multi-hop-qa

DocSkill提出一种带技能编排策略的智能体框架，用于多语言多跳文档问答，通过MCTS离线探索轨迹并结合SFT/DPO训练策略，同时发布包含60K QA对、覆盖10种语言的MMHop-DocBench数据集。

该论文提出两阶段诊断评估方法，检验以图结构表示的LLM推理链的结构可靠性，发现图的可靠性随拓扑、提示设计和模型族而变化，且图结构对不确定性估计的校准和节点级诊断有帮助，但作为提示上下文时收益有限。

Nebius AI 提出 Caddie 方法，通过强化学习从任务成败结果训练 LLM agent 的自然语言批评器（critic），无需逐步标注。Qwen3-4B critic 在 MuSiQue 上将成功率提升逾 25 个百分点，并可跨基座模型与任务域泛化。

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.

CANOPY 提出一种自适应粒度的多模态 RAG 检索后证据压缩框架，通过层级表征与父节点相对精炼选择多粒度证据区域，并在压缩不足时触发定向再检索，在 5 个 QA 基准上取得更高平均准确率，同时将读取器输入 token 减少 14.2–27.7%。

This paper analyzes the evolution of attention routing in recurrent language models and proposes WISE, a training-free inference method that reuses stabilized sparse attention support to achieve up to 1.76× attention speedup while preserving performance on multi-hop QA benchmarks.

CoVeR is a coverage-based routing method that reduces LLM verifier calls by 62-68% in agentic retrieval systems while maintaining accuracy on multi-hop QA benchmarks.

HyperProve proposes a retrieval-augmented QA framework for multi-hop questions using answer-guided hypergraph expansion, achieving superior performance over baselines in benchmarks.

CoG is a cognitive-inspired, training-free framework for adaptive knowledge exploration in retrieval-augmented generation, achieving state-of-the-art performance on multi-hop QA benchmarks through plan-explore-reflect cycles and bidirectional graph-text synergy.

This paper proposes a context-augmented training framework for multi-hop question-answering, showing that combining context graphs with knowledge graphs and using reinforcement learning improves performance in biomedical domains.

PARSER introduces a decoupled architecture using scatter-gather subagents for parallel chunk reading and iterative reasoning, significantly improving long-context multi-hop accuracy and reducing latency in LLM agents.

R2Adapter is a lightweight routing and rewriting adapter that dynamically allocates queries between vanilla and graph-based RAG, reducing graph-based RAG usage by up to 59% while maintaining accuracy in multi-hop QA benchmarks.

This paper proposes Evidence Sufficiency Boundary Training to enhance selective answering in grounded multi-hop QA systems by learning when to abstain from answering due to insufficient evidence, leading to improved boundary localization and reduced unsupported-answer rates on benchmarks like HotpotQA and MuSiQue.

ISO-RAG is a training-free, topology-driven retrieval-augmented generation framework that uses isoperimetric noise control to improve multi-hop question answering by reducing semantic drift and enhancing efficiency.

Hi-Q is an evidence-conditioned framework that dynamically refines multi-hop queries into hierarchical trees based on corpus support signals, improving retrieval and answer accuracy in multi-hop question answering.

ReliableRAG is a reliability-driven framework designed to combat misinformation in Retrieval-Augmented Generation systems for multi-hop question answering by evaluating fine-grained information triples to ensure accurate and faithful answers.

The article introduces PACE, a training-free framework for optimizing RAG systems by reordering candidates based on evidence coverage and dynamically adjusting reranking budgets, improving evidence recall and reducing latency.

The paper introduces AgenticRAG-FP, an interventional benchmark for causal failure attribution in agentic RAG systems, analyzing how failures propagate across retrieval hops and evaluating diagnosis performance.

This paper proposes SAG, a SQL-retrieval augmented generation architecture that organizes documents into event-entity hyperedges without building a global knowledge graph, enabling query-time dynamic linking of evidence chunks for multi-hop QA. It reports state-of-the-art retrieval and QA performance on HotpotQA, 2WikiMultiHopQA, and MuSiQue benchmarks.

Proposes Co-E, a training-free system that synchronizes graph and text memory for multi-hop question answering, improving over comparable training-free baselines on six benchmarks.