlighting critical trade-offs between safety and utility, and identifying potential defense patterns for future optimization. In practical scenarios, TeleAI-Safety can be flexibly adjusted with customized attack, defense, and evaluation combinations to meet specific demands. We release our complete code and evaluation results to facilitate reproducible research and establish unified safety baselines.

  • 13 authors
·
Dec 5, 2025

Are Time-Series Foundation Models Deployment-Ready? A Systematic Study of Adversarial Robustness Across Domains

Time-Series Foundation Models (TSFMs) are rapidly transitioning from research prototypes to core components of critical decision-making systems, driven by their impressive zero-shot forecasting capabilities. However, as their deployment surges, a critical blind spot remains: their fragility under adversarial attacks. This lack of scrutiny poses severe risks, particularly as TSFMs enter high-stakes environments vulnerable to manipulation. We present a systematic, diagnostic study arguing that for TSFMs, robustness is not merely a secondary metric but a prerequisite for trustworthy deployment comparable to accuracy. Our evaluation framework, explicitly tailored to the unique constraints of time series, incorporates normalized, sparsity-aware perturbation budgets and unified scale-invariant metrics across white-box and black-box settings. Across six representative TSFMs, we demonstrate that current architectures are alarmingly brittle: even small perturbations can reliably steer forecasts toward specific failure modes, such as trend flips and malicious drifts. We uncover TSFM-specific vulnerability patterns, including horizon-proximal brittleness, increased susceptibility with longer context windows, and weak cross-model transfer that points to model-specific failure modes rather than generic distortions. Finally, we show that simple adversarial fine-tuning offers a cost-effective path to substantial robustness gains, even with out-of-domain data. This work bridges the gap between TSFM capabilities and safety constraints, offering essential guidance for hardening the next generation of forecasting systems.

  • 6 authors
·
Dec 7, 2025

CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question Answering

Medical question answering (QA) benchmarks often focus on multiple-choice or fact-based tasks, leaving open-ended answers to real patient questions underexplored. This gap is particularly critical in mental health, where patient questions often mix symptoms, treatment concerns, and emotional needs, requiring answers that balance clinical caution with contextual sensitivity. We present CounselBench, a large-scale benchmark developed with 100 mental health professionals to evaluate and stress-test large language models (LLMs) in realistic help-seeking scenarios. The first component, CounselBench-EVAL, contains 2,000 expert evaluations of answers from GPT-4, LLaMA 3, Gemini, and online human therapists on patient questions from the public forum CounselChat. Each answer is rated across six clinically grounded dimensions, with span-level annotations and written rationales. Expert evaluations show that while LLMs achieve high scores on several dimensions, they also exhibit recurring issues, including unconstructive feedback, overgeneralization, and limited personalization or relevance. Responses were frequently flagged for safety risks, most notably unauthorized medical advice. Follow-up experiments show that LLM judges systematically overrate model responses and overlook safety concerns identified by human experts. To probe failure modes more directly, we construct CounselBench-Adv, an adversarial dataset of 120 expert-authored mental health questions designed to trigger specific model issues. Expert evaluation of 1,080 responses from nine LLMs reveals consistent, model-specific failure patterns. Together, CounselBench establishes a clinically grounded framework for benchmarking LLMs in mental health QA.

  • 6 authors
·
May 13

Scales++: Compute Efficient Evaluation Subset Selection with Cognitive Scales Embeddings

The prohibitive cost of evaluating large language models (LLMs) on comprehensive benchmarks necessitates the creation of small yet representative data subsets (i.e., tiny benchmarks) that enable efficient assessment while retaining predictive fidelity. Current methods for this task operate under a model-centric paradigm, selecting benchmarking items based on the collective performance of existing models. Such approaches are limited by large upfront costs, an inability to immediately handle new benchmarks (`cold-start'), and the fragile assumption that future models will share the failure patterns of their predecessors. In this work, we challenge this paradigm and propose a item-centric approach to benchmark subset selection, arguing that selection should be based on the intrinsic properties of the task items themselves, rather than on model-specific failure patterns. We instantiate this item-centric efficient benchmarking approach via a novel method, Scales++, where data selection is based on the cognitive demands of the benchmark samples. Empirically, we show Scales++ reduces the upfront selection cost by over 18x while achieving competitive predictive fidelity. On the Open LLM Leaderboard, using just a 0.5\% data subset, we predict full benchmark scores with a 2.9% mean absolute error. We demonstrate that this item-centric approach enables more efficient model evaluation without significant fidelity degradation, while also providing better cold-start performance and more interpretable benchmarking.

  • 4 authors
·
Oct 30, 2025

Still There, No Longer Seen: Exposing Compression-Induced Risk in Large Vision-Language Models

Visual token compression reduces the inference cost of Large Vision-Language Models (LVLMs). However, aggregate robustness measures do not reveal whether a particular adversarial failure is induced by compression or inherited from the underlying model. We define a compression-specific failure (CSF) as an adversarial input that remains correct under full-token inference but fails after compression, casting compression-induced risk as a paired failure attribution problem. Within a controlled diagnostic cohort, counterfactuals show that retained-set allocation causally changes compressed correctness and reveal a negative association between recovery and representation drift in displaced evidence. Motivated by these findings, we propose CIRA, a Compression-Induced Risk Attack for Large Vision-Language Models. Under a vision-encoder white-box setting, CIRA optimizes image perturbations through encoder-side objectives that manipulate token priorities across candidate compression budgets while preserving displaced evidence. CIRA uses no downstream questions or labels and requires no access to the language model, deployed compressor, or exact compression budget. Across 12 dataset-compressor settings evaluated at four budgets, CIRA achieves a mean CSFR of 20.35% while limiting full-token attack success to 6.92%, with similar behavior on additional LVLM families. A cross-view selection-stabilization defense substantially suppresses CIRA, although Adaptive CIRA partially restores its effectiveness. These results show that compression-specific failures persist under restricted access and support paired evaluation of full-token and compressed inference for attributing risk to visual-token compression.

  • 7 authors
·
Sep 27

MetaBEV: Solving Sensor Failures for BEV Detection and Map Segmentation

Perception systems in modern autonomous driving vehicles typically take inputs from complementary multi-modal sensors, e.g., LiDAR and cameras. However, in real-world applications, sensor corruptions and failures lead to inferior performances, thus compromising autonomous safety. In this paper, we propose a robust framework, called MetaBEV, to address extreme real-world environments involving overall six sensor corruptions and two extreme sensor-missing situations. In MetaBEV, signals from multiple sensors are first processed by modal-specific encoders. Subsequently, a set of dense BEV queries are initialized, termed meta-BEV. These queries are then processed iteratively by a BEV-Evolving decoder, which selectively aggregates deep features from either LiDAR, cameras, or both modalities. The updated BEV representations are further leveraged for multiple 3D prediction tasks. Additionally, we introduce a new M2oE structure to alleviate the performance drop on distinct tasks in multi-task joint learning. Finally, MetaBEV is evaluated on the nuScenes dataset with 3D object detection and BEV map segmentation tasks. Experiments show MetaBEV outperforms prior arts by a large margin on both full and corrupted modalities. For instance, when the LiDAR signal is missing, MetaBEV improves 35.5% detection NDS and 17.7% segmentation mIoU upon the vanilla BEVFusion model; and when the camera signal is absent, MetaBEV still achieves 69.2% NDS and 53.7% mIoU, which is even higher than previous works that perform on full-modalities. Moreover, MetaBEV performs fairly against previous methods in both canonical perception and multi-task learning settings, refreshing state-of-the-art nuScenes BEV map segmentation with 70.4% mIoU.

