Vision-centric retrieval for VQA requires retrieving images to supply missing visual cues and integrating them into the reasoning process. However, selecting the right images and integrating them effectively into the model's reasoning remains challenging. To address this challenge, we propose R3G, a modular Reasoning-Retrieval-Reranking framework. It first produces a brief reasoning plan that specifies the required visual cues, then adopts a two-stage strategy, with coarse retrieval followed by fine-grained reranking, to select evidence images. On MRAG-Bench, R3G improves accuracy across six MLLM backbones and nine sub-scenarios, achieving state-of-the-art overall performance. Ablations show that sufficiency-aware reranking and reasoning steps are complementary, helping the model both choose the right images and use them well. We release code and data at https://github.com/czh24/R3G.
Zhuohong Chen, Zhengxian Wu, Zirui Liao et al.· 0 citations
Training robust visual surveillance models requires large-scale datasets with precise spatial annotations, yet collecting real surveillance data is costly, privacy-sensitive, and often legally constrained. Synthetic data generation offers a compelling alternative, but existing methods lack fine-grained 3D control over object placement and appearance, limiting geometric consistency across camera viewpoints. We introduce POCI-Diff (Positioning Objects Consistently and Interactively), a framework that generates annotated synthetic scenes from explicit 3D bounding-box layouts with per-object semantic control. By integrating Blended Latent Diffusion with depth-conditioned ControlNet, POCI-Diff synthesises complex multi-object scenes in a single forward pass, binding individual text descriptions to specific 3D locations. We further propose a warping-free editing pipeline supporting object insertion, removal, and transformation via regeneration, enabling efficient scene variation for data augmentation. Object identity across edits is preserved by conditioning on reference images via IP-Adapter, ensuring appearance consistency throughout interactive scene manipulation. Experiments show that POCI-Diff outperforms state-of-the-art 3D layout-guided generation methods in visual fidelity and layout adherence, while eliminating warping-induced geometric artifacts.
Andrea Rigo, Luca Stornaiuolo, Weijie Wang et al.· 0 citations
AI agents are autonomous systems that combine LLMs with external tools to solve complex tasks. While such tools extend capability, improper tool permissions introduce security risks such as indirect prompt injection and tool misuse. We characterize these failures as unbalanced tool-driven agency. Agents may retain unnecessary permissions (excessive agency) or fail to invoke required tools (insufficient agency), amplifying the attack surface and reducing performance. We introduce AgenTRIM, a framework for detecting and mitigating tool-driven agency risks without altering an agent's internal reasoning. AgenTRIM addresses these risks through complementary offline and online phases. Offline, AgenTRIM reconstructs and verifies the agent's tool interface from code and execution traces. At runtime, it enforces per-step least-privilege tool access through adaptive filtering and status-aware validation of tool calls. Evaluating on the AgentDojo benchmark, AgenTRIM substantially reduces attack success while maintaining high task performance. Additional experiments show robustness to description-based attacks and effective enforcement of explicit safety policies. Together, these results show that AgenTRIM provides a practical, capability-preserving approach to safer tool use in LLM-based agents.
Roy Betser, Amit Giloni, Shamik Bose et al.· 0 citations
The proliferation of Large Language Models (LLMs) necessitates valid evaluation methods to provide guidance for both downstream applications and actionable future improvements. The Item Response Theory (IRT) model with Computerized Adaptive Testing has recently emerged as a promising framework for evaluating LLMs via their response accuracy. Beyond simple response accuracy, LLMs' chain of thought (CoT) lengths serve as a vital indicator of their reasoning ability. To leverage the CoT length information to assist LLM evaluation, we propose the \textbf{La}tency-\textbf{R}esponse \textbf{T}heory (LaRT) model, which jointly models both the response accuracy and CoT length by introducing a key correlation parameter between the latent ability and the latent speed. We derive an efficient stochastic approximation Expectation-Maximization algorithm for parameter estimation. We establish rigorous identifiability results for the latent ability and latent speed parameters to ensure the statistical validity of their estimation. Through both theoretical asymptotic analyses and simulation studies, we demonstrate LaRT's advantages over IRT in terms of superior estimation accuracy and shorter confidence intervals for latent trait estimation. To evaluate LaRT in real data, we collect responses from diverse LLMs on popular benchmark datasets. We find that LaRT yields different LLM rankings than IRT and outperforms IRT across multiple key evaluation metrics including predictive power, item efficiency, ranking validity, and LLM evaluation efficiency. Code and data are available at https://github.com/Toby-X/Latency-Response-Theory-Model
Zhiyu Xu, Jia Liu, Yixin Wang et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
In web environments, user preferences are often refined progressively as users move from browsing broad categories to exploring specific items. However, existing generative recommenders overlook this natural refinement process. Generative recommendation formulates next-item prediction as autoregressive generation over tokenized user histories, where each item is represented as a sequence of discrete tokens. Prior models typically fuse heterogeneous attributes such as ID, category, title, and description into a single embedding before quantization, which flattens the inherent semantic hierarchy of items and fails to capture the gradual evolution of user intent during web interactions. To address this limitation, we propose CoFiRec, a novel generative recommendation framework that explicitly incorporates the Coarse-to-Fine nature of item semantics into the tokenization process. Instead of compressing all attributes into a single latent space, CoFiRec decomposes item information into multiple semantic levels, ranging from high-level categories to detailed descriptions and collaborative filtering signals. Based on this design, we introduce the CoFiRec Tokenizer, which tokenizes each level independently while preserving structural order. During autoregressive decoding, the language model is instructed to generate item tokens from coarse to fine, progressively modeling user intent from general interests to specific item-level interests. Experiments across multiple public benchmarks and backbones demonstrate that CoFiRec outperforms existing methods, offering a new perspective for generative recommendation. Theoretically, we prove that structured tokenization leads to lower dissimilarity between generated and ground truth items, supporting its effectiveness in generative recommendation. Our code is available at https://github.com/YennNing/CoFiRec.
