Tag
Omni-Woodpecker提出一种免训练框架，通过证据-先验重平衡来纠正Omni多模态大模型中的跨模态幻觉，在CMM和AVHBench等基准上显著提升性能。
提出SOAR（Source-Aware Reinforcement Learning），通过将外部证据与内部知识解耦为可观测的并行流并引入Source-Consistent Reward，实现搜索代理的精确信用分配；在3B和7B模型上分别提升7.3%和5.4%，并将证据伪造率降低74.8%。
HARPO is a reinforcement learning framework from Huawei Technologies that jointly optimizes hallucination mitigation and creative writing via a Hallucination-Aware Generative Reward Model, a selective activation mechanism, and a data curriculum, reducing hallucination rates while boosting creative-writing quality on Qwen models.
该论文揭示了音频-视觉大语言模型中"源混淆接地幻觉"的内在机制——问题中继（question-relay）会携带干扰模态的线索，并提出训练无关的SECRET方法，在保持生成位置通路的同时将问题状态导向所需的源证据，在CMM和AVHBench基准上相比基线模型最高提升18.0和7.1个百分点。
AFDBench is a benchmark for evaluating generative meteorological reasoning in National Weather Service forecast discussions, using reinforcement learning to improve LLM performance on accuracy, style adherence, and data fidelity.
This survey paper reviews decoding methods for large language models and vision-language models, focusing on inference-time approaches to enhance efficiency and control generation. It identifies paradigms, highlights challenges, and discusses future research directions.
CLAIR-Fin is an adversarial multi-agent framework designed to enhance faithfulness in cross-modal financial question answering by decomposing questions into atomic claims and using adaptive debate and verification to reduce hallucinations.
This paper investigates whether layer-wise differences exist in Mixture-of-Experts (MoE) models and proposes EAACD, an expert-aware adaptive contrast decoding method that leverages expert activation patterns in higher layers to reduce hallucinations in LLMs for QA tasks.
This paper introduces SSC-GRPO, a step-level reward optimization method that uses self-consistency scores across multiple rollouts to reduce context-sensitive factual hallucinations in LLM reasoning, achieving state-of-the-art performance on math reasoning and hallucination benchmarks.
Hallo4D is a model-agnostic framework that leverages large multimodal language models to detect and correct spatial and temporal hallucinations in 3D and 4D generation, improving consistency across viewpoints and time without requiring retraining.
This paper proposes VaseMuseum, a lightweight multimodal agent framework that combines 3D digitization with vision-language models to create an interactive digital museum for ancient Greek pottery, addressing challenges of evidence grounding and hallucination through source-level and response-level reliability control.
This paper studies a lightweight prompting-based recovery approach for LLM dialogue agents when backend database calls fail, showing that the Guided-Retry strategy reduces hallucination by 50% on MultiWOZ and 42% on SGD across six model families.
This paper reveals that hallucination in large vision-language models is caused by a dynamic structural misalignment where certain attention heads act as risky mediators, decoupling from visual evidence to lock onto language priors. The authors propose Fox, a training-free causal intervention framework that diagnoses and physically severs these pathological shortcuts, achieving state-of-the-art performance in faithful decoding.
CaVe-VLM-CoT is a modular reflection-based agentic-RAG framework for vision-language models that enforces evidence-grounded reasoning through a five-stage pipeline, achieving 87.1% accuracy on ScienceQA and proposing a suite of 23 metrics for evaluation.
Introduces MODE-RAG, a multi-agent system using Variational Free Energy and Monte Carlo Tree Search to dynamically gate interventions for mitigating hallucinations in Multimodal Retrieval-Augmented Generation systems, along with the ModeVent evaluation dataset.
This paper proposes a multi-agent 'Trust but Verify' system to reduce medical hallucinations in LLMs. It tests three open-access models on clinical questions about banned drugs and achieves a 53% reduction in hallucination error rate.
Discusses various methods to optimize DiffusionGemma inference, reduce hallucination, and improve performance for tool use and agents, including entropy-bounded sampling, schema scaffolding, and retrieval during denoising.
This paper proposes NTS-CoT, a novel framework that uses Chain-of-Thought reasoning to mitigate hallucinations in LLM-based news timeline summarization. It introduces three modules—Element-CoT, Date Selection, and Causal-CoT—to improve faithfulness and reduce omissions, outperforming state-of-the-art baselines on three benchmarks.
This paper introduces MGAP, a training-free decoding method that reduces hallucinations in Multimodal Large Language Models by adaptively suppressing only the harmful parts of language priors while preserving the model's semantic manifold. The method outperforms prior baselines on POPE and CHAIR benchmarks.
This paper demonstrates that Whisper's hallucination failures on silence, noise, or music can be detected and mitigated purely from internal activations using sparse autoencoders, achieving large reductions in hallucination rate without fine-tuning.