Research map / Large language models / Retrieval and knowledge
Thinking Forward and Backward: Multi-Objective Reinforcement Learning for Retrieval-Augmented Reasoning
Wenda Wei, Yu-An Liu, Ruqing Zhang, Jiafeng Guo, Lixin Su, Shuaiqiang Wang, Dawei Yin, Maarten de Rijke, Xueqi Cheng
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 backward reward and interpolate with forward model (ablation) | Forward-RAR (Qwen2.5-3B-Base) · 7 QA benchmarks Avg. | Exact Match: 0.333 -> 0.338 | Improved |
| add backward reward and interpolate with forward model (ablation) | Forward-RAR (Qwen2.5-3B-Instruct) · 7 QA benchmarks Avg. | Exact Match: 0.384 -> 0.397 | Improved |
| add forward reward and interpolate with backward model (ablation) | Backward-RAR (Qwen2.5-3B-Base) · 7 QA benchmarks Avg. | Exact Match: 0.321 -> 0.338 | Improved |
| add forward reward and interpolate with backward model (ablation) | Backward-RAR (Qwen2.5-3B-Instruct) · 7 QA benchmarks Avg. | Exact Match: 0.377 -> 0.397 | Improved |
| use forward (answer-side) reward instead of backward reward (ablation) | Single-reward RAR (Base) · 7 QA benchmarks Avg. | Exact Match: 0.321 -> 0.333 | Improved |
| use forward reward instead of backward reward (ablation) | Single-reward RAR (Instruct) · 7 QA benchmarks Avg. | Exact Match: 0.377 -> 0.384 | Improved |
| replace outcome-only reward with bidirectional information-distance rewards | Search-R1 (Base) · 7 QA benchmarks Avg. | Exact Match: 0.312 -> 0.338 | Improved |
| replace outcome-only reward with bidirectional information-distance rewards | Search-R1 (Instruct) · 7 QA benchmarks Avg. | Exact Match: 0.336 -> 0.397 | Improved |
| replace outcome-only reward with bidirectional rewards | Search-R1 (Instruct) · HotpotQA | Exact Match: 0.331 -> 0.391 | Improved |
| replace outcome-only reward with bidirectional rewards | Search-R1 (Instruct) · 2WikiMultiHopQA | Exact Match: 0.310 -> 0.402 | Improved |
| replace outcome-only reward with bidirectional rewards | Search-R1 (Base) · Bamboogle | Exact Match: 0.128 -> 0.188 | Improved |
| replace outcome-only reward with bidirectional rewards | Search-R1 (Instruct) · Bamboogle | Exact Match: 0.232 -> 0.363 | Improved |
Closest papers by meaning
- Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning (NeurIPS 2025 · similarity 0.90)
- Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs (AAAI 2025 · similarity 0.90)
- Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Scent of Knowledge: Optimizing Search-Enhanced Reasoning with Information Foraging (NeurIPS 2025 · similarity 0.89)
- SD-Search: Hindsight Self-Distillation for Search-Augmented Reasoning (NeurIPS 2026 · similarity 0.89)
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
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.