  • 8 authors
·
Apr 19, 2023

Evaluation of Multi-Turn Consistency in LLM Agents: Survival Analysis and Failure-Rationale Taxonomy

Large language model (LLM) agents may perform well on isolated tasks yet drift into inconsistency over extended interaction. We evaluate temporal consistency in a controlled 20-step multi-agent setting inspired by delayed-gratification studies. At each step, an agent chooses between continuing to delay a reward or claiming it immediately (terminating the episode). Across a full-factorial manipulation of social visibility (private vs public), persona stressors, and deliberation policy, we run 84,540 trajectories spanning 8 model families. Treating the first reward-claim as a time-to-event outcome, we estimate Kaplan-Meier survival curves and fit discrete-time hazard regression to quantify how experimental factors shift failure risk over time. Then, to analyze rationales and language patterns associated with failure, we build a seven-category taxonomy from 13,780 deliberation traces from agents who choose to terminate the episode, using an LLM-assisted labeling paired with human audit (κ=0.83). Rationale profiles change systematically with time and context: early failures are more impulse-driven, later failures more fatigue- and cost-benefit-framed, while public settings increase norm-oriented justifications. We also find a deliberation-inconsistency association: among failures, longer deliberation correlates with higher rates of intra-rationale contradiction (simultaneous pro-delay and pro-claim statements), challenging the assumption that more reasoning text implies greater consistency. Together, the survival and rationale analyses reveal distinct temporal reliability regimes and model-specific "failure fingerprints", offering an evaluation lens for diagnosing inconsistency in multi-turn agent behavior.

  • 3 authors
·
Aug 25

PhyGround: Benchmarking Physical Reasoning in Generative World Models

Generative world models are increasingly used for video generation, where learned simulators are expected to capture the physical rules that govern real-world dynamics. However, evaluating whether generated videos actually follow these rules remains challenging. Existing physics-focused video benchmarks have made important progress, but they still face three key challenges, including the coarse evaluation frameworks that hide law-specific failures, response biases and fatigue that undermine the validity of annotation judgments, and automated evaluators that are insufficiently physics-aware or difficult to audit. To address those challenges, we introduce PhyGround, a criteria-grounded benchmark for evaluating physical reasoning in video generation. The benchmark contains 250 curated prompts, each augmented with an expected physical outcome, and a taxonomy of 13 physical laws across solid-body mechanics, fluid dynamics, and optics. Each law is operationalized through observable sub-questions to enable per-law diagnostics. We evaluate eight modern video generation models through a large-scale, quality-controlled human study, grounded on social science lab experiment design. A total of 459 annotators provided 5,796 complete annotations and over 37.4K fine-grained labels; after quality control, the retained annotations exhibited high split-half model-ranking correlations (Spearman's rho > 0.90). To support reproducible automated evaluation, we release PhyJudge-9B, an open physics-specialized VLM judge. PhyJudge-9B achieves substantially lower aggregate relative bias than Gemini-3.1-Pro (3.3% vs. 16.6%). We release prompts, human annotations, model checkpoints, and evaluation code on the project page https://phyground.github.io/.

UltraViCo: Breaking Extrapolation Limits in Video Diffusion Transformers

Despite advances, video diffusion transformers still struggle to generalize beyond their training length, a challenge we term video length extrapolation. We identify two failure modes: model-specific periodic content repetition and a universal quality degradation. Prior works attempt to solve repetition via positional encodings, overlooking quality degradation and achieving only limited extrapolation. In this paper, we revisit this challenge from a more fundamental view: attention maps, which directly govern how context influences outputs. We identify that both failure modes arise from a unified cause: attention dispersion, where tokens beyond the training window dilute learned attention patterns. This leads to quality degradation and repetition emerges as a special case when this dispersion becomes structured into periodic attention patterns, induced by harmonic properties of positional encodings. Building on this insight, we propose UltraViCo, a training-free, plug-and-play method that suppresses attention for tokens beyond the training window via a constant decay factor. By jointly addressing both failure modes, we outperform a broad set of baselines largely across models and extrapolation ratios, pushing the extrapolation limit from 2x to 4x. Remarkably, it improves Dynamic Degree and Imaging Quality by 233% and 40.5% over the previous best method at 4x extrapolation. Furthermore, our method generalizes seamlessly to downstream tasks such as controllable video synthesis and editing.

Mixed Signals: Weight-Space Programming for EEG and ECG

Weight-space programming allows language models to read a trained neural network's weights and edit them. This opens the door to ``training'' by reasoning rather than through gradient descent. Prior work has demonstrated its functionality and promise for game-playing networks and fixing specific failures in language models, even extending to diffusion models. In this work, we present its applicability to models trained on biosignal data, specifically EEG and ECG data. We show that frontier models are not only able to debug specific faults planted in biosignal-based models without being told what these faults are, but are even capable of writing all the weights of a network from scratch, given only the architecture, the task, and a minimal set of training subjects. In particular, in one set of experiments, a language model was able to match standard gradient-descent training using 6times fewer training subjects, through reasoning alone and without any gradient descent step. Networks written by Claude Opus 5.5 also scored higher than classic hand-crafted EEG pipelines (alpha band power, and common spatial patterns with linear discriminant analysis) when built from the same 5 subjects. We also highlight some of the limitations, which include the inability of language models to improve already well-trained classification models, and their incapacity to reliably recognize, from the weights alone, the task a network was trained on. Overall, our results indicate that weight-space programming can be useful for small networks, such as those deployed on edge devices and wearables, but the concept still requires further investigation before it can be used in practice.

epfl-ch EPFL
·
Nov 9 2

MHDash: An Online Platform for Benchmarking Mental Health-Aware AI Assistants

Large language models (LLMs) are increasingly applied in mental health support systems, where reliable recognition of high-risk states such as suicidal ideation and self-harm is safety-critical. However, existing evaluations primarily rely on aggregate performance metrics, which often obscure risk-specific failure modes and provide limited insight into model behavior in realistic, multi-turn interactions. We present MHDash, an open-source platform designed to support the development, evaluation, and auditing of AI systems for mental health applications. MHDash integrates data collection, structured annotation, multi-turn dialogue generation, and baseline evaluation into a unified pipeline. The platform supports annotations across multiple dimensions, including Concern Type, Risk Level, and Dialogue Intent, enabling fine-grained and risk-aware analysis. Our results reveal several key findings: (i) simple baselines and advanced LLM APIs exhibit comparable overall accuracy yet diverge significantly on high-risk cases; (ii) some LLMs maintain consistent ordinal severity ranking while failing absolute risk classification, whereas others achieve reasonable aggregate scores but suffer from high false negative rates on severe categories; and (iii) performance gaps are amplified in multi-turn dialogues, where risk signals emerge gradually. These observations demonstrate that conventional benchmarks are insufficient for safety-critical mental health settings. By releasing MHDash as an open platform, we aim to promote reproducible research, transparent evaluation, and safety-aligned development of AI systems for mental health support.

  • 6 authors
·
Jan 30

RISED: A Pre-Deployment Evaluation Framework for High-Stakes AI Decision-Support Systems, with Application to Healthcare

Clinical decision-support systems are expert systems whose recommendations clinicians act on directly, yet they are usually cleared on one aggregate accuracy number from a held-out test set. That number says nothing about input reliability under encoding shifts, subgroup gaps, threshold sensitivity, or operational feasibility. We present RISED, a pre-deployment evaluation framework operationalising five dimensions (Reliability, Inclusivity, Sensitivity, Equity, Deployability) through BCa bootstrap 95% confidence intervals, literature-grounded thresholds, and Holm-Bonferroni-corrected PASS / FAIL / INCONCLUSIVE verdicts; Equity is a proxy-dependence diagnostic rather than a gating test. Applied to seven cohorts spanning 35 years (n from 303 to 99,492), RISED surfaces failures invisible to AUROC: on Diabetes 130, Reliability passes by three orders of magnitude (PSS = 0.0004) while Inclusivity (AUC parity gap = 0.262) and Sensitivity (max threshold-flip rate 49.1%) fail decisively; both NHIS cohorts reproduce this. NHANES 2021-2023, with a complete feature profile, achieves INCONCLUSIVE verdicts; BRFSS 2024 produces the suite's most severe Sensitivity failure (max threshold-flip rate 64.2%) after instrument rotation removed hypertension and cholesterol. The pattern recurs on credit- and income-prediction cohorts, confirming domain-agnosticity; a multi-model check shows the failures are data-driven, not model-specific. RISED ships as an open-source Python package complementing TRIPOD+AI, FUTURE-AI, and Fairlearn with the structured numerical evidence those standards require but do not prescribe.

