evidence-grounding

Open-o3-Video-v2 将视频推理建模为统一的智能体证据接地过程，模型主动识别缺失信息、获取额外观测并将其时空证据写入推理轨迹。基于 Qwen3-VL-4B 骨干，该模型在 V-STaR 上取得 40.4% mAM 和 56.3% mLGM，显著超越基线。

该论文提出Image Factuality Assessment任务及EviScene多模态ReAct智能体，通过视觉检查、外部检索与证据分析的交替推理来判定图像所描绘场景是否事实成立，并发布了包含10,174张图像的EviLens专家审阅基准，在Qwen3.5-9B上实现了79.05%的平衡准确率。

这篇论文提出一个用于可解释自杀风险评估的框架，包含风险评估、证据定位和心理社会因素识别三个模块，并在 IEEE BigData 2026 Explainable Suicide Risk Detection Challenge 中报告了相应的 F1 结果。

The paper introduces Nautil, a 731-case audited dataset and evaluation suite for teaching LLM investigators when evidence is sufficient to close an investigation, showing fine-tuning and RL on closure decisions sharply reduces evidence overstatement even for frontier models.

TRACE introduces target-aware retrieval, attributed evidence localization, and contract-constrained extraction to close the 'grounding contract gap' in scientific LitTraceQA, scoring 0.760613 on the official 71-question test set with strong paper F1 (0.9728) but weaker table extraction performance.

PathAnchor 是一个基于路径结构证据工作区的科学推理系统，通过 Material–Sensor–Signal–System 轨迹组织证据，在 120 个柔性传感器问题上以 82.6% 得分领先六个基线系统，并显著提升来源召回率与引用完整性。

BioPhys-Bridge introduces a benchmark dataset for evaluating language models on evidence-grounded scientific reasoning in biophysics, focusing on interdisciplinary tasks with quantitative and mechanistic grounding.

The paper examines correctness-gated multi-teacher distillation, finding decision shifts, lost label functionality, and an inconclusive grounding audit, with no incremental benefit over hard filtering.

The paper introduces PACEShop, a benchmark dataset, and PACEJudge, a judging protocol for evaluating personalized, actionable, compositional, and evidence-grounded shopping assistants, addressing gaps in current evaluation methods.

Presents LitTraceQA, a benchmark for scientific question answering that requires systems to retrieve relevant papers, locate supporting evidence, and produce verified answers in multiple formats.

Presents an evidence-grounded pipeline for constraint checking in construction documents, evaluating resolution-breadth tradeoffs in PDF evidence allocation on expert-referenced tasks.

EarlyDx is a new large-scale benchmark for evaluating LLMs on open-ended, evidence-supported diagnosis generation at emergency department admission, built from 154,834 MIMIC-IV encounters. It reveals that even frontier and medical-specialized models struggle to synthesize admission-time evidence, with post-training only partially improving inference-dependent recall.

Scope3Trace proposes an evidence-grounded information extraction framework that leverages large language models to identify and extract Scope 3 greenhouse gas emissions from sustainability reports, contributing a dual-level multimodal dataset and achieving high extraction accuracy.

ResearchStudio-Idea is a skill suite that combines literature search, novelty checking, and pattern-guided generation to produce traceable research proposals, built from analysis of 1,947 machine learning conference papers.

Data2Story is a multi-agent framework that transforms datasets into verifiable multimodal stories, acting as an end-to-end data journalist with evidence grounding and multimodal generation.

Introduces EHRNote-ChatQA, a benchmark for evidence-grounded multi-turn clinical question answering over multiple discharge summaries, constructed with expert validation. Benchmarking 22 LLMs reveals challenges in evidence grounding and multi-turn error accumulation.

This paper introduces FullCite, a framework for generating structured inline citations that link each claim to both its source document and specific evidence spans. Evaluated on three QA benchmarks (ASQA, BioASQ, ExpertQA), it finds that while LLMs are good at document-level attribution, they struggle with precise evidence span identification.

MedicalBench is a new benchmark for evaluating large language models on medical concept extraction from electronic health records, focusing on implicit reasoning and evidence grounding. It includes 823 expert-annotated examples and shows that current models perform modestly, highlighting the difficulty of extracting implicitly stated medical concepts.

This paper proposes a framework for parallel chunk-level processing of long documents with LLMs to reduce cumulative bias and improve evidence traceability, achieving significant reductions in omission errors and unsupported claims.