arXiv subjects

Benjamin Van Durme

Publications and source records attributed to Benjamin Van Durme.

At least 19 recordsLinked to original sources

arXiv2026

Tracing the Scientific Discovery Capacity of LLMs to Their Training Data

Large language models (LLMs) show growing promise for scientific discovery, yet basic questions remain about what enables reliable discovery. How can we trace the scientific discovery capacity of LLMs to their training data? We study this question as one of missing probability mass and develop a theoretical account in which monofacts-the facts observed once during training-set the discovery capacity. In materials science and regulatory genomics, we directly manipulate monofact rates in training sets and find that more monofacts consistently yield more discoveries across two open-weight model families, sampling strategies, and reasoning budgets. Further experiments show the relationship cannot be explained by training-set diversity alone and can transfer across scientific domains when paired with in-domain data. Together, our results identify monofact rate in training data as a key factor governing discoveries in language models.

cs.CL↗

arXiv2026

FOCUS: Active Critic Learning for Long-Horizon Language Model Agents

Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions. Learned critics provide value estimates for intermediate states, but these estimates can be inaccurate, while estimating intermediate values with repetitive continuation samples from points within a trajectory incurs substantial costs. To mitigate these issues, we introduce FOCUS, an RL method that uses selective branching to help the critic network learn its value estimator. The rollout controller facilitates active learning for the critic by sampling continuations from states that are expected to most reduce the critic's estimation error. These refined value estimates then score the segment between consecutive states with a TD($λ$) advantage, while the sampled evidence supplies training targets for the critic. FOCUS supports long-horizon agentic reinforcement learning through a modular context-management interface that decouples context management from rollout allocation. Our experiments show that using compaction-induced segment boundaries for value estimation and credit assignment, FOCUS outperforms GRPO and VinePPO with matched per-step generation budgets on legal reasoning, deep search, and code localization benchmarks, improving held-out accuracy at Qwen3.5-9B scale by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs. 48.2) and by 4.1 points on DeonticBench (56.5 vs. 52.4).

cs.LG↗

arXiv2026

AdaPRA: Process Reward Agents with Adaptive Rubrics for Data Science

Data science agents can automate end-to-end machine learning workflows, yet they remain far from competitive on human leaderboards. Agents make process errors such as insufficient exploration, unsound validation, coding bugs, and poor use of a fixed budget (re-running the same exploratory data analysis and baseline models in every attempt). We propose AdaPRA to address these issues: a persistent coding agent harness that accumulates knowledge across attempts on the same task, with rubric-guided process supervision that identifies and corrects errors. Our process reward agent audits the agent's trajectory and workspace by executing checks, and adapts its rubrics to each task. We find that rubrics seeded from easier tasks generalize to harder tasks, with rubric adaptation further improving performance. With five attempts per task, AdaPRA improves an 80B Qwen model's leaderboard percentile by 24.5 points on DSPredict-Easy and 14.8 points on DSPredict-Hard while reducing process errors. We also show that supervision with rubrics mined from one base model's runs improves another (DeepSeek-V3.2), and that AdaPRA beats DataPRM, a competitive PRM for data science. We also share our code for reproducibility.

cs.AI↗

arXiv2026

Attention Sinks in VLMs Are Borrowed, Not Seen

We study how to detect and mitigate visual attention sinks in vision-language models (VLMs). Visual attention sinks are image tokens that absorb a large share of attention over the image regardless of the query or task, and they mislead methods that treat attention maps as a record of where the model ``looked.'' Prior work has detected sinks through massive activation dimensions and proposed redistributing their attention, but none have addressed what makes a token become a sink. Comparing six candidate detectors on ten VLMs from four families, we find that only the centered cosine similarity between an image token's key and the key of the language model's text-sink detects visual sinks. Thus, a visual sink is an image token whose key aligns with the text-sink's key. Moving a sink's key component along this direction to the median of the image keys sends its attention back to the text-sink (\textbf{Are Borrowed}) while leaving model answers unchanged~(\textbf{Not Seen}). Building on this, we propose a training-free, entropy-triggered re-look that first cleans the attention map, then crops the most-attended region of the cleaned map and inserts it into the reasoning trace. Cleaning adds negligible computation to the forward pass, and the re-look improves accuracy with a mean gain of $+4.3$ points on fine-detail VQA benchmarks across model families.

