DynaKRAG introduces learnable evidence control for multi-hop RAG

The authors introduce DynaKRAG, a unified framework that formulates multi-hop retrieval-augmented generation as state-conditioned control over atomic evidence operations. At each step, a validity layer constructs the executable action set while a learned controller selects the next operation to update the evidence state.

• DynaKRAG uses Qwen2.5-7B-Instruct and achieves F1 scores of 0.5998 on HotpotQA, 0.5340 on 2Wiki, and 0.3061 on MuSiQue.
• The system outperforms 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.
• Removing sufficiency feedback negatively impacts performance across all datasets.

These results demonstrate the benefit of coordinating retrieval, diagnosis, and gap-directed acquisition under an evolving evidence state.