Research map / Large language models / Agents, code and math
Benchmarking and Enhancing Rule Knowledge-Driven Reasoning of Large Language Models
Zijie Xu, Wenjun Ke, Peng Wang, Guozheng Li, Qingjian Ni, Jiajun Liu, Ziyu Shang, Jing Zhou
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 RAG (retrieve top-K rules) | vanilla LLMs · RULER | extrinsic hallucination / reliability: mean gain +1.9 extrinsic, +2.1 reliability | Improved |
| add few-shot demonstrations | vanilla LLMs · RULER | intrinsic hallucination score: mean gain +0.3 | No clear effect |
| add self-training on self-generated QA pairs | vanilla LLMs · RULER application and complex reasoning | composite score: gains on memorization (max 7.5) but application max 6.4, complex reasoning max 4.8 | Mixed |
| add DPO on top of self-training | Self-Training · RULER | extrinsic hallucination / reliability: +0.5 extrinsic, +0.8 reliability (mean) | Improved |
| existing enhancement methods (RAG, self-training) on complex reasoning vs memorization | Qwen3-8B, Qwen3-14B · RULER hard items | gain over vanilla: max gain +4.0 (knowledge) vs +1.4 (complex reasoning) | Hurt |
| add majority voting at test time | Qwen3 with enhancement methods · RULER complex reasoning | composite score: Qwen3-8B Self-Training + DPO 4.9 -> 5.5 | Improved |
| add outcome reward model best-of-N | Qwen3 with enhancement methods · RULER complex reasoning | composite score: Qwen3-14B Self-Training + DPO 5.4 -> 6.4 | Improved |
| add RAMPS rule-aware MCTS process reward with beam search | Qwen3-14B Self-Training + DPO · RULER complex reasoning | composite score: 5.4 -> 7.7 (+2.3) | Improved |
| train PRM on advantage signal instead of expected score MC (ablation) | RAMPS · RULER complex reasoning | composite score: MC variant declines by 0.2-0.6 points | Improved |
| use beam search instead of Best-of-N at inference (ablation) | RAMPS · RULER complex reasoning | composite score: Best-of-N mean degradation -0.4 | Improved |
| increase per-step samples k from 8 to 12 (Qwen3-8B) (ablation) | RAMPS beam search · RULER complex reasoning | composite score: +0.4 on average at about 1.5x cost | Improved |
| increase per-step samples k beyond 16, up to 32 (ablation) | RAMPS beam search · RULER complex reasoning | composite score: +0.07 for k=32 at almost double cost | No clear effect |
Closest papers by meaning
- LLM+AL: Bridging Large Language Models and Action Languages for Complex Reasoning About Actions (AAAI 2025 · similarity 0.88)
- RuAG: Learned-rule-augmented Generation for Large Language Models (ICLR 2025 · similarity 0.88)
- Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment (ICLR 2025 · similarity 0.88)
- RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation (ICML 2025 · similarity 0.88)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.87)
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.