cs.AI↗

arXiv2026

TeleTune: Evolving Agent Skills From Offline Telemetry

Computer-use agents need to capture procedural knowledge of how people use software. User telemetry offers a scalable source of this knowledge. However, learning reusable skills from these logs requires addressing three challenges: (1) Goal Underspecification, since logs do not record the goal behind each action; (2) Non-Replayability, since past activity cannot be replayed to evaluate skill updates; and (3) Interleaved Trajectories, since logs may mix several tasks without marking their boundaries. To address these, we introduce TeleTune, a framework for learning a textual skill library from offline logs without recorded goals, cannot be replayed during optimization, and may interleave tasks. TeleTune uses action-prediction errors on logged trajectories to propose library edits and keep only those that improve held-out action-prediction accuracy, which we call skill-guided progress. The learned workflows also enable retrieval of demonstrations that cover the subgoals of a new task. At test time, the agent is provided with the learned library and the workflow-based retrieved demonstrations. Experiments on WorkArena and Online-Mind2Web show that TeleTune outperforms random retrieval, Agent Workflow Memory (AWM), and their combination. We find that the best baseline varies by setting, whereas TeleTune achieves average success rates of 77.1% and 80.6%, respectively, improving over the strongest baseline on each benchmark by 6.7% and 7.7%. Under the heaviest perturbation of the WorkArena training data,TeleTune keeps the highest average success rate at 68.5%, 6.3% above the strongest baseline. Our analyses show (1) skill optimization and workflow-based retrieval are complementary, (2) optimizing on fixed logs costs 5 to 75 times fewer tokens than validating the same edits with live episodes, (3) skill-guided progress tracks the live success rate.

cs.AI↗

arXiv2026

SIGMA: Self-Improving Alignment Generalization from a Model Spec

LLM agents are increasingly capable of executing complex tasks and of recursively improving themselves on easy-to-verify objectives such as software engineering and mathematics. Since alignment is much harder to verify, this creates a growing risk of capabilities increasing without appropriate safety alignment, especially as capabilities expand to auto-research and cybersecurity. Existing approaches focus on capability self-improvement using verifiable feedback or on alignment training with supervision from stronger models or curated data, creating an external supervision bottleneck for alignment. We ask whether current models can improve their own safety alignment, and propose SIGMA, a data generation and training pipeline enabling alignment self-improvement that generalizes to out-of-distribution settings. Given only a "Model Spec" stating the model's desired behavior, SIGMA leverages a model's reasoning capabilities to strengthen its own safety reasoning. SIGMA first performs spec-guided task synthesis, using the candidate model as a task designer agent to generate diverse alignment dilemma scenarios and convert them into training tasks that stress-test its understanding of the Model Spec. Next, SIGMA conducts self-judged alignment training through supervised fine-tuning and rubric-based reinforcement learning with the model itself as the reward model. Despite training only on single-turn chat data, SIGMA improves safety alignment in multi-turn agentic environments (AgentHarm harmfulness decreases from 22.6 to 14.8; Agentic Misalignment decreases from 79.1 to 3.8), outperforms Deliberative Alignment and Constitutional AI baselines, and retains general capability. Analyses show that a Model Spec balancing harmlessness and helpfulness, test-time reasoning for safety deliberation, and high-quality rubrics from SIGMA's task designer agent are crucial for effective self-improvement.

cs.AI↗

arXiv2026

MultiVENT-Raw: A Benchmark for Retrieval and Reasoning over Raw Videos

Online information is increasingly consumed in video format. Much of this comes in the form of *raw video*: continuous footage taken on a cell phone, with a hand-held camera, or via CCTV, which is then directly uploaded to social media platforms and content sharing services. Whereas professional or even amateur-edited footage tends to feature scripted speech, chyrons, graphics, and metadata that help contextualize its subject matter, raw video typically contains none of these things, making it a much more challenging medium for information retrieval and machine understanding. To facilitate progress in this domain, we release MultiVENT-Raw, a multilingual collection of nearly 120,000 primarily raw videos (over 5,300 total hours), paired with 130 events and 222 event-centric queries, along with human-annotated video relevance judgments and human-extracted key facts for relevant videos. MultiVENT-Raw supports both a retrieval task---to identify videos in the collection relevant to a query event---and a generation task---to summarize event-related videos into a coherent report for a target user. We benchmark strong baselines on MultiVENT-Raw, showing both tasks to be challenging even for some of the latest multimodal models.

cs.CV↗

arXiv2026

Unified Multimodal Uncertain Inference

We introduce Unified Multimodal Uncertain Inference (UMUI), a multimodal inference task spanning text, audio, and video, where models must produce calibrated probability estimates of hypotheses conditioned on a premise in any modality or combination. While uncertain inference has been explored in text, extension to other modalities has been limited to single-modality binary entailment judgments, leaving no framework for fine-grained probabilistic reasoning in or across other modalities. To address this, we curate a human-annotated evaluation set with scalar probability judgments across audio, visual, and audiovisual settings, and additionally evaluate on existing text and audio benchmarks. We introduce CLUE (Calibrated Latent Uncertainty Estimation), which combines self-consistent teacher calibration and distribution-based confidence probing to produce calibrated predictions. We demonstrate that our 3B-parameter model achieves equivalent or stronger performance than zero-shot baselines up to 32B parameters across all modalities.

cs.CV↗

arXiv2026

Knowledge Pull Requests for Continual Document Authoring

We introduce Knowledge Pull Requests (KPRs), a framework for continual document authoring that makes each change interpretable. Documents require ongoing revision as new knowledge surfaces from other sources, languages, or times, but existing approaches either edit with no account of what knowledge changed or regenerate from scratch. A KPR integrates new knowledge into a document by extracting claims, filtering and routing them to sections, and flagging conflicts with existing content, producing a ChangeLog that separates what knowledge changes (claim proposal) from how the text changes (document diff). We evaluate KPRs on revising Wikipedia across languages and updating query-driven reports on RAGTIME. KPRs integrate more information and better preserve existing content than rewriting from sources or regenerating from scratch, while adding the most information per token generated. A KPR-revised article also grounds question answering better than a frontier model with search, which does not surface knowledge documented only in other languages.

cs.CL↗

arXiv2026

Mining Legal Arguments in U.S. Corporate Case Law

Legal argument mining supports passage classification, retrieval, and argument completion. This work introduces an expert-annotated dataset of 42 U.S. federal tax opinions on corporate reorganizations under I.R.C. §368. To our knowledge, it is the first expert-annotated, tree-structured argument corpus for this domain. Explicit spans receive one of five functional labels: Rule, Analysis, Conclusion, Background Facts, and Procedural History. Rule, Analysis, and Conclusion spans can be linked into directed support trees, while Background Facts and Procedural History serve a contextual function. The corpus provides span-based, sentence-based, flat, and tree-structured representations. Agreement analysis shows that functional node labels are more reliable than directed support edges and implicit intermediate conclusions. Directed-path agreement is stronger than direct-edge agreement, which indicates that broad reachability is more stable than exact local decomposition. Classification experiments show that functional labels are learnable under case-disjoint evaluation. Retrieval experiments show that supervised fine-tuning improves within-case retrieval. However, cross-case generalization remains weak. The dataset supports legal passage classification and provides a conservative benchmark for structured argument mining in U.S. federal tax case law.

cs.CL↗

arXiv2026

Are Finer Citations Always Better? Rethinking Granularity for Attributed Generation

Citation granularity -- whether to cite individual sentences, paragraphs, or documents -- is a critical design choice in attributed generation. While fine-grained citations are commonly preferred for precise human verification, their impact on model performance remains under-explored. We analyze four model scales (8B-120B) and demonstrate that enforcing fine-grained (sentence-level) citations forfeits gains of 2-97% (median 40%) relative to the best-performing granularity, and up to 338% on individual tasks. Strikingly, setting citation granularity to its optimal value (based on attribution quality) unlocks these substantial gains while leaving overall answer correctness essentially unchanged (between -2.3% and +4.4%). We observe a consistent pattern where attribution quality peaks at intermediate (paragraph-level) granularities: finer citations appear to sever the semantic dependencies needed to ground a claim, while excessively coarse citations introduce distracting noise. Importantly, this performance gap varies with scale: when a claim rests on a small or moderate amount of evidence, it disproportionately penalizes larger models by disrupting the multi-sentence information synthesis at which they excel. Fine-grained citation rests on the premise that a sentence is a sufficient unit of evidence on its own. Our results indicate that it often is not, and that this is a property of the model rather than of the citation standard. Standards fixed for human verifiability may therefore paradoxically degrade the very attribution they aim to ensure; effective attribution requires matching granularity to the model's semantic scope rather than fixing it by convention.

