Tag
该论文提出三阶段流程识别多语言大模型中的"Bridge Routing Heads"（BRH），发现五种语言对应的头集合几乎完全互斥，证明多语言多跳推理由语言特异电路实现；消融实验显示因果效应，并通过激活级干预无需训练即可挽救高达51.7%的跨语言推理失败。
The paper introduces horizontal and vertical proactivity in tool-using agents and proposes Q&D, a training method that rewards a questioner for retrieving required evidence without a reward model or judge. On multi-hop QA benchmarks and a simulated customer-service setting, the trained questioner beats a prompted model 15x its size on two of three benchmarks, showing that what agents pursue and when they stop matters as much as whether they act unprompted.
LatWeave introduces a deterministic multi-hop question answering system using knowledge lattices to eliminate hallucination and ensure auditable evidence chains through meet, compare, and abstain operators.
This paper presents FedV-KGQA, a federated system for multi-hop question answering over vertically partitioned knowledge graphs, sharing empirical findings, design lessons, and an interactive prototype.
Mosaic is a training-free framework for query-aware exploration in GraphRAG that adapts retrieval policies based on query requirements, showing significant improvements in answer correctness and efficiency on standard benchmarks.
DocHop introduces a benchmark for evaluating multi-hop reasoning in information-dense documents, focusing on chart-context integration in multimodal large language models, with experiments showing a significant gap between model and human performance.
PRO-Step introduces a step-level process reward optimization method for retrieval-augmented generation to improve multi-hop reasoning by evaluating logical validity and evidential grounding at each step.
SelfGraphRAG introduces a framework that generates synthetic question-answer pairs from knowledge graphs to address the supervision gap in graph-based retrieval-augmented generation, enhancing retrieval precision and reasoning performance.
SSE-Bio proposes a structured self-evolving agent with a trainable retrieval policy to improve multi-hop biomedical reasoning, demonstrating significant performance gains over baselines.
SABET-QA introduces an iterative framework for temporal knowledge graph question answering that enhances multi-hop reasoning through bidirectional entity-temporal scoring and contextualization, showing consistent improvements over baselines on benchmarks like CronQuestions and TimeQuestions.
MolBioKG is a two-layer system that grounds unseen molecules in biomedical knowledge graphs via multi-resolution structural anchoring, enabling out-of-graph link recovery and multi-hop reasoning from SMILES strings. It significantly improves Hits@10 and out-of-graph target recall over baselines while maintaining traceable evidence.
This paper introduces D2F-ReAG, a novel paradigm for multi-hop reasoning-augmented generation that dynamically controls reasoning depth by judging root-level reasoning reliability and decomposing questions into sub-questions when needed, improving accuracy on multi-hop benchmarks.
This paper introduces a new benchmark for evaluating machine unlearning, focusing on multi-hop reasoning consistency and recovery robustness. Experiments show existing unlearning methods face an 'impossible triangle' trade-off among forget quality, robustness, and utility preservation.
Introduces Multimodal CoLRAG-TF, a four-axis fusion architecture for retrieval-augmented generation over complex PDFs, integrating dense text, BM25, knowledge-graph triple filtering, and image similarity. Achieves significant improvements in multi-hop reasoning on a benchmark of Japanese disaster lesson PDFs.
Telco-GAIA is a bilingual, multi-modal benchmark for evaluating tool-using agents in the telecom domain, comprising 100 human-verified tasks requiring multi-hop reasoning over heterogeneous sources, with objective scoring via exact string matching.
This paper presents Search-on-Graph-R1 (SoG-R1), which trains an 8B LLM to navigate knowledge graphs by first scaffolding a frontier teacher with gold SPARQL queries to produce grounded trajectories, then applying supervised fine-tuning and reinforcement learning. The compact model surpasses frozen frontier systems on WebQSP, CWQ, and GrailQA, notably achieving the best results on CWQ among compared methods.
DeLIVeR is a framework that uses a reinforced planner LLM to decompose claims into question sets for structured knowledge graph traversal, improving fact-checking accuracy over static RAG baselines by 10-15% on benchmark datasets.
Introduces GRASP, a reinforcement learning framework that trains agents to adaptively coordinate semantic search, keyword search, and paragraph reading during multi-step reasoning, improving retrieval recall and question answering performance on multi-hop benchmarks.
This paper introduces RSF-GLLM, a framework that decouples differentiable graph reasoning from LLM generation to address the semantic gap in multi-hop knowledge graph question answering, achieving competitive performance with superior inference efficiency.
A developer reflects on when GraphRAG is preferable to standard RAG for AI agents, noting it becomes valuable for multi-hop reasoning and relationship understanding, but adds significant complexity.