Tianxin Wei, Xuying Ning, Xuxing Chen et al.· 0 citations
Recent advancements in multimodal large language models (MLLMs) have shown strong understanding of driving scenes, drawing interest in their application to autonomous driving. However, high-level reasoning in safety-critical scenarios, where avoiding one traffic risk can create another, remains a major challenge. Such reasoning is often infeasible with only a single front view and requires a comprehensive view of the environment, which we achieve through multi-view inputs. We define Safety-Critical Reasoning as a new task that leverages multi-view inputs to address this challenge. Then, we distill Safety-Critical Reasoning into two stages: first resolve the immediate risk, then mitigate the decision-induced downstream risks. To support this, we introduce WaymoQA, a dataset of 35,000 human-annotated question-answer pairs covering complex, high-risk driving scenarios. The dataset includes multiple-choice and open-ended formats across both image and video modalities. Experiments reveal that existing MLLMs underperform in safety-critical scenarios compared to normal scenes, but fine-tuning with WaymoQA significantly improves their reasoning ability, highlighting the effectiveness of our dataset in developing safer and more reasoning-capable driving agents. Our code and data are provided in https://github.com/sjyu001/WaymoQA
Seungjun Yu, Seonho Lee, Namho Kim et al.· 0 citations
Current vision-language models (VLMs) in medicine are primarily designed for categorical question answering (e.g., "Is this normal or abnormal?") or qualitative descriptive tasks. However, clinical decision-making often relies on quantitative assessments, such as measuring the size of a tumor or the angle of a joint, from which clinicians draw their own diagnostic conclusions. This quantitative reasoning capability remains underexplored and poorly supported in existing VLMs. In this work, we introduce MedVision, a large-scale dataset and benchmark specifically designed to evaluate and improve VLMs on quantitative medical image analysis. MedVision spans 22 public datasets covering diverse anatomies and modalities, with 29.0K 3D images, 11.2M annotated 2D slices, and 24.3M single-instance annotations. We focus on three representative quantitative tasks: (1) detection of anatomical structures and abnormalities, (2) tumor/lesion (T/L) size estimation, and (3) angle/distance (A/D) measurement. We show that current off-the-shelf VLMs perform poorly on these tasks. However, supervised and reinforcement fine-tuning (RFT) on MedVision significantly enhances performance across detection, T/L size estimation, and A/D measurement, yielding MedVision-V0 as a strong open baseline. In the RFT stage, we design and evaluate the efficacy of process rewards, multiplicative reward composition, and multi-task RFT with curriculum learning. MedVision provides a foundation for developing VLMs with robust quantitative reasoning capabilities in medical imaging.
Yongcheng Yao, Yongshuo Zong, Raman Dutt et al.· 0 citations
This paper examines the role of public interest litigation in promoting accountability for AI and automated decision-making (ADM) in Australia. Since ADM regulation faces political and geopolitical headwinds, effective governance will have to rely on the enforcement of existing laws. Drawing on interviews with Australian public interest litigators, technology policy activists, and technology law scholars, the paper positions public interest litigation as part of a larger ecosystem for transparency, accountability and justice with respect to ADM. The paper explores the tactics and strategies of what one participant described as 'retrofitting' old laws to ADM. These go beyond creative legal argumentation, to encompass practices of community-building, collaboration on theories of change, canny selection of clients and causes of action, and aligning the interests of stakeholders in litigation. Naturally, the paper also contends with the limits of these strategies, and of the Australian legal system. Where limits are capable of being overcome, the paper presents findings on urgent needs: the enabling institutional arrangements without which effective litigation and accountability will falter. The paper is relevant to law and technology scholars, individuals and groups harmed by ADM, public interest litigators and technology lawyers, civil society and advocacy organisations, and policymakers.