cs.CL↗

arXiv2026

Seeing Through the MiRAGE: Evaluating Multimodal Retrieval Augmented Generation

We introduce MiRAGE, an evaluation framework for retrieval-augmented generation (RAG) from multimodal sources. As audiovisual media becomes a more prevalent source of information online, RAG systems must integrate such media into generation. Yet, existing evaluation methods for RAG are largely text-centric and do not readily transfer to multimodal settings. MiRAGE is a claim-centric approach to multimodal RAG evaluation, consisting of InfoF1, which assesses factuality and information coverage, and CiteF1, which assesses citation support and completeness. We show that, when applied by humans, MiRAGE strongly aligns with extrinsic judgments of output quality. We additionally introduce an automatic implementation of MiRAGE and compare it to multimodal variants of three prominent text-centric RAG metrics---ALCE, ARGUE, and RAGAS---finding that MiRAGE outperforms all three on text while being the only one to generalize to multimodal sources. We release open-source implementations and outline evaluation methods for multimodal RAG.

cs.CL↗

arXiv2026

FrankenReport: Early Exiting in Long-Form Generation Using Expected Value of Computation

While deep research systems address interactive information-seeking needs impressively, their real-world deployments face latency and resource-consumption challenges. We present FrankenReport, an interface for long-form knowledge-seeking report generation that supports adaptive early exiting per section: it evaluates intermediate outputs during generation and predicts whether further targeted computation will yield significant quality gains. In a simulation study, FrankenReport outperforms random allocation baselines by a large margin (up to 4x) under low budgets and smoothly recovers full-pipeline quality as the budget grows, showing that future quality gains are predictable from intermediate drafts. Through experiments and user studies, we further show that despite varying preferences across users and topics, FrankenReport adapts to simple, natural user feedback as efficiently as methods requiring much costlier supervision such as generated drafts and explicit rationales.

cs.HC↗

arXiv2026

Many-Tier Instruction Hierarchy in LLM Agents

Large language model agents receive instructions from many sources-system messages, user prompts, tool outputs, other agents, and more-each carrying different levels of trust and authority. When these instructions conflict, agents must reliably follow the highest-privilege instruction to remain safe and effective. The dominant paradigm, instruction hierarchy (IH), assumes a fixed, small set of privilege levels (typically fewer than five) defined by rigid role labels (e.g., system > user). This is inadequate for real-world agentic settings, where conflicts can arise across far more sources and contexts. In this work, we propose Many-Tier Instruction Hierarchy (ManyIH), a paradigm for resolving instruction conflicts among instructions with arbitrarily many privilege levels. We introduce ManyIH-Bench, the first benchmark for ManyIH. ManyIH-Bench requires models to navigate up to 12 levels of conflicting instructions with varying privileges, comprising 853 agentic tasks (427 coding and 426 instruction-following). ManyIH-Bench composes constraints developed by LLMs and verified by humans to create realistic and difficult test cases spanning 46 real-world agents. Our experiments show that even the current frontier models perform poorly (~40% accuracy) when instruction conflict scales. This work underscores the urgent need for methods that explicitly target fine-grained, scalable instruction conflict resolution in agentic settings.

cs.CL↗

arXiv2026

Can Coding Agents Reproduce Findings in Computational Materials Science?

