AI Paper Atlas

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

AAAI 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
add RAG (retrieve top-K rules)vanilla LLMs · RULERextrinsic hallucination / reliability: mean gain +1.9 extrinsic, +2.1 reliabilityImproved
add few-shot demonstrationsvanilla LLMs · RULERintrinsic hallucination score: mean gain +0.3No clear effect
add self-training on self-generated QA pairsvanilla LLMs · RULER application and complex reasoningcomposite score: gains on memorization (max 7.5) but application max 6.4, complex reasoning max 4.8Mixed
add DPO on top of self-trainingSelf-Training · RULERextrinsic hallucination / reliability: +0.5 extrinsic, +0.8 reliability (mean)Improved
existing enhancement methods (RAG, self-training) on complex reasoning vs memorizationQwen3-8B, Qwen3-14B · RULER hard itemsgain over vanilla: max gain +4.0 (knowledge) vs +1.4 (complex reasoning)Hurt
add majority voting at test timeQwen3 with enhancement methods · RULER complex reasoningcomposite score: Qwen3-8B Self-Training + DPO 4.9 -> 5.5Improved
add outcome reward model best-of-NQwen3 with enhancement methods · RULER complex reasoningcomposite score: Qwen3-14B Self-Training + DPO 5.4 -> 6.4Improved
add RAMPS rule-aware MCTS process reward with beam searchQwen3-14B Self-Training + DPO · RULER complex reasoningcomposite score: 5.4 -> 7.7 (+2.3)Improved
train PRM on advantage signal instead of expected score MC (ablation)RAMPS · RULER complex reasoningcomposite score: MC variant declines by 0.2-0.6 pointsImproved
use beam search instead of Best-of-N at inference (ablation)RAMPS · RULER complex reasoningcomposite score: Best-of-N mean degradation -0.4Improved
increase per-step samples k from 8 to 12 (Qwen3-8B) (ablation)RAMPS beam search · RULER complex reasoningcomposite score: +0.4 on average at about 1.5x costImproved
increase per-step samples k beyond 16, up to 32 (ablation)RAMPS beam search · RULER complex reasoningcomposite score: +0.07 for k=32 at almost double costNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy