---
title: "DynaKRAG: A Unified Framework for Learnable Evidence Control in Multi-Hop Retrieval-Augmented Generation"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.06507v1"
permalink: "https://feed.astralor.com/items/itm_a6c7e498cfad366d9418a77b"
published: "2026-07-07T17:09:36.000Z"
score: 77
channel: "ai"
tags: ["DynaKRAG","Qwen2.5-7B-Instruct","Retrieval-Augmented Generation","Multi-hop QA","State-conditioned control"]
---

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

## AI commentary

> This paper elevates evidence selection in multi-hop RAG from fixed pipelines to learnable state-conditioned control—a very elegant idea. It is especially worth reading for researchers working on QA systems and retrieval-augmented generation, particularly those interested in model autonomy in decision-making.

## At a glance

Multi-hop retrieval-augmented generation (RAG) requires sequential evidence acquisition, but existing methods typically rely on fixed pipelines or predefined control topologies, lacking learnable control over evidence operations. DynaKRAG formulates evidence acquisition as state-conditioned control, using a learned controller to select atomic operations (retrieval, query reformulation, evidence critique, sufficiency judging) and dynamically update the evidence state. Based on Qwen2.5-7B-Instruct, DynaKRAG achieves state-of-the-art F1 scores on three benchmarks including HotpotQA, demonstrating the benefit of coordinating retrieval, diagnosis, and gap-directed acquisition under evolving evidence state.

## Key points

- Proposes DynaKRAG, a unified framework that formulates multi-hop evidence acquisition as state-conditioned control over atomic evidence operations.
- Defines atomic operations (retrieval, query reformulation, evidence critique, sufficiency judging) and constructs executable action sets via a validity layer.
- Outperforms strong baselines on HotpotQA (F1 0.5998), 2Wiki (0.5340), and MuSiQue (0.3061).
- Ablation studies show the learned controller and sufficiency feedback are crucial; additional retrieval is not uniformly beneficial.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_a6c7e498cfad366d9418a77b