Large language models are increasingly deployed as autonomous coding agents and have achieved remarkably strong performance on software engineering benchmarks. However, it is unclear whether such success transfers to computational scientific workflows, where tasks require not only strong coding ability, but also the ability to navigate complex, domain-specific procedures and to interpret results in the context of scientific claims. To address this question, we present AutoMat, a benchmark for evaluating LLM-based agents' ability to reproduce claims from computational materials science. AutoMat poses three interrelated challenges: recovering underspecified computational procedures, navigating specialized toolchains, and determining whether the resulting evidence supports a claim. By working closely with subject matter experts, we curate a set of claims from real materials science papers to test whether coding agents can recover and execute the end-to-end workflow needed to support (or undermine) such claims. We then evaluate multiple representative coding agent settings across several foundation models. Our results show that current LLM-based agents obtain low overall success rates on AutoMat, with the best-performing setting achieving a success rate of only 53%. Error analysis further reveals that agents perform worst when workflows must be reconstructed from paper text alone and that they fail primarily due to incomplete procedures, methodological deviations, and execution fragility. Taken together, these findings position AutoMat as both a benchmark for computational scientific reproducibility and a tool for diagnosing the current limitations of agentic systems in AI-for-science settings.

cs.CL↗

arXiv2026

Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge

Large language models are increasingly trained and deployed with long contexts that span documents, code repositories, and interaction histories. This scaling reflects the implicit assumption that training on longer contexts will only help the model by exposing it to richer evidence. We challenge this view by studying how the context window shapes a model's mode of learning, shifting it between parametric internalization and contextualization. We propose the Information Abundance Paradox, which hypothesizes that abundant relevant information in the training context can reduce the incentive to encode that information parametrically, thereby increasing reliance on context. In pretraining with long documents, increasing the context window improves language modeling, natural language understanding, and closed-book MCQA only up to an intermediate optimum, after which performance consistently declines. In supervised fine-tuning, more task-relevant train-time context improves performance with supporting context, but reduces robustness when context is absent or misleading at test time. Our analysis suggests that this behavior arises when longer context provides a lower complexity solution. Mechanistically, training with informative context shifts gradient pressure from feed-forward networks, often linked to parametric knowledge, toward attention modules, and causal interventions show that this shift increases reliance on context during inference. Overall, these findings support the Information Abundance Paradox and suggest that scaling toward near-infinite context is not simply a matter of supplying more data, even when high-quality long-context data is abundant.

cs.CL↗

arXiv2026

Process Supervision of Confidence Margin for Calibrated LLM Reasoning

Scaling test-time computation with reinforcement learning (RL) has emerged as a reliable path to improve large language models (LLM) reasoning ability. Yet, outcome-based reward often incentivizes models to be overconfident, leading to hallucinations, unreliable confidence-based control, and unnecessary compute allocation. We introduce Reinforcement Learning with Confidence Margin (RLCM), a calibration-aware RL framework that jointly optimizes correctness and confidence reliability via a margin-enhanced process reward over intermediate-budget completions. Rather than aligning confidence to correctness likelihoods, RLCM encourages the model to widen the confidence margin between correct and incorrect steps within a single reasoning trajectory. Across mathematical, code, logic and science benchmarks, our method substantially improves calibration while maintaining or improving accuracy. We further show that, with calibrated confidence signals, the resulting models enable more efficient conformal risk control and effective confidence-weighted aggregation

cs.LG↗

arXiv2026

ARIA: A Causal-Aware Framework for Rescuing LLM Reasoning in Trustworthy Materials Discovery

Generative models have revolutionized the process of materials discovery, yet they often fail to satisfy underlying physical causality. Through an analysis of Large Language Models (LLMs) augmented with knowledge graphs derived from current literature, we uncover a phenomenon termed contextual tunneling, where models "over-anchor" on narrow, retrieved evidence while suppressing global physical reasoning. To address this problem, we introduce ARIA, a causal-aware framework that conditions knowledge use on mechanistic completeness. ARIA routes each query through a three-tier cascade: (i) direct causal reasoning when complete evidence chains of Process-Structure-Property (PSP) are available, (ii) physics-informed analogical transfer for sparse or novel material systems, and (iii) explicit parametric fallback when external evidence is incomplete. As a proof of concept, we construct a Knowledge Graph (KG) containing 2,839 extracted PSP relations from peer-reviewed articles in the materials literature and evaluate ARIA on forward prediction and inverse design tasks for two-dimensional (2D) materials. ARIA mitigates contextual tunneling, improves over unaugmented and naive KG-augmented baselines, and provides further gains when an online literature search is used for evidence enrichment. Crucially, ARIA produces auditable causal traces, enabling physically grounded and trustworthy AI-assisted materials discovery.

cs.AI↗