Tag
Microsoft AI introduces TalentRL, a strategy-guided reinforcement learning framework for hierarchical routing over procedural memory in LLM agents, where coarse-grained 'talents' guide selection of fine-grained skills. It improves success rates across seven QA benchmarks and achieves 99.3% on ALFWorld and 90.7 on WebShop across Qwen and Llama backbones.
提出 CLIK，一个上下文条件化的参数生成器，通过单次前向传播将上下文映射为 LoRA 参数，把知识内化到模型权重中，无需测试时梯度更新，在文档问答和多轮对话上优于现有超网络方法。
本文提出了一种基于标签保持扰动的输入歧义检测框架，先检测再消歧，包含黑盒规则库与白盒重要性引导两种检测器，在四类问答数据集上优于输出侧不确定性基线与LLM改写扰动基线。
提出证据状态校准（evidence-state calibration）概念与SuffiCE方法，通过微调模型下一token分布中的置信度，使其随证据变化而准确反映答对概率，在临床诊断和多跳问答上将Brier分数降低最高60%、AUROC提升最高0.16。
GEMS-RAG 是一个面向工程标准文档的图驱动多模态检索增强生成框架，通过层级解析、视觉实体抽取、多模态知识图谱构建与图驱动混合检索，在 MUTCD 数据集上相较文本、多模态和图基 RAG 基线提升了检索与问答效果。
MoveBench introduces the first rubric-based benchmark for evaluating large language models in sports and health coaching, covering safety, scientific validity, and personalization across multiple-choice and open-ended tasks. An evaluation of 15 LLMs on 809 questions with 5,012 rubric criteria reveals significant gaps in safety awareness and personalized program design.
论文提出联合约束聚合（JCA）方法，通过将问题要求表示为共享约束并逐一核对候选答案，缓解检索偏向导致的限定词敏感问答失败，在三种评测设置和多个指令微调模型上提升准确率。
This paper introduces evidence-based generation (EBG) with a joint Ans+Evid metric and reformed benchmarks (MultiHop-RAG, QASPER-MP) to show that RAG's similarity-driven chunked retrieval—not reading—is the key bottleneck for long-context multi-hop question answering across frontier and open-weight LLMs.
The author explores how RAG evaluations should distinguish three different reasons a system answers "I don't know" — retrieval failure, missing evidence, and conflicting sources — and proposes evals that label the expected unresolved state rather than just scoring answers as correct, incorrect, or refused.
提出一个无奖励、无裁判、无答案键的自改进代理框架，通过双层循环在小型精确DSL上进行算子归纳，实现精确问答能力的自我扩展，并在多个领域验证其可靠性与跨模型迁移能力。
该研究通过交叉验证证据与推理轨迹，探讨答案配置（续接自身思考 vs. 独立答案调用）如何影响模型对证据的遵循。结果表明独立调用在证据遵循上比续接高出29.9至60.6个百分点，但两者在去除轨迹后的准确率下降幅度相近，说明从轨迹中获益与顺从轨迹是两种不同属性。
该论文提出了 Knowledge Transfer Probing (KTP) 协议，研究通过图像生成训练是否能扩展预训练统一多模态模型的知识，并引入文本对齐与图文交错对齐方法以提升文本端知识访问能力。
JevGPT: Thinking Fast and Slow in Autonomous Research 提出了一种用于自主研究的双系统架构，将快速的证据处理（系统一）与审慎的研究规划（系统二）耦合，在固定语料问答和文档排序任务上实现了领先的性能，同时比同类研究代理更快、成本更低。
提出CofFeQA，一种基于反事实样本的少样本知识蒸馏方法，用于开放域问答任务；在TriviaQA等基准上，相比仅用事实样本的蒸馏基线在T5设置下以32个少样本对提升约8.77 EM和9.98 F1分。
Researchers introduce GeopoliticalQA, a benchmark of 5,525 QA pairs over 415 geopolitical and strategic intelligence reports, evaluating 15 retrieval systems and six open-weight LLMs. They find mid-size encoder bge-m3 matches or beats much larger and commercial embedding models at chunk-level retrieval, while answer generation remains hard even with gold evidence for inferential and risk-oriented questions.
ArcticQA/ArcticAbstain introduces a dataset and paired benchmark evaluating whether LLMs appropriately abstain on Arctic science multiple-choice questions, finding large baseline differences and only modest increases in abstention when correct answers are removed.
The paper proposes a pipeline that automatically generates ontology-grounded, reasoner-verified multiple-choice benchmarks for evaluating LLM logical reasoning in scientific AI, validated on three ontologies (Pizza, PMDco, DOID) across six LLMs.
A message-intervention study showing that rationales passed between reasoner and verifier LLMs barely improve answer accuracy but substantially alter support judgments, especially when corrupted — suggesting rationale sharing should be evaluated as a verification mechanism rather than an accuracy booster.
该论文研究了在单语检索增强生成（RAG）中推理语言与查询/检索文档语言对齐的作用，构建了一个基于虚构桌游世界《The DarkEye》的德语 RAG 问答测试集，发现对齐推理语言能提升表现，但仅能达到模型原生英语推理的水平。
HalluScoring 2026 presents the first shared task on LLM hallucination detection and factual verification for Arabic question answering, with 13 teams competing across four subtasks covering binary hallucination detection and answer selection on the HalluScore and HalluTruthQA datasets.