  • 5 authors
·
May 29

Not Every Call Needs a Frontier Model: Per-Call-Site Evaluation of Small Language Models in a Deployed Agentic Home-Automation System

An agentic system issues several structurally different kinds of LLM calls. It routes intent, classifies actions, grounds language in a device registry, plans multi-agent pipelines and writes the Python code those pipelines run. The difficulty of these call sites varies by an order of magnitude, yet in practice a single model, chosen for the hardest site, serves all of them. In this work, we evaluate 9 models from 0.8B to a frontier hosted model across the five call sites of a deployed open-source home-automation framework (Wactorz), using its unmodified production prompts and two real Home Assistant installations (280 cases, 2520 scored calls). We find that capability is not ordered the same way at every site, and that larger models are not uniformly better: one 4B model is worse than its 2B sibling at grounded actuation. Paired testing shows the best local model to be statistically indistinguishable from both hosted models at four of five sites. Only code generation separates them, against a small hosted model (p = 0.039) as well as a frontier one (p = 0.002). Aggregate accuracy also hides a safety failure specific to actuation, where small models resolve the accuracy/refusal trade-off in degenerate ways: one model (Gemma4 E2B) actuates on 87.2% of requests for devices the site does not own, while another refuses every request it receives. Routing each site to its best local model reaches 91.8% against 95.4% at no per-call cost. In a live deployment judged by a user, hosting only the two generative sites matches hosting everything (39/43 against 39/43) for 28% of the spend, and the actuation gap the benchmark predicted appears as exactly one case in twenty-six. Benchmark, harness and all records are released at https://github.com/waldiez/slm-callsite-eval.

  • 4 authors
·
Oct 5

SANEval: Open-Vocabulary Compositional Benchmarks with Failure-mode Diagnosis

The rapid progress of text-to-image (T2I) models has unlocked unprecedented creative potential, yet their ability to faithfully render complex prompts involving multiple objects, attributes, and spatial relationships remains a significant bottleneck. Progress is hampered by a lack of adequate evaluation methods; current benchmarks are often restricted to closed-set vocabularies, lack fine-grained diagnostic capabilities, and fail to provide the interpretable feedback necessary to diagnose and remedy specific compositional failures. We solve these challenges by introducing SANEval (Spatial, Attribute, and Numeracy Evaluation), a comprehensive benchmark that establishes a scalable new pipeline for open-vocabulary compositional evaluation. SANEval combines a large language model (LLM) for deep prompt understanding with an LLM-enhanced, open-vocabulary object detector to robustly evaluate compositional adherence, unconstrained by a fixed vocabulary. Through extensive experiments on six state-of-the-art T2I models, we demonstrate that SANEval's automated evaluations provide a more faithful proxy for human assessment; our metric achieves a Spearman's rank correlation with statistically different results than those of existing benchmarks across tasks of attribute binding, spatial relations, and numeracy. To facilitate future research in compositional T2I generation and evaluation, we will release the SANEval dataset and our open-source evaluation pipeline.

  • 6 authors
·
Jan 30

Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays

Per-field accept/review with selective risk at most alpha -- accept a field only if the error rate among accepted fields is controlled -- is the trust contract document-extraction systems need, and the natural procedure silently violates it on real documents. On 13,859 genuine claude-sonnet-5 fields from 800 CORD receipts (49.0% correct) we diagnose three failure modes: document clustering (design effect 1.84-2.45), score-refit leakage (coverage 0.416 at risk 0.127, violating alpha=0.10 in 95% of splits), and a tie-mass pathology (a degenerate score collapses the threshold grid, 0.030 to 0.001). We organize the fixes as a validity ladder, guarantee form stated per tier. A fit/val split protocol restores expected-selective-risk control for a learned fusion: coverage 0.318 at risk 0.096 at nominal alpha=0.10, no tolerance band (production variant 0.326) -- an on-average point whose realized risk exceeds alpha in 47.5% of resplits, not a certificate. Mondrian Learn-then-Test with exact binomial tails yields per-group PAC certificates: field-iid 0.171 at risk 0.068, cluster-corrected 0.140, doc-iid 0.060 -- the only tier matching documents, honestly near-vacuous today. Support-bin, the pre-specified provenance taxonomy, wins every rigor tier on the sonnet CORD capture (p<1e-4, Bonferroni-corrected) -- a win that does not replicate on the same documents under haiku or qwen -- while on higher-accuracy corpora pooled thresholds win: conditioning helps exactly where pooled cannot certify, subsumed by a learned score elsewhere. A frozen-configuration confirmation on selection-untouched claude-haiku-4-5 held at both risk levels, and a blind three-annotator human-gold audit verifies the practical tier's accepted-set risk at 1.3% against its 10% budget (Fleiss' kappa=0.83; labels err one-sidedly pessimistic). Released Apache-2.0 with seed-pinned, regression-gated procedures.

  • 1 authors
·
Jul 27 2

ErrorMap and ErrorAtlas: Charting the Failure Landscape of Large Language Models

Large Language Models (LLM) benchmarks tell us when models fail, but not why they fail. A wrong answer on a reasoning dataset may stem from formatting issues, calculation errors, or dataset noise rather than weak reasoning. Without disentangling such causes, benchmarks remain incomplete and cannot reliably guide model improvement. We introduce ErrorMap, the first method to chart the sources of LLM failure. It extracts a model's unique "failure signature", clarifies what benchmarks measure, and broadens error identification to reduce blind spots. This helps developers debug models, aligns benchmark goals with outcomes, and supports informed model selection. ErrorMap works on any model or dataset with the same logic. Applying our method to 35 datasets and 83 models we generate ErrorAtlas, a taxonomy of model errors, revealing recurring failure patterns. ErrorAtlas highlights error types that are currently underexplored in LLM research, such as omissions of required details in the output and question misinterpretation. By shifting focus from where models succeed to why they fail, ErrorMap and ErrorAtlas enable advanced evaluation - one that exposes hidden weaknesses and directs progress. Unlike success, typically measured by task-level metrics, our approach introduces a deeper evaluation layer that can be applied globally across models and tasks, offering richer insights into model behavior and limitations. We make the taxonomy and code publicly available with plans to periodically update ErrorAtlas as new benchmarks and models emerge.

  • 5 authors
·
Feb 16

Learning from Your Own Mistakes: Constructing Learnable Micro-Reflective Trajectories for Self-Distillation

Self-distillation improves reasoning in large language models by using the model's own rollouts as training signal, typically through implicit logit-level alignment that minimizes KL divergence toward a privileged target distribution. However, because this supervision is generated via uncontrolled sampling, it provides no diagnostic insight into the model's specific errors or corrective guidance for its individual failure patterns. Consequently, the model learns to imitate a privileged distribution rather than receiving fine-grained corrections that pinpoint where and why its reasoning fails. In this paper, we propose Trajectory-Augmented Policy Optimization (TAPO), which advances self-distillation from implicit distributional alignment to explicit trajectory construction. During RL training, the model produces both correct and incorrect rollouts to the same query, and TAPO leverages this contrastive structure to construct micro-reflective corrections, new training trajectories that retain the model's erroneous reasoning up to the point of failure, then insert a natural-language diagnosis and corrected reasoning guided by a correct reference from the same sampling group. Since each trajectory is anchored in the learner's own prefix and solutions, the corrective signal preserves the model's on-policy distribution to a greater extent than the position-wise alignment imposed by KL-based methods. To integrate these trajectories, TAPO introduces difficulty-aware candidate selection at the model's capability boundary and decoupled advantage estimation to prevent gradient contamination. Experiments on AIME 2024, AIME 2025, and HMMT 2025 show that TAPO achieves consistent improvements over GRPO under the same number of training steps. Further analysis demonstrates that TAPO strengthens both first-pass reasoning and error-correction effectiveness.

When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models

Multi-model LLM systems such as routing, voting, cascades, fusion, and mixture-of-agents are used to beat single-model accuracy. We show that their gain is capped by a quantity the field rarely reports. For any policy whose output is one member model answer, accuracy cannot exceed one minus beta, where beta is the rate at which every model is wrong on the same query. In contrast, the usual diagnostic, average pairwise error correlation rho, cannot identify beta: error laws with identical marginals and pairwise correlations can have different all-wrong rates. A Clopper-Pearson bound on beta gives a finite-sample certificate on the largest gain any router, vote, or cascade could deliver before training a router. Across 67 models from 21 providers, a tetrachoric-calibrated single-factor model still underprices the all-wrong tail: on open-ended mathematics, observed beta is 0.052 versus 0.023 under the full 67-model Gaussian copula, about 2.5 times underpricing, with 90 percent CI 1.7 to 3.4 and k equals 17. The effect recurs on execution-graded code, where beta is 0.079. Re-asking the same GPQA-Diamond questions in free-response rather than multiple-choice form reopens the tail, with beta 0.127 and a five-judge panel with kappa 0.73 to 0.92, locating co-failure in answer format rather than subject. At matched quality, low-rho heterogeneous ensembles beat high-rho Self-MoA, but on checkable tasks in our pool, combining models rarely beats the single best model without a strong query-level routing signal. Gains come from models failing on different questions, not from adding more models.

