DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation

2026-07-07

AI Summary

This work addresses the limitations of existing approaches in multi-hop retrieval-augmented generation, which rely on fixed pipelines and lack dynamic control over evidence manipulation. The authors propose the first unified state-conditioned control framework, modeling multi-hop evidence acquisition as a sequence of atomic operations conditioned on the current reasoning state. A validity filtering layer constructs a feasible action set, from which a learnable controller adaptively selects the optimal operation. Integrating state-conditioned policy learning with the Qwen2.5-7B-Instruct model, the method is optimized end-to-end and achieves F1 scores of 0.5998, 0.5340, and 0.3061 on HotpotQA, 2WikiMultihopQA, and MuSiQue, respectively—significantly outperforming existing controllable baselines. Ablation studies confirm the critical contributions of both the learned controller and the sufficiency-based feedback mechanism.

Abstract

Multi-hop retrieval-augmented generation (RAG) acquires evidence sequentially, with each new document potentially revealing missing facts, bridge entities, query defects, or sufficient support for answering. Existing methods provide useful operations such as iterative retrieval, query reformulation, evidence critique, and sufficiency judging, but typically organize them within method-specific pipelines or predefined control topologies. This leaves underexplored how to learn a shared state-conditioned policy that chooses among currently valid evidence operations. We introduce DynaKRAG, which formulates multi-hop evidence acquisition as state-conditioned control over atomic evidence operations. At each step, a validity layer constructs the executable action set, and a learned controller selects the next operation. The resulting transition updates the evidence state and may enable new operations at subsequent steps. With Qwen2.5-7B-Instruct, DynaKRAG achieves F1 scores of 0.5998 on HotpotQA, 0.5340 on 2Wiki, and 0.3061 on MuSiQue, outperforming the strongest controlled baseline on all three benchmarks. Replacing the learned controller with a uniform-valid policy reduces F1 by 3.96--5.78 points, while removing sufficiency feedback hurts all three datasets. Controlled retrieval-cap experiments further show that additional retrieval is not uniformly beneficial. Together, these results demonstrate the benefit of coordinating retrieval, diagnosis, and gap-directed acquisition under an evolving evidence state.

Problem

• multi-hop retrieval-augmented generation
• evidence control
• learnable policy
• state-conditioned control
• retrieval-augmented generation

Innovation

• learnable control
• multi-hop RAG
• evidence state
• atomic evidence operations
• dynamic retrieval