Using a large language model to play the user is now standard in scalable evaluation. It has a repeatedly diagnosed failure: the simulated user is excessively cooperative, so a system under test can score by the sheer number of questions it asks rather than by making the user willing to speak. We answer with a disclosure gate conditioning information release on the companion agent's behaviour: its state is a ladder of five ordered gates, merged onto three observable depth layers. We specify, ablate, and audit it, and train a user simulator against that specification. Gating behaviour is learned from the training corpus's synthetic branch, while the real branch supplies how people speak and react; after training, the simulator need not be told at runtime which gate each item sits behind. The gate is a load-bearing component of the environment: on the English corpus of a published companion-agent benchmark (CompanionBench), once training no longer states per example which gate each item sits behind, the largest rank displacement across 12 systems under test exceeds the noise band set by re-running that environment under a new seed, while per-system scores show no detectable change. We state two acceptance criteria: a ranking must be order-preserving, and absolute scores must be scale-stable. Of the candidates we examine, only one passes both -- the simulator we release -- and its leaderboard correlates at 0.993 with the benchmark's original simulator. By contrast, prompting a frontier model as the simulator barely moves the ranking while shifting every score upward -- a shift invisible to anyone checking the ranking alone. The environment we specify is the one that benchmark already used. That publication describes the mechanism in about four hundred words, and we supply what it lacked: specification, ablations, human studies, negative controls, and downstream sensitivity analysis.
We present the zbMATH Open Knowledge Graph, a large-scale RDF knowledge graph (KG) covering more than 250 years of mathematical scholarship. Unlike existing scholarly knowledge graphs that primarily capture bibliographic metadata and citation structures, the zbMATH Open KG integrates expert-curated semantic content, including reviews, keywords, subject classifications, software references, and disambiguated authorship. This combination of domain-specific representation of mathematical knowledge and extensive temporal coverage supports analyses that require fine-grained exploration of mathematical concepts, research fields, and scholarly relationships over time. The resulting graph comprises 34 million entities and 168 million RDF triples represented using established Semantic Web vocabularies, supporting interoperability and FAIR data principles. We further demonstrate its capabilities through query-driven historically grounded scholarly exploration use cases, illustrating how the knowledge graph can surface relationships and patterns that may be difficult to identify from bibliographic and citation information alone. The zbMATH Open KG provides an open semantic infrastructure for studying the development of mathematical knowledge and tracing scholarly connections across centuries of scholarship.
Multi-turn tool calling is a core evaluation scenario for large language model (LLM) agents. On public tool-calling benchmarks, open-weight models now approach or even surpass closed-source frontier models in aggregate accuracy. However, this metric averages over many different multi-turn situations and obscures whether progress is balanced across them. We propose an action-class-oriented diagnostic framework that decomposes multi-turn failures into two orthogonal modes: action-class miscalibration and action-execution failure. The framework operates over a four-class action space (TOOL_CALL/ASK/REFUSE/CONFIRM) and introduces a self-revealing upper bound Acc <= GAR (Gold Action Recall); the two modes show up as bound violation (Acc > GAR, exposing state-grader masking of miscalibration) and large bound slack (GAR >> Acc, localizing execution failure within TOOL_CALL). We validate it on a panel of tool-calling models across multiple multi-turn benchmarks. Across our panel, the diagnostic reveals action-class miscalibration as a substantial failure mode the state grader cannot see. This gap inflates standing for heavily tool-trained families, which our diagnostic separates from families with context-appropriate action choice. Calibration is reshapable through context-only perturbations, but the reshape is heterogeneous: a single perturbation moves accuracy in opposite directions across families (up to +11.5 vs -21.0 pp on the same scenario), and its effect further depends on the perturbation mechanism. We argue that multi-turn tool-calling evaluations should supplement aggregate accuracy with action-class diagnostics that expose what the model actually does in each scenario.
Kangjia Zhao, Jiajun Li, Haozhan Shen et al.· 0 citations
Vision-language models (VLMs) have demonstrated strong performance in visual question answering with natural images. However, they continue to struggle with scientific diagrams, which are designed to convey functional or relational meaning rather than literal scenes. We therefore introduce a framework for generating large-scale diagram-grounded instruction data by leveraging terminology derived from scientific curricula. Our approach systematically extracts domain concepts, synthesizes atomic facts, retrieves relevant diagrams from the web, and generates multimodal supervision in the form of diagram captions and multiple-choice questions. Using this pipeline, we construct SciGram, a dataset of over 194K diagrams and 1.4M visual instructions across life, earth, and physical sciences. Despite relying on noisy web data and synthetic annotations, models fine-tuned on SciGram achieve substantial improvements on diagram-centric benchmarks, including TQA, ScienceQA, and AI2D, outperforming or matching state-of-the-art VLMs while using fewer training instances. Furthermore, augmenting existing models such as LLaVA OneVision with SciGram establishes new state-of-the-art performance on diagram question answering. Our results highlight the effectiveness of terminology-grounded instruction generation as a general strategy for improving vision-language reasoning in scientific domains. To support future research in scientific diagram understanding, we release both the SciGram dataset and models.