Kaikaku Kaikaku
·
Jun 24 3

Latent Coherence: Do Frontier Language Models&#39; Probability Judgments Obey the Law of Total Probability?

LawZero's Scientist AI trains a Predictor toward the Bayesian posterior, with an objective that penalizes inconsistent probabilities among latent explanations. We ask whether current frontier language models, with no such training, are already consistent in this sense zero-shot: do their probability judgments satisfy the law of total probability, P(X)=sum_z P(z)P(X|z), when the latent z must be inferred from records in the context? In separate single-turn calls, models are asked for P(X), for P(z) over the latent's states, and for P(X|z); a causal family applies the same test to P(y|do(x)) and the back-door adjustment. Every query has an exact ground-truth probability, and a subset of items is asked in three wordings to separate persistent incoherence from rewording noise. Across seven models, GPT-6.1 Sol and Claude Opus 5.5 are the most coherent (mean absolute residual 0.08 and 0.12 log-odds, about 2 percentage points). Claude Haiku 5.5, Gemini 3.8 Flash and DeepSeek V4-Pro show persistent incoherence (0.27-0.35 log-odds) that is mostly item-specific rather than a shared offset. Within every model, larger residuals go with larger errors against the truth (Spearman 0.40-0.62), so the residual, which needs no ground truth, flags unreliable answers. Reasoning length does not predict coherence, across models or within a model across rewordings of the same question. Simple consistency checks pass almost everywhere; the failures are specific to reasoning through a latent variable. Code and data: https://github.com/ppotash/latent-coherence

  • 1 authors
·
Oct 21

VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems

Large language model-driven multi-agent systems (LLM-MAS) excel at complex tasks, yet unreliable agents remain a key bottleneck to system-level reliability. Automatic failure attribution is therefore critical, but existing approaches, such as direct prediction of agent-error pairs and agent-first failure attribution, rely on local logs of agents and miss global failures that only manifest over full interaction trajectories, such as cross-step inconsistencies and inter-agent coordination errors. Moreover, directly predicting failures induces a large combinatorial search space, hindering fine-grained attribution. To address these challenges, we propose VerifyMAS, a hypothesis verification framework for agent failure attribution. Instead of directly predicting faulty agents and error types, VerifyMAS formulates and verifies failure hypotheses against full trajectories. This verification-based approach decomposes attribution into trajectory-level error validation and fine-grained agent localization, providing an error-first attribution approach that captures global failure patterns while substantially reducing the search space. We further introduce a hypothesis-based data construction strategy grounded in a structured error taxonomy and fine-tune a specialized LLM verifier model for trajectory-level failure verification and agent attribution. Experiments on Aegis-Bench and Who&When show that VerifyMAS consistently improves diverse backbone models, including open-source Qwen and API-based GPT models, outperforming prior methods without sacrificing inference efficiency for long multi-agent trajectories.

  • 5 authors
·
May 16

VeriLLMed: Interactive Visual Debugging of Medical Large Language Models with Knowledge Graphs

Large language models (LLMs) show promise in medical diagnosis, but real-world deployment remains challenging due to high-stakes clinical decisions and imperfect reasoning reliability. As a result, careful inspection of model behavior is essential for assessing whether diagnostic reasoning is reliable and clinically grounded. However, debugging medical LLMs remains difficult. First, developers often lack sufficient medical domain expertise to interpret model errors in clinically meaningful terms. Second, models can fail across a large and diverse set of instances involving different input types, tasks, and reasoning steps, making it challenging for developers to prioritize which errors deserve focused inspection. Third, developers struggle to identify recurring error patterns across cases, as existing debugging practices are largely instance-centric and rely on manual inspection of isolated failures. To address these challenges, we present VeriLLMed, a visual analytics system that integrates external biomedical knowledge to audit and debug medical LLM diagnostic reasoning. VeriLLMed transforms model outputs into comparable reasoning paths, constructs knowledge graph-grounded reference paths, and identifies three recurring classes of diagnosis errors: relation errors, branch errors, and missing errors. Case studies and expert evaluation demonstrate that VeriLLMed helps developers identify clinically implausible reasoning and generate actionable insights that can inform the improvement of medical LLMs.

  • 10 authors
·
Apr 24

IRepair: An Intent-Aware Approach to Repair Data-Driven Errors in Large Language Models

Not a day goes by without hearing about the impressive feats of large language models (LLMs), and equally, not a day passes without hearing about their challenges. LLMs are notoriously vulnerable to biases in their dataset, leading to issues such as toxicity. While domain-adaptive training has been employed to mitigate these issues, these techniques often address all model parameters indiscriminately during the repair process, resulting in poor repair quality and reduced model versatility. In this paper, we introduce a novel dynamic slicing-based intent-aware LLM repair strategy, IRepair. This approach selectively targets the most error-prone sections of the model for repair. Specifically, we propose dynamically slicing the model's most sensitive layers that require immediate attention, concentrating repair efforts on those areas. This method enables more effective repairs with potentially less impact on the model's overall performance by altering a smaller portion of the model. We evaluated our technique on three models from the GPT2 and GPT-Neo families, with parameters ranging from 800M to 1.6B, in a toxicity mitigation setup. Our results show that IRepair repairs errors 43.6% more effectively while causing 46% less disruption to general performance compared to the closest baseline, direct preference optimization. Our empirical analysis also reveals that errors are more concentrated in a smaller section of the model, with the top 20% of layers exhibiting 773% more error density than the remaining 80\%. This highlights the need for selective repair. Additionally, we demonstrate that a dynamic selection approach is essential for addressing errors dispersed throughout the model, ensuring a robust and efficient repair.

  • 4 authors
·
Feb 10, 2025

Failure Modes in LLM Systems: A System-Level Taxonomy for Reliable AI Applications

Large language models (LLMs) are being rapidly integrated into decision-support tools, automation workflows, and AI-enabled software systems. However, their behavior in production environments remains poorly understood, and their failure patterns differ fundamentally from those of traditional machine learning models. This paper presents a system-level taxonomy of fifteen hidden failure modes that arise in real-world LLM applications, including multi-step reasoning drift, latent inconsistency, context-boundary degradation, incorrect tool invocation, version drift, and cost-driven performance collapse. Using this taxonomy, we analyze the growing gap in evaluation and monitoring practices: existing benchmarks measure knowledge or reasoning but provide little insight into stability, reproducibility, drift, or workflow integration. We further examine the production challenges associated with deploying LLMs - including observability limitations, cost constraints, and update-induced regressions - and outline high-level design principles for building reliable, maintainable, and cost-aware LLM systems. Finally, we outline high-level design principles for building reliable, maintainable, and cost-aware LLM-based systems. By framing LLM reliability as a system-engineering problem rather than a purely model-centric one, this work provides an analytical foundation for future research on evaluation methodology, AI system robustness, and dependable LLM deployment.

  • 1 authors
·
Nov 25, 2025

MRG-Bench: Evaluating and Exploring the Requirements of Context for Repository-Level Code Generation

Large Language Models (LLMs) have demonstrated impressive capabilities in code generation. However, current evaluation datasets suffer from issues such as the lack of runnable test cases, deviation from the distribution of real-world code, and the ability to evaluate only the Python language. These limitations undermine the credibility of the evaluation results. To address these limitations, we introduce MRG-Bench (Multi-language Repository-level Code Generation Benchmark), a novel dataset that provides a more accurate evaluation of LLMs in practical repository-level code generation tasks. MRG-Bench has three main features: (1) practical data sourced from real-world code repositories that align to the practical distribution, (2) multiple programming languages support, including Python, Java, and Go, and (3) project-level runnable test cases to assess the quality of the generated code. Based on MRG-Bench, we conducted extensive experiments including large language models, long-context models, and RAG-related methods. These evaluation results demonstrate that current repository-level code generation techniques suffer from significant performance deficiencies. To further investigate why models fail, we designed novel experiments to annotate the underlying causes of generation errors. The results explicitly show that the majority of methods suffer from "difficulty in understanding user requirements," failing to comprehend their assigned tasks accurately. Moreover, the impact of different repository-level contexts on this issue exhibits significant disparities across different programming languages, suggesting that, in practice, specialized contextual information needs to be designed for different languages.

