factuality

该论文评估了单一LLM基线与三种多智能体拓扑在临床任务中的表现，发现协作能提升任务指标，但并不必然降低幻觉，且模型本身而非拓扑结构决定了幻觉的变化方向。

该论文对八个闭源和开源大模型在英文与孟加拉语XL-Sum测试集上的零样本抽象摘要能力进行了标准化评测，结合词法、语义指标、事实性验证和LLM-as-a-Judge评估，发现英文表现优异并不自动迁移到低资源语言，且不同评估方法产生不同的模型排名。

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。

This arXiv paper introduces a pre-registered evaluation measuring cross-jurisdiction substitution in LLM recall of U.S. state Medicaid policy facts, finding that Claude Sonnet 5.5 and GPT-5.6 Sol reproducibly return real values belonging to other states, while showing that naive attribution inflates such errors severalfold unless asked-state reference values are fully checked.

This paper applies split-conformal claim filtering to multi-hop RAG, showing that at a 95% conformal target, the fraction of responses with fully supported claims rises to 95.80%-97.20% across HotpotQA, Natural Questions, and TriviaQA using Llama 3.1 8B and GPT-4o-mini — but with strong abstention and low claim retention, highlighting that reliability gains trade off against answer coverage.

This study evaluates Jev, a System One model, for measuring factual differences in AI-generated radiology reports, demonstrating its effectiveness and efficiency compared to other evaluation methods.

Introducing GPT-6 Sol and Luna models with stronger coding, computer use, improved factuality, and alignment, outperforming Claude on AutomationBench.

The paper proposes Style-Debiased DPO (SD-DPO), a method for factuality-aware synthetic preference data to improve knowledge elicitation in large language models, addressing issues where standard DPO may incorrectly penalize correct responses due to stylistic differences.

This paper introduces TKFQA, a counterfactual benchmark of 10,130 QA pairs over tables, texts, and knowledge graphs for evaluating LLM factuality consistency and order-robust reasoning, and proposes ORLF, a training framework that improves reasoning-chain accuracy and reduces input-order sensitivity.

Proposes three attention-guided strategies for layer selection in contrastive decoding for large language models, improving factuality on TruthfulQA over the DoLa baseline.

This paper proposes a novel online reinforcement learning method to improve factuality in reasoning LLMs by designing a reward function that balances factual precision, detail, and relevance, achieving a 23.1 percentage point reduction in hallucination rate on six benchmarks.

This paper investigates whether neural network probes that predict correctness of language model outputs actually capture objective correctness or the model's own self-judgement, using conflict cases where the two disagree. The authors find that transferable directions predominantly preserve self-judgement polarity, challenging the interpretation of correctness readouts.

Arena.ai has added Factuality to model rankings, supporting weighting of human preference and factuality, and showing changes in model rankings.

This paper proposes a semantic framework to describe AI systems, distinguishing justified claims from misleading outputs, and defines common failures such as hallucination and unsupported assertions.

This paper systematically examines cross-lingual disparities in LLM-based healthcare question answering across five languages, finding significant gaps in factual alignment and proposing the MultiWikiHealthCare dataset.

Introduces MARGO, a reinforcement learning framework that uses mixed-mode advantage regularization to mitigate thinking-induced hallucinations in large reasoning models by comparing thinking and non-thinking trajectories.

This paper proposes D2R-RAG, a model-agnostic and resource-aware framework that diagnoses and repairs factual errors in RAG systems under latency and VRAM constraints, achieving better accuracy-efficiency trade-offs on FEVER and HotpotQA.

ConflictScore is a new metric that quantifies how well language models acknowledge conflicting evidence in their grounding documents, decomposing responses into atomic claims and measuring conflict balance. The paper also introduces ConflictBench, a benchmark covering diverse conflict forms, and shows the metric can improve truthfulness on TruthfulQA.

Report claims that GPT-5.5 Instant shows significant improvements in factual accuracy, particularly in high-stakes fields like medicine, law, and finance.

MoshiRAG combines a compact full-duplex speech language model with asynchronous retrieval-augmented generation to improve factuality while maintaining real-time interactivity. The approach leverages natural temporal gaps in conversation to retrieve external knowledge without disrupting the natural flow of dialogue.