Research map / Large language models / Retrieval and knowledge
HiPRAG: Hierarchical Process Rewards for Efficient Agentic Retrieval Augmented Generation
Peilin Wu, Mian Zhang, Kun Wan, Wentian Zhao, Kaiyu He, Xinya Du, Zhiyu Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HiPRAG hierarchical process reward | Qwen2.5-7B-Instruct GRPO · 7 QA benchmarks | Avg. CEM: 61.2 -> 67.2 | Improved |
| add HiPRAG process reward | Qwen2.5-7B-Instruct PPO · 7 QA benchmarks avg | Avg. CEM: 53.3 -> 64.5 | Improved |
| add HiPRAG process reward | Llama-3.2-3B-Instruct PPO · 7 QA benchmarks avg | Avg. CEM: 56.4 -> 64.8 | Improved |
| add HiPRAG process reward | Qwen2.5-3B-Instruct GRPO · 7 QA benchmarks avg | Avg. OSR: 8.4 -> 4.1 | Improved |
| use HiPRAG parsable step output format with Search-R1 outcome reward (ablation) | Search-R1 · 7 QA benchmarks avg | Avg. CEM: 60.3 -> 61.2 | No clear effect |
| use HiPRAG parsable step output format with beta-GRPO reward (ablation) | beta-GRPO · 7 QA benchmarks avg | Avg. CEM: 62.5 -> 62.1 | No clear effect |
| add over-search-only penalty instead of full process reward (ablation) | HiPRAG Qwen2.5-3B-Instruct PPO · 7 QA benchmarks avg | Avg. CEM: 64.1 -> 58.8 | Hurt |
| add over-search-only penalty instead of full process reward (ablation) | HiPRAG Qwen2.5-3B-Instruct PPO · 7 QA benchmarks avg | Avg. USR: 38.1 -> 52.7 | Hurt |
| add under-search-only penalty instead of full process reward (ablation) | HiPRAG Qwen2.5-3B-Instruct PPO · 7 QA benchmarks avg | Avg. CEM: 64.1 -> 63.3 | Hurt |
| add under-search-only penalty instead of full process reward (ablation) | HiPRAG Qwen2.5-3B-Instruct PPO · 7 QA benchmarks avg | Avg. OSR: 4.9 -> 6.6 | Hurt |
| process bonus coefficient 0.2 instead of 0.4 (ablation) | HiPRAG Qwen2.5-3B-Instruct PPO · 7 QA benchmarks avg | Avg. CEM: 64.1 -> 59.6 | Hurt |
| process bonus coefficient 0.6 instead of 0.4 (ablation) | HiPRAG Qwen2.5-3B-Instruct PPO · 7 QA benchmarks avg | Avg. CEM: 64.1 -> 62.5 | Hurt |
Closest papers by meaning
- Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning (NeurIPS 2025 · similarity 0.93)
- Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.92)
- MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.91)
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.91)
- SD-Search: Hindsight Self-Distillation for Search-Augmented Reasoning (NeurIPS 2026 · similarity 0.90)
- Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