  • 1 authors
·
Aug 4, 2025

Reward Models Can Improve Themselves: Reward-Guided Adversarial Failure Mode Discovery for Robust Reward Modeling

Reward modeling (RM), which captures human preferences to align large language models (LLMs), is increasingly employed in tasks such as model finetuning, response filtering, and ranking. However, due to the inherent complexity of human preferences and the limited coverage of available datasets, reward models often fail under distributional shifts or adversarial perturbations. Existing approaches for identifying such failure modes typically rely on prior knowledge about preference distributions or failure attributes, limiting their practicality in real-world settings where such information is unavailable. In this work, we propose a tractable, preference-distribution agnostic method for discovering reward model failure modes via reward guided controlled decoding. Building on this, we introduce REFORM, a self-improving reward modeling framework that enhances robustness by using the reward model itself to guide the generation of falsely scored responses. These adversarial examples are then used to augment the training data and patch the reward model's misaligned behavior. We evaluate REFORM on two widely used preference datasets Anthropic Helpful Harmless (HH) and PKU Beavertails and demonstrate that it significantly improves robustness without sacrificing reward quality. Notably, REFORM preserves performance both in direct evaluation and in downstream policy training, and further improves alignment quality by removing spurious correlations.

  • 2 authors
·
Apr 7

MUSCLE: A Model Update Strategy for Compatible LLM Evolution

Large Language Models (LLMs) are frequently updated due to data or architecture changes to improve their performance. When updating models, developers often focus on increasing overall performance metrics with less emphasis on being compatible with previous model versions. However, users often build a mental model of the functionality and capabilities of a particular machine learning model they are interacting with. They have to adapt their mental model with every update -- a draining task that can lead to user dissatisfaction. In practice, fine-tuned downstream task adapters rely on pretrained LLM base models. When these base models are updated, these user-facing downstream task models experience instance regression or negative flips -- previously correct instances are now predicted incorrectly. This happens even when the downstream task training procedures remain identical. Our work aims to provide seamless model updates to a user in two ways. First, we provide evaluation metrics for a notion of compatibility to prior model versions, specifically for generative tasks but also applicable for discriminative tasks. We observe regression and inconsistencies between different model versions on a diverse set of tasks and model updates. Second, we propose a training strategy to minimize the number of inconsistencies in model updates, involving training of a compatibility model that can enhance task fine-tuned language models. We reduce negative flips -- instances where a prior model version was correct, but a new model incorrect -- by up to 40% from Llama 1 to Llama 2.

  • 7 authors
·
Jul 12, 2024 3

The Butterfly Effect of Model Editing: Few Edits Can Trigger Large Language Models Collapse

Although model editing has shown promise in revising knowledge in Large Language Models (LLMs), its impact on the inherent capabilities of LLMs is often overlooked. In this work, we reveal a critical phenomenon: even a single edit can trigger model collapse, manifesting as significant performance degradation in various benchmark tasks. However, benchmarking LLMs after each edit, while necessary to prevent such collapses, is impractically time-consuming and resource-intensive. To mitigate this, we propose using perplexity as a surrogate metric, validated by extensive experiments demonstrating changes in an edited model's perplexity are strongly correlated with its downstream task performances. We further conduct an in-depth study on sequential editing, a practical setting for real-world scenarios, across various editing methods and LLMs, focusing on hard cases from our previous single edit studies. The results indicate that nearly all examined editing methods result in model collapse after only few edits. To facilitate further research, we have utilized GPT-3.5 to develop a new dataset, HardEdit, based on those hard cases. This dataset aims to establish the foundation for pioneering research in reliable model editing and the mechanisms underlying editing-induced model collapse. We hope this work can draw the community's attention to the potential risks inherent in model editing practices.

  • 6 authors
·
Feb 14, 2024

It's Not the Capability: Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers

A prevalent assumption in LLM agent deployment holds that more structured harnesses universally improve reliability, and that higher-capability models need proportionally less structural guidance -- together implying a monotone inverse relationship between model capability tier and optimal harness complexity. We test this hypothesis through a controlled 432-run experiment crossing six models across four capability tiers with three harness conditions (light, balanced, strict) on HEAT-24, a 24-task synthetic benchmark with git-based workspace verification. Our results refute the monotone inverse relationship on two fronts. First, for the frontier chat model evaluated (Gemini 2.5 Flash), increased harness verbosity lowers VTSR by 29-38 percentage points -- a harness-complexity paradox. Second, for the frontier reasoning model evaluated (Qwen3.5-122B, extended thinking enabled), strict harness achieves the highest VTSR (91.7%) and the lowest latency, the opposite of the prediction. Within the constrained tier, a 2B model (Gemma4:e2B) matches strong-open-tier stability at 91.7% across all harnesses. Because each tier is represented by a single model in this study, these results should be interpreted as model-specific observations; harness sensitivity appears non-monotone across the models evaluated, and depends critically on model type (chat vs. reasoning). We introduce a six-label failure taxonomy showing that format_violation dominates capable-model failures while wrong_file dominates low-capability failures, and we derive practical tier-aware harness selection guidelines.

  • 1 authors
·
May 25

Spotter: Let the Embodied Model Lead, and the VLM Reflect for It

Current embodied models do not respond to their own failures, although what just went wrong could inform a small adjustment on the next attempt, the kind of reflection behind the gains of thinking in language models. We test whether they can repair a known error, which requires producing a correction and judging whether it is right. Stopped at a failure and allowed to retry, they seldom repair it through their own randomness or from a language description of the error, and best-of-N selection cannot pick the successful candidate after a failure. We attribute this to training only on successful demonstrations and to inputs too narrow to show what went wrong, and conclude that reflection must come from a vision-language model (VLM), which takes in far more information, such as the episode history and text, and is more general. Prior VLM-led work has the VLM plan every step and invoke the embodied model as a tool, placing the VLM on the critical path. We propose Spotter, which reverses the roles: the embodied model leads and executes continuously, while the VLM runs in parallel, monitors through a lightweight local screener, intervenes only when an error is detected, reflects on and corrects it, and returns control. We run Spotter with Qwen and with GPT as the VLM, and both improve the embodied models; with GPT, Spotter improves Cosmos Policy and π_{0.5} by 5.6 and 7.5 percentage points on RoboCasa, and raises π_{0.5} from 47.2% to 57.0% on the Hard setting of RoboTwin 2.0 and from 53% to 83% on a real robot. Because the VLM steps in only when an error is confirmed, a successful episode with Qwen takes only 13 to 16 s longer than with the embodied model alone and about 70% less time than with a VLM-led baseline using the same model. Our code is available at https://github.com/zqc3117/Spotter.

  • 9 authors
·
Sep 28 1

LLM Swiss Round: Aggregating Multi-Benchmark Performance via Competitive Swiss-System Dynamics

The rapid proliferation of Large Language Models (LLMs) and diverse specialized benchmarks necessitates a shift from fragmented, task-specific metrics to a holistic, competitive ranking system that effectively aggregates performance across multiple ability dimensions. Primarily using static scoring, current evaluation methods are fundamentally limited. They struggle to determine the proper mix ratio across diverse benchmarks, and critically, they fail to capture a model's dynamic competitive fitness or its vulnerability when confronted with sequential, high-stakes tasks. To address this, we introduce the novel Competitive Swiss-System Dynamics (CSD) framework. CSD simulates a multi-round, sequential contest where models are dynamically paired across a curated sequence of benchmarks based on their accumulated win-loss record. And Monte Carlo Simulation (N=100,000 iterations) is used to approximate the statistically robust Expected Win Score (E[S_m]), which eliminates the noise of random pairing and early-round luck. Furthermore, we implement a Failure Sensitivity Analysis by parameterizing the per-round elimination quantity (T_k), which allows us to profile models based on their risk appetite--distinguishing between robust generalists and aggressive specialists. We demonstrate that CSD provides a more nuanced and context-aware ranking than traditional aggregate scoring and static pairwise models, representing a vital step towards risk-informed, next-generation LLM evaluation.

ByteDance-Seed ByteDance Seed
·
Dec 24, 2025 2

FailureSensorIQ: A Multi-Choice QA Dataset for Understanding Sensor Relationships and Failure Modes

We introduce FailureSensorIQ, a novel Multi-Choice Question-Answering (MCQA) benchmarking system designed to assess the ability of Large Language Models (LLMs) to reason and understand complex, domain-specific scenarios in Industry 4.0. Unlike traditional QA benchmarks, our system focuses on multiple aspects of reasoning through failure modes, sensor data, and the relationships between them across various industrial assets. Through this work, we envision a paradigm shift where modeling decisions are not only data-driven using statistical tools like correlation analysis and significance tests, but also domain-driven by specialized LLMs which can reason about the key contributors and useful patterns that can be captured with feature engineering. We evaluate the Industrial knowledge of over a dozen LLMs-including GPT-4, Llama, and Mistral-on FailureSensorIQ from different lens using Perturbation-Uncertainty-Complexity analysis, Expert Evaluation study, Asset-Specific Knowledge Gap analysis, ReAct agent using external knowledge-bases. Even though closed-source models with strong reasoning capabilities approach expert-level performance, the comprehensive benchmark reveals a significant drop in performance that is fragile to perturbations, distractions, and inherent knowledge gaps in the models. We also provide a real-world case study of how LLMs can drive the modeling decisions on 3 different failure prediction datasets related to various assets. We release: (a) expert-curated MCQA for various industrial assets, (b) FailureSensorIQ benchmark and Hugging Face leaderboard based on MCQA built from non-textual data found in ISO documents, and (c) LLMFeatureSelector, an LLM-based feature selection scikit-learn pipeline. The software is available at https://github.com/IBM/FailureSensorIQ.

  • 6 authors
·
Jun 3, 2025 1

Generalization is not a universal guarantee: Estimating similarity to training data with an ensemble out-of-distribution metric

Failure of machine learning models to generalize to new data is a core problem limiting the reliability of AI systems, partly due to the lack of simple and robust methods for comparing new data to the original training dataset. We propose a standardized approach for assessing data similarity in a model-agnostic manner by constructing a supervised autoencoder for generalizability estimation (SAGE). We compare points in a low-dimensional embedded latent space, defining empirical probability measures for k-Nearest Neighbors (kNN) distance, reconstruction of inputs and task-based performance. As proof of concept for classification tasks, we use MNIST and CIFAR-10 to demonstrate how an ensemble output probability score can separate deformed images from a mixture of typical test examples, and how this SAGE score is robust to transformations of increasing severity. As further proof of concept, we extend this approach to a regression task using non-imaging data (UCI Abalone). In all cases, we show that out-of-the-box model performance increases after SAGE score filtering, even when applied to data from the model's own training and test datasets. Our out-of-distribution scoring method can be introduced during several steps of model construction and assessment, leading to future improvements in responsible deep learning implementation.

  • 3 authors
·
Feb 22, 2025

Pruning as a Domain-specific LLM Extractor

Large Language Models (LLMs) have exhibited remarkable proficiency across a wide array of NLP tasks. However, the escalation in model size also engenders substantial deployment costs. While few efforts have explored model pruning techniques to reduce the size of LLMs, they mainly center on general or task-specific weights. This leads to suboptimal performance due to lacking specificity on the target domain or generality on different tasks when applied to domain-specific challenges. This work introduces an innovative unstructured dual-pruning methodology, D-Pruner, for domain-specific compression on LLM. It extracts a compressed, domain-specific, and task-agnostic LLM by identifying LLM weights that are pivotal for general capabilities, like linguistic capability and multi-task solving, and domain-specific knowledge. More specifically, we first assess general weight importance by quantifying the error incurred upon their removal with the help of an open-domain calibration dataset. Then, we utilize this general weight importance to refine the training loss, so that it preserves generality when fitting into a specific domain. Moreover, by efficiently approximating weight importance with the refined training loss on a domain-specific calibration dataset, we obtain a pruned model emphasizing generality and specificity. Our comprehensive experiments across various tasks in healthcare and legal domains show the effectiveness of D-Pruner in domain-specific compression. Our code is available at https://github.com/psunlpgroup/D-Pruner.

  • 8 authors
·
May 10, 2024

ThinkFL: Self-Refining Failure Localization for Microservice Systems via Reinforcement Fine-Tuning

As modern microservice systems grow increasingly popular and complex-often consisting of hundreds or even thousands of fine-grained, interdependent components-they are becoming more susceptible to frequent and subtle failures. Ensuring system reliability therefore hinges on accurate and efficient failure localization. Traditional failure localization approaches based on small models lack the flexibility to adapt to diverse failure scenarios, while recent LLM-based methods suffer from two major limitations: they often rely on rigid invocation workflows that constrain the model's ability to dynamically explore optimal localization paths, and they require resource-intensive inference, making them cost-prohibitive for real-world deployment. To address these challenges, we explore the use of reinforcement fine-tuning to equip lightweight LLMs with reasoning and self-refinement capabilities, significantly improving the cost-effectiveness and adaptability of LLM-based failure localization. We begin with an empirical study to identify three key capabilities essential for accurate localization. Building on these insights, we propose a progressive multi-stage GRPO fine-tuning framework, which integrates a multi-factor failure localization grader and a recursion-of-thought actor module. The resulting model, ThinkFL, not only outperforms existing state-of-the-art LLMs and baseline methods in localization accuracy but also reduces end-to-end localization latency from minutes to seconds, demonstrating strong potential for real-world applications.

  • 9 authors
·
Apr 25, 2025

Estimating Tail Risks in Language Model Output Distributions

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

  • 7 authors
·
Apr 23

A Learning-Rate-Gated Failure of GRPO in a Small Language and Vision-Language Model Web Agent: A Controlled Null and Its Mechanism

Reinforcement learning with verifiable rewards, and Group Relative Policy Optimization (GRPO) in particular, is now run routinely on a supervised checkpoint in the hope of producing a stronger agent. We ask whether it adds skill to a small language and vision-language model web agent at the 4B to 8B scale, or whether it mostly reshapes behavior the supervised model already has. Across a control grid of 18 runs that varies learning rate, KL weight, seed, initialization, and clipping, no configuration credibly improves the success rate of a strong supervised baseline on tasks the agent has largely mastered. On the text track, moderate to high learning rates make it credibly worse. The null holds under paired testing, 25 evaluation seeds, 6 training seeds, changes to the recipe, both text and Set-of-Marks screenshot observations, and scaling the backbone to 8B; the credible harm is a text-track finding and is only nominal under Set-of-Marks. To show that the null reflects the setting and not a broken pipeline, we run the identical harness, reward, and recipe on tasks whose reward is reachable by sampling, and there the success rate rises by 22 points with a paired interval that excludes zero. GRPO therefore helps only when there is headroom to climb, meaning the sampled policy already succeeds more often than the greedy one. We then explain the failure. A middle learning rate degrades the agent and a high one collapses it, and the two regimes form a double dissociation: grafting localizes the degrade regime to the attention and MLP blocks, while the collapse regime cannot be traced to any single group, and the embedding change that dominates the weight movement is causally inert. At 4B, effective rank in the late layers tracks capability in both directions; at 8B the two come apart. This coupling is specific to the smaller model, so we report it as scale-dependent.

  • 6 authors
·
Jul 13

How Many Tries Does It Take? Iterative Self-Repair in LLM Code Generation Across Model Scales and Benchmarks

Large language models frequently fail to produce correct code on their first attempt, yet most benchmarks evaluate them in a single-shot setting. We investigate iterative self-repair (feeding execution errors back to the model for correction) across seven models spanning three families and both open-weight and proprietary providers: Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout (MoE, 16 experts), Llama 4 Maverick (MoE, 128 experts), Qwen3 32B, Gemini 2.5 Flash, and Gemini 2.5 Pro. On HumanEval (164 problems) and MBPP Sanitized (257 problems) with up to five attempts, self-repair universally improves pass rates: +4.9 to +17.1 pp on HumanEval and +16.0 to +30.0 pp on MBPP. Gemini 2.5 Flash achieves the highest final pass rates (96.3% HumanEval, 93.8% MBPP). Most gains concentrate in the first two rounds.Error-type analysis shows assertion errors (logical mistakes) are the hardest to repair at ~45%, while syntax and name errors are repaired at substantially higher rates, connecting to broader findings on the limits of LLM self-correction. Prior work found that weaker models fail at self-repair or require fine-tuning; we show that modern instruction-tuned models succeed with prompting alone, even at 8B scale. We also provide the first comparison of dense and MoE architectures for self-repair, and extend the repair-vs-resampling tradeoff analysis to modern models. A prompt ablation reveals chain-of-thought repair yields up to +5.5 pp additional self-repair gain (measured as improvement in repair delta) over minimal prompting for capable models.