Raul Ortega, Jos\'e Manuel G\'omez-P\'erez· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Conditional independence tests (CITs) test for conditional dependence between two random objects $X$ and $Y$ given a third random object $Z$. Existing CITs have limited applicability to high-dimensional data, especially multimodal data like text. However, we show that such tests are of interest for large language model (LLM) outputs, where we test whether an output $X$ generated from a source text $Z$ carries information about an attribute $Y$ beyond $Z$ itself. For this purpose, we propose embedded CITs (eCITs), which embed $X$ and $Z$ and apply an existing CIT to the resulting representations and to $Y$. We show that, provided the embedding of $Z$ is sufficient, i.e. retains the information $Z$ carries about either $Y$ or the representation of $X$, the null hypothesis transfers from $X$ and $Z$ to their representations, so that a CIT valid for the embedded hypothesis is valid for the original one. We further give conditions for equivalence of the two hypotheses, and show that sufficiency weakens to mean sufficiency when the embedded test targets conditional mean independence. We propose a semi-synthetic simulation design to assess type I error (T1E) control and power of the eCITs for given embedding maps on a specific dataset and task, and use it to evaluate them on our application. Applying the eCITs to German Parliament speeches, we find for all combinations of embedding maps considered that the summaries of two LLMs contain information about the speaker's faction and gender beyond the speech they were generated from.
Marco Simnacher, Georg Keilbar, Benjamin König et al.· 0 citations
Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation. However, these agents suffer from severe overconfidence, making their expressed confidence unreliable for user trust and downstream abstention. To address this, we augment the Deep Research pipeline with step confidence elicitation after each retrieval, building on the commonly used post-answer verbalized confidence. Interestingly, we find that Evidence Confidence (E-Conf), elicited after the final retrieval step, provides a stronger uncertainty signal than Answer Confidence (A-Conf), elicited after answer generation, and that A-Conf is largely shaped by E-Conf. Based on these findings, we propose DualStake, a dual-path calibration method that applies margin-clipped, confidence-dependent stake rewards to jointly align E-Conf and A-Conf with answer correctness while limiting extreme confidence optimization. Experiments on Qwen2.5-7B, Qwen2.5-7B-Instruct, and Qwen3-4B across 8 QA benchmarks demonstrate that DualStake consistently improves calibration without sacrificing answer accuracy. The code is available at https://github.com/FloXXXt/DualStake.
Yinuo Xu, Yuwei Liang, Jianjie Cheng et al.· 0 citations
Language models can ignore prompt evidence when it conflicts with memorized knowledge. Post-training can make models follow such evidence more reliably, but it is unclear whether these gains require new machinery or strengthen machinery already present. We compare nine post-training arms spanning GRPO, SFT, and DPO from one starting checkpoint, with key comparisons extended across scales and families. We estimate a grounding direction from that checkpoint before training. Across five tested GRPO variants, grounding gains are small. For the two variants replicated across seeds, equivalence tests bound their effects below the conflict-SFT gain even as the rewarded metric improves. Conflict-SFT improves grounding moderately, while DPO drives grounding near ceiling on its matched distribution. Conflict-SFT and DPO largely use the same causal attention-head set as the starting model. Subtracting the starting-model direction suppresses both gains, while adding it to the starting model recovers 35% of DPO's gain at a dose passing all stated side-effect checks. After a supervised warm start makes the context answer appear in more rollouts, the same GRPO recipe adds essentially no further grounding gain. In our setting, grounding gains largely depend on machinery already present in the starting model.
Monocular videos record 3D scenes as sequences of 2D image-plane projections, obscuring depth and spatial relationships. Multi-object trackers localize and associate objects primarily using appearance and geometry observed only in the image plane, inheriting these ambiguities. To address this limitation, we introduce PLANET, an end-to-end multi-object tracker designed to move beyond the image plane. As an enabling step, we lift existing 2D tracking datasets into 3D. We then form world-grounded queries by embedding reconstructed 3D scene geometry into the features and positional encodings used during query formation. An auxiliary 3D location prediction task further encourages the queries to encode object positions during training. A complementary dual-resolution temporal memory preserves this evidence across longer temporal gaps. As a result, PLANET achieves state-of-the-art performance across three diverse benchmarks.