Wide access to advanced experimental methods in materials science has given rise to an abundance of procedural knowledge, which is scattered across decades of scientific literature and recorded in unstructured formats that are challenging to analyze systematically. In this work, we present LeMat-Synth Parser, a modular, open-source, and multi-modal extraction toolbox that utilizes large language models (LLMs) and vision language models (VLMs) to automatically structure synthesis protocols and performance metrics extracted from both text and figures of publications. Applying LeMat-Synth Parser to 81K open-access publications, we curate LeMat-Synth, an extensive dataset of 58K synthesis procedures and to our knowledge the largest and most diverse structured inorganic materials synthesis dataset to date, covering 35 synthesis methods and 16 material classes based on a domain-specific ontology. We validate extraction quality against annotations by domain experts and a scalable LLM-as-a-judge framework, and benchmark a suite of models to identify optimal configurations and characterize cross-model biases. To demonstrate the extensibility of LeMat-Synth Parser, we apply it to two distinct domains. First, we link synthesis protocols and catalyst identity to thermocatalytic performance across a corpus of ammonia-decomposition publications. Second, we cross-validate text- and figure-reported critical transition temperatures across 1,384 superconductivity papers, then use the validated pipeline to recover the critical transition temperature for every composition in a sample series. We release LeMat-Synth Parser and the LeMat-Synth dataset openly on GitHub and Hugging Face
Magdalena Lederbauer, Siddharth Betala, Valerie Gentzke et al.· 0 citations
Code-capable large language model (LLM) agents are embedded in software engineering workflows where they can read, write, and execute code, raising "jailbreak" stakes beyond text-only settings. Prior evaluations emphasize refusal or harmful-text detection, leaving open whether agents compile and run malicious programs. We present JAWS-Bench (Jailbreaks Across WorkSpaces), a benchmark spanning three escalating workspace regimes mirroring attacker capability: empty (JAWS-0), single-file (JAWS-1), and multi-file (JAWS-M). We pair it with a hierarchical, executable-aware Judge Framework that tests (i) compliance, (ii) attack success, (iii) syntactic correctness, and (iv) runtime executability to measure deployable harm. Across seven LLM backends from five families, prompt-only attacks in JAWS-0 achieve 61% compliance; 58% are harmful, 52% parse, and 27% run end-to-end. In JAWS-1, compliance reaches ~100% for stronger models with a mean ASR (Attack Success Rate) ~71%; JAWS-M raises mean ASR to ~75%, with 32% runnable attack code. Wrapping an LLM in an agent increases ASR by 1.6$\times$, by overturning initial refusals during planning and tool use. Additional evaluations with SWE-Agent and OpenAI Codex exhibit similar trends, indicating that JAWS-Bench can be reused across multiple agent frameworks. Category analyses identify which attack classes are most vulnerable and deployable, motivating execution-aware defenses and refusal-preserving agent designs.
Shoumik Saha, Jifan Chen, Sam Mayers et al.· 0 citations
The systematic assessment of AI systems is increasingly vital as these technologies enter high-stakes domains. To address this, the EU's Artificial Intelligence Act introduces AI Regulatory Sandboxes (AIRS): supervised environments where AI systems can be tested under the oversight of Competent Authorities (CAs), balancing innovation with compliance, particularly for startups and SMEs. Yet significant challenges remain: assessment methods are fragmented, tests lack standardisation, and feedback loops between developers and regulators are weak. This paper operationalises the AIRS lifecycle. We map the sandbox journey into 29 concrete activities, from pre-participation guidance through application, preparation, participation, exit, and post-participation monitoring, and we distinguish between a Core AIRS centred on regulatory oversight and an Extended AIRS that additionally embeds structured technical testing through an AI Technical Sandbox (AITS). From this mapping we derive 15 infrastructural and governance requirements that an AITS must satisfy, each linked to the activities it supports and, for high-risk systems, to the provider obligations set out in Articles 9-15 of the AI Act. The framework aims to address multiple stakeholders: CAs gain structured workflows for applying legal obligations; technical experts can integrate robust evaluation methods; and AI providers access a transparent pathway to compliance. We conclude by outlining the Sandbox Configurator, an open-source framework intended to instantiate AITS environments from these requirements, and by discussing how a shared technical foundation can support a scalable and innovation-friendly European infrastructure for trustworthy AI governance.
Alessio Buscemi, Thibault Simonetto, Daniele Pagani et al.· 0 citations
Jailbreak attacks bypass the safety alignment of large language models (LLMs) to elicit harmful outputs, yet the vast parameter space makes diagnosing the underlying failure mechanisms extremely challenging. We present NeuroBreak, a visual analytics system that helps experts progressively unpack jailbreak mechanisms from layer-level semantics down to neuron-level behaviors. A layer-wise probing pipeline traces how harmful representations evolve across layers, while a dual-dimensional character--behavior categorization reveals each safety-related neuron's inherent tendency and contextual contribution. These analyses are made interpretable through tailored visualization designs: a task-driven probing projection that reveals safety decision boundaries, a dual-stream semantic evolution flow that traces cross-layer semantic shifts, and a character--behavior chord graph that unifies neuron roles, attribution scores, and collaborative relations in a single view with in-situ causal verification. Quantitative evaluations and case studies show that NeuroBreak uncovers safety failure causes and provides actionable insights for strengthening LLM defenses.
Chuhan Zhang, Ye Zhang, Bowen Shi et al.· 0 citations