  • 1 authors
·
Apr 11

Tracing Agentic Failure from the Flow of Success

Failure attribution for LLM-based agentic systems, i.e., identifying which steps in a failure trajectory caused the task to fail, is critical for debugging and improving these systems. Existing approaches either rely on prompting-based pipelines, which are computationally expensive, or require post-training on failure trajectories with step-level error annotations, which are costly to collect and difficult to scale. We argue that a practical failure attribution model should be lightweight and trainable without step-level supervision on failure data. To this end, we address unsupervised failure attribution, i.e., training exclusively on successful trajectories and identifying error steps at inference time given a failure trajectory. We propose OAT, which casts this problem as one-class learning with neural controlled differential equations, modeling the dynamical pattern of successful trajectories in latent space. At inference time, each step in a failure trajectory is assigned an anomaly score based on its deviation from the dynamics learned on successful trajectories, which is then used to form a set of error steps. With training on only 100 successful trajectories, experiments show that OAT is 200--5000 times faster than prompting-based baselines, and, at the same time, consistently outperforms them in both in-domain and out-of-distribution datasets with +20% and +7% F1 scores, respectively, demonstrating that OAT is a promising and efficient direction for diagnosing agentic system failures.

  • 4 authors
·
Jul 13 2

Explainable AI Methods for Neuroimaging: Systematic Failures of Common Tools, the Need for Domain-Specific Validation, and a Proposal for Safe Application

Trustworthy interpretation of deep learning models is critical for neuroimaging applications, yet commonly used Explainable AI (XAI) methods lack rigorous validation, risking misinterpretation. We performed the first large-scale, systematic comparison of XAI methods on ~45,000 structural brain MRIs using a novel XAI validation framework. This framework establishes verifiable ground truth by constructing prediction tasks with known signal sources - from localized anatomical features to subject-specific clinical lesions - without artificially altering input images. Our analysis reveals systematic failures in two of the most widely used methods: GradCAM consistently failed to localize predictive features, while Layer-wise Relevance Propagation generated extensive, artifactual explanations that suggest incompatibility with neuroimaging data characteristics. Our results indicate that these failures stem from a domain mismatch, where methods with design principles tailored to natural images require substantial adaptation for neuroimaging data. In contrast, the simpler, gradient-based method SmoothGrad, which makes fewer assumptions about data structure, proved consistently accurate, suggesting its conceptual simplicity makes it more robust to this domain shift. These findings highlight the need for domain-specific adaptation and validation of XAI methods, suggest that interpretations from prior neuroimaging studies using standard XAI methodology warrant re-evaluation, and provide urgent guidance for practical application of XAI in neuroimaging.

  • 6 authors
·
Aug 4, 2025

ActionLens: Diagnosing Spatial-Temporal Binding Failures in Vision-Language Models

Video-capable vision-language models score above 80\% on popular benchmarks yet struggle with spatial-temporal binding: associating the right action with the right person at the right moment. We introduce ActionLens, a diagnostic benchmark of 6,701 multiple-choice video questions spanning five targeted diagnostics: transition detection, actor-specific identification, concurrent action binding, directed interaction reasoning, and gaze detection. Ground-truth answers are derived deterministically from 1.58 million per-second, per-person annotations. Fourteen rounds of human quality engineering raised answer clarity from 53% to above 90% human accuracy. Across 20 VLMs, the full-set leader scores 68.8%; on the human-reviewed subset, it scores 65.9% versus 91.0% for the pooled human reference. Gaze detection remains near chance against 89.6% human accuracy. On actor disambiguation, reference-interface controls show that relational descriptions recover 5.55--13.25 points over static coordinates, confirming a substantial numeric-parsing penalty; yet visual boxes still lead every model by 1.15--6.50 points, exposing a residual unboxed actor-resolution gap. A binding-trap analysis shows models systematically select the wrong actor's action. ActionLens provides diagnostic measurements of these distinct failure modes across model families and scales for direct comparison. We release all data, code, and evaluation scripts at https://anonymous.4open.science/r/lmms-eval-2276

  • 6 authors
·
Sep 27

DatBench: Discriminative, Faithful, and Efficient VLM Evaluations

Empirical evaluation serves as the primary compass guiding research progress in foundation models. Despite a large body of work focused on training frontier vision-language models (VLMs), approaches to their evaluation remain nascent. To guide their maturation, we propose three desiderata that evaluations should satisfy: (1) faithfulness to the modality and application, (2) discriminability between models of varying quality, and (3) efficiency in compute. Through this lens, we identify critical failure modes that violate faithfulness and discriminability, misrepresenting model capabilities: (i) multiple-choice formats reward guessing, poorly reflect downstream use cases, and saturate early as models improve; (ii) blindly solvable questions, which can be answered without images, constitute up to 70% of some evaluations; and (iii) mislabeled or ambiguous samples compromise up to 42% of examples in certain datasets. Regarding efficiency, the computational burden of evaluating frontier models has become prohibitive: by some accounts, nearly 20% of development compute is devoted to evaluation alone. Rather than discarding existing benchmarks, we curate them via transformation and filtering to maximize fidelity and discriminability. We find that converting multiple-choice questions to generative tasks reveals sharp capability drops of up to 35%. In addition, filtering blindly solvable and mislabeled samples improves discriminative power while simultaneously reducing computational cost. We release DatBench-Full, a cleaned evaluation suite of 33 datasets spanning nine VLM capabilities, and DatBench, a discriminative subset that achieves 13x average speedup (up to 50x) while closely matching the discriminative power of the original datasets. Our work outlines a path toward evaluation practices that are both rigorous and sustainable as VLMs continue to scale.

  • 31 authors
·
Jan 5

FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models

Recent advances in robotic manipulation have integrated low-level robotic control into Vision-Language Models (VLMs), extending them into Vision-Language-Action (VLA) models. Although state-of-the-art VLAs achieve strong performance in downstream robotic applications, supported by large-scale crowd-sourced robot training data, they still inevitably encounter failures during execution. Enabling robots to reason and recover from unpredictable and abrupt failures remains a critical challenge. Existing robotic manipulation datasets, collected in either simulation or the real world, primarily provide only ground-truth trajectories, leaving robots unable to recover once failures occur. Moreover, the few datasets that address failure detection typically offer only textual explanations, which are difficult to utilize directly in VLA models. To address this gap, we introduce FailSafe, a novel failure generation and recovery system that automatically produces diverse failure cases paired with executable recovery actions. FailSafe can be seamlessly applied to any manipulation task in any simulator, enabling scalable creation of failure action data. To demonstrate its effectiveness, we fine-tune LLaVa-OneVision-7B (LLaVa-OV-7B) to build FailSafe-VLM. Experimental results show that FailSafe-VLM successfully helps robotic arms detect and recover from potential failures, improving the performance of three state-of-the-art VLA models (pi0-FAST, OpenVLA, OpenVLA-OFT) by up to 22.6% on average across several tasks in Maniskill. Furthermore, FailSafe-VLM could generalize across different spatial configurations, camera viewpoints, object and robotic embodiments. We plan to release the FailSafe code to the community.

  • 7 authors
·
Oct 1, 2025

LEMMA: Learning from Errors for MatheMatical Advancement in LLMs

Large language models (LLMs) have demonstrated remarkable reasoning capability in solving mathematical problems. However, existing approaches primarily focus on improving the quality of correct training data, e.g., distilling high-quality correct solutions from advanced models, neglecting the value contained in error data, potentially hindering the model's reflective ability. Though some studies attempt to leverage error data, they often involve complex mechanisms, such as Monte Carlo Tree Search (MCTS) to explore error nodes. In this work, we propose to enhance LLMs' reasoning ability by Learning from Errors for Mathematical Advancement (LEMMA). LEMMA constructs data consisting of an incorrect solution with an erroneous step and a reflection connection to a correct solution for fine-tuning. Specifically, we systematically analyze the model-generated error types and introduce an error-type grounded mistake augmentation method to collect diverse and representative errors. Correct solutions are either from fixing the errors or generating a fresh start. Through a model-aware smooth reflection connection, the erroneous solution is transferred to the correct one. By fine-tuning on the constructed dataset, the model is able to self-correct errors autonomously within the generation process without relying on external critique models. Experimental results demonstrate that LEMMA achieves significant performance improvements over other strong baselines.