Orcun Cetintas, Guillem Brasó, Tim Meinhardt et al.· 0 citations
Obtaining labeled data for semantic segmentation in applied settings (e.g., autonomous driving, industrial waste sorting) is expensive and often infeasible at scale. We present a cross-modal pseudo-labeling pipeline that enables unsupervised domain adaptation without any target-domain annotations. The pipeline is built on two core foundation models: SAM generates class-agnostic region proposals, and EVA-CLIP assigns semantic labels based on region-text similarity, with confidence filtering ensuring that only reliable pseudo-labels are used for self-training a segmentation model. As an optional extension, BLIP provides language-grounded verification for ambiguous regions, thereby improving pseudo-label quality without altering the overall pipeline. Evaluated on two domain shifts, synthetic-to-real autonomous driving and, with a primary focus, lab-to-factory industrial waste sorting, the pipeline consistently improves over source-only baselines. Our results demonstrate that pseudo-label quality, not quantity, is a decisive factor in self-training under domain shift, and that cross-modal language grounding offers a practical path to reliable automatic annotation in deployment-critical applications.
Udo Schlegel, Shubhangi, Gabriel Dax et al.· 0 citations
The rapid advancement of vision-language models (VLMs) has accelerated progress in computational pathology; however, whole-slide image (WSI)-based pathology report generation remains limited by the scarcity of large-scale WSI--report datasets and the complexity of mapping spatially distributed visual patterns to structured clinical text. To address this, we introduce a clinically curated Pan-Asia WSI--report dataset of approximately 10,500 pairs from five institutions and establish the REG 2025 benchmark through a MICCAI challenge for systematic evaluation of multimodal models. We analyze submitted methods spanning pretrained VLMs, multiple-instance learning frameworks, hierarchical expert models, retrieval-augmented generation, and cross-modal Transformers. Rather than indicating that VLM use alone was sufficient for superior performance, the results suggest that top-performing methods benefited from structured report representations, hierarchical diagnostic decomposition, and effective multimodal grounding. We identify key limitations, including instability in quantitative attribute estimation (e.g., numeric hallucination) and a tendency toward diagnostic overspecification, with some errors resembling known diagnostic pitfalls in routine pathology. These findings establish REG 2025 as a benchmark for evaluating WSI-based structured report generation and vision-language understanding in computational pathology, providing insights for the design of clinically grounded multimodal pathology models.
Yu-Mi Lee, Harim Oh, Hyo-yun Kim et al.· 0 citations
Fault localization can focus a code model's repair on the statements a failing test implicates, but a targeted edit may succeed merely because it is small, and a second model call may succeed without using the failure at all. We separate these explanations with three arms applied to the same failed candidate: blind whole-solution resampling, spectrum-based localization followed by suspect-span infilling, and same-length infilling at a disjoint random code span. Across three frozen 26-32B models, three benchmarks and 488 failing candidates, plus a separately declared 24B fourth model from a third family, three results follow. First, localization is rarely available: only 9.0% of failing candidates expose a failing public test with a usable spectrum. Second, among the 177 candidates localizable from a strong suite, localized infilling loses decisively to blind resampling at a matched attempt count (3:40, p = 3.0 x 10^-9), opposite to our hypothesis; the loss replicates in a third family at -11.3 points (95% CI [-16.6, -6.8]), and widening the edit does not rescue it. Third, against the random-span placebo localized infilling leads pooled (11:1, Holm-adjusted p = .019), but that lead resolves in no individual model under the analysis our shipped plan designates primary (best Holm p = .087), so we report the location effect as suggestive rather than established. Re-pricing attempts as tokens narrows but does not overturn this: a span attempt spends 21.7 generated tokens against 371.1, yet 16 localized attempts reach 6.8% while one blind attempt already reaches 10.1%. Infilling reproduces the removed span verbatim in 48.9% of attempts, which is why more budget does not help. We restrict every localization conclusion to the 24-32B models tested.
InfRared Small Target Detection (IRSTD) is a challenging task. Relying solely on pixel-level information, vision-only methods struggle to distinguish targets from clutter. Current multimodal methods typically describe both targets and backgrounds with a single textual prompt. Such an approach lacks dedicated regional guidance and ignores infrared semantic asymmetry. Consequently, it provides insufficient background suppression information and introduces severe feature optimization conflicts, overwhelming small targets with noise. To address these issues, we propose a novel Asymmetric Dual-text Guided Network (ADGNet). Specifically, accounting for the infrared semantic asymmetry, we first design the Asymmetric Dual-text Prompt (ADP), comprising an image-agnostic abstract target prompt and an image-specific detailed background prompt. To leverage these prompts, we introduce an Asymmetric Dual-Branch Interaction (ADBI) module to separately guide visual features with their respective text priors, protecting targets from noise while suppressing background clutter. Subsequently, we introduce an Adaptive Feature Aggregation (AFA) module to dynamically fuse features from the two branches. Furthermore, we construct a multimodal Asymmetric Image-Text Infrared (AITIR) dataset by providing asymmetric text annotations for three public datasets (IRSTD-1K, NUDT-SIRST, and SIRST). Extensive experiments demonstrate that ADGNet outperforms 21 state-of-the-art (SOTA) methods. Code is available at https://github.com/iLearn-Lab/MM26-ADGNet.
Tongtong Wang, Mingzhu Xu, Chenglong Yu et al.· 0 citations