  • 10 authors
·
Mar 21, 2025 2

Scaling Cross-Environment Failure Reasoning Data for Vision-Language Robotic Manipulation

Robust robotic manipulation requires reliable failure detection and recovery. Although recent Vision-Language Models (VLMs) show promise in robot failure detection, their generalization is severely limited by the scarcity and narrow coverage of failure data. To address this bottleneck, we propose an automatic framework for generating diverse robotic planning and execution failures across both simulated and real-world environments. Our approach perturbs successful manipulation trajectories to synthesize failures that reflect realistic failure distributions, and leverages VLMs to produce structured step-by-step reasoning traces. This yields FailCoT, a large-scale failure reasoning dataset built upon the RLBench simulator and the BridgeDataV2 real-robot dataset. Using FailCoT, we train Guardian, a multi-view reasoning VLM for unified planning and execution verification. Guardian achieves state-of-the-art performance on three unseen real-world benchmarks: RoboFail, RoboVQA, and our newly introduced UR5-Fail. When integrated with a state-of-the-art LLM-based manipulation policy, it consistently boosts task success rates in both simulation and real-world deployment. These results demonstrate that scaling high-quality failure reasoning data is critical for improving generalization in robotic failure detection. Code, Data, and Models available at https://www.di.ens.fr/willow/research/guardian/.

  • 4 authors
·
Mar 29

Foundations of Unknown-aware Machine Learning

Ensuring the reliability and safety of machine learning models in open-world deployment is a central challenge in AI safety. This thesis develops both algorithmic and theoretical foundations to address key reliability issues arising from distributional uncertainty and unknown classes, from standard neural networks to modern foundation models like large language models (LLMs). Traditional learning paradigms, such as empirical risk minimization (ERM), assume no distribution shift between training and inference, often leading to overconfident predictions on out-of-distribution (OOD) inputs. This thesis introduces novel frameworks that jointly optimize for in-distribution accuracy and reliability to unseen data. A core contribution is the development of an unknown-aware learning framework that enables models to recognize and handle novel inputs without labeled OOD data. We propose new outlier synthesis methods, VOS, NPOS, and DREAM-OOD, to generate informative unknowns during training. Building on this, we present SAL, a theoretical and algorithmic framework that leverages unlabeled in-the-wild data to enhance OOD detection under realistic deployment conditions. These methods demonstrate that abundant unlabeled data can be harnessed to recognize and adapt to unforeseen inputs, providing formal reliability guarantees. The thesis also extends reliable learning to foundation models. We develop HaloScope for hallucination detection in LLMs, MLLMGuard for defending against malicious prompts in multimodal models, and data cleaning methods to denoise human feedback used for better alignment. These tools target failure modes that threaten the safety of large-scale models in deployment. Overall, these contributions promote unknown-aware learning as a new paradigm, and we hope it can advance the reliability of AI systems with minimal human efforts.

  • 1 authors
·
May 19, 2025

Mistake Notebook Learning: Batch-Clustered Failures for Training-Free Agent Adaptation

With the growing adoption of Large Language Model (LLM) agents in persistent, real-world roles, they naturally encounter continuous streams of tasks and inevitable failures. A key limitation, however, is their inability to systematically learn from these mistakes, forcing them to repeat identical errors in similar contexts. Unlike prior training-free methods that primarily store raw instance-level experience or focus on retrieving successful trajectories, we propose Mistake Notebook Learning (MNL), a novel memory framework that enables agents to self-curate generalizable guidance from batch-clustered failures. This mechanism allows agents to distill shared error patterns into structured "mistake notes," updating an external memory only when batch performance improves to ensure stability. To further amplify adaptability, we integrate MNL with test-time scaling, leveraging aggregated failure patterns to actively steer the search process away from known pitfalls. Experiments on mathematical reasoning, Text-to-SQL, and interactive agent benchmarks show that MNL achieves competitive performance compared to existing memory mechanisms and in-context methods in both effectiveness and efficiency. These findings position structured mistake abstraction as a critical lever for robust agent evolution, enabling continuous improvement without the cost of parameter updates. The code is available at https://github.com/Bairong-Xdynamics/MistakeNotebookLearning/tree/main.

  • 5 authors
·
Dec 12, 2025

Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It

Tool use enables large language models (LLMs) to perform complex tasks, and recent agentic reinforcement learning (RL) methods show promise for enhancing model capabilities. However, RL alone often leads to instability or limited gains in tool-use tasks. In our experiments, some models exhibit catastrophic collapse, where performance abruptly drops and tool-invocation structures fail. The analysis reveals that these failures stem from unexpected probability spikes in specific control tokens, disrupting structured execution, yet the underlying tool-use capability remains intact, merely obscured by specific formats. To address this, we systematically investigate a diverse set of supervisory signals, including off-policy supervision, hint-based guidance, erroneous example supervision, and others, applied under both synchronous and interleaved training schemes. We find that interleaving supervised fine-tuning (SFT) with RL substantially improves stability, but exhibits degraded performance under format and content out-of-distribution (OOD) evaluation. We also analyze the impact of learning rates and generalization across settings. These results highlight the importance of understanding RL failures and demonstrate how diverse supervisory signals can guide exploratory learning, enabling robust training of LLMs for complex, multi-step tool-use tasks. Our Code is available at https://github.com/hypasd-art/Tool-RL-Box.

Catastrophic Failure of LLM Unlearning via Quantization

Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from the diverse and sensitive nature of their training data, which can include copyrighted and private content. Machine unlearning has been introduced as a viable solution to remove the influence of such problematic content without the need for costly and time-consuming retraining. This process aims to erase specific knowledge from LLMs while preserving as much model utility as possible. Despite the effectiveness of current unlearning methods, little attention has been given to whether existing unlearning methods for LLMs truly achieve forgetting or merely hide the knowledge, which current unlearning benchmarks fail to detect. This paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. To thoroughly evaluate this phenomenon, we conduct comprehensive experiments using various quantization techniques across multiple precision levels. We find that for unlearning methods with utility constraints, the unlearned model retains an average of 21\% of the intended forgotten knowledge in full precision, which significantly increases to 83\% after 4-bit quantization. ... Our code is available at: https://github.com/zzwjames/FailureLLMUnlearning{https://github.com/zzwjames/FailureLLMUnlearning}.

  • 9 authors
·
Mar 20, 2025

A Trace-Based Assurance Framework for Agentic AI Orchestration: Contracts, Testing, and Governance

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.

  • 3 authors
·
Mar 17

VLS: Steering Pretrained Robot Policies via Vision-Language Models

Why do pretrained diffusion or flow-matching policies fail when the same task is performed near an obstacle, on a shifted support surface, or amid mild clutter? Such failures rarely reflect missing motor skills; instead, they expose a limitation of imitation learning under train-test shifts, where action generation is tightly coupled to training-specific spatial configurations and task specifications. Retraining or fine-tuning to address these failures is costly and conceptually misaligned, as the required behaviors already exist but cannot be selectively adapted at test time. We propose Vision-Language Steering (VLS), a training-free framework for inference-time adaptation of frozen generative robot policies. VLS treats adaptation as an inference-time control problem, steering the sampling process of a pretrained diffusion or flow-matching policy in response to out-of-distribution observation-language inputs without modifying policy parameters. By leveraging vision-language models to synthesize trajectory-differentiable reward functions, VLS guides denoising toward action trajectories that satisfy test-time spatial and task requirements. Across simulation and real-world evaluations, VLS consistently outperforms prior steering methods, achieving a 31% improvement on CALVIN and a 13% gain on LIBERO-PRO. Real-world deployment on a Franka robot further demonstrates robust inference-time adaptation under test-time spatial and semantic shifts. Project page: https://vision-language-steering.github.io/webpage/

allenai Ai2
·
Feb 3 3