Research map / Large language models / Reasoning and chain of thought
Structured Reasoning for LLMs: A Unified Framework for Efficiency and Explainability
Yubo Dong, Hehe Fan, Linchao Zhu, Yi Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Structured Reasoning, a framework inspired by cognitive science and neurosymbolic AI, aimed at improving LLM complex reasoning (logical deduction, planning) beyond token-level probability relationships, with claimed gains in efficiency and explainability.
Problem
LLMs struggle with complex reasoning such as logical deduction and planning because they rely primarily on token-level probability relationships.
Approach
Introduces Structured Reasoning, a unified framework that moves reasoning beyond token-level probabilities, drawing on cognitive science and neurosymbolic AI to improve both efficiency and explainability. Details are not available from the abstract alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MaxFlow reward to GRPO | DeepSeek-R1-Distill-Qwen-1.5B · 9-benchmark average, 8K length | avg pass@1: 43.28 -> 48.89 | Improved |
| add MaxFlow reward to GRPO | DeepSeek-R1-Distill-Qwen-7B · 9-benchmark average, 8K length | avg pass@1: 58.33 -> 61.17 | Improved |
| add LCS reward to GRPO | DeepSeek-R1-Distill-Qwen-7B · 9-benchmark average, 8K length | avg pass@1: 58.33 -> 57.65 | Hurt |
| add LCS reward to GRPO at short lengths | DeepSeek-R1-Distill-Qwen-7B · 9-benchmark average, 1K length | avg pass@1: 29.19 -> 35.06 | Improved |
| use MaxFlow reward | DeepSeek-R1-Distill-Qwen-1.5B · temperature robustness, MATH500 | accuracy std: ±0.79 -> ±0.53 | Improved |
| use attention-based Max-Flow step importance instead of perplexity (ablation) | step filtering · Error filtering efficiency, 9 steps | EFE: 68.07 -> 76.67 | Improved |
| use gold-order structured steps instead of shuffled (ablation) | attention matrices · Entailment Trees task 1 | alignment: 28.48% -> 71.27% | Improved |
| add structured guidance prompting without training | DeepSeek-R1 671B · MATH500 | avg tokens: 2945 -> 1577 | Improved |
| add structured guidance prompting without training | DeepSeek-R1 671B · AMC23 | avg tokens: 1716 -> 2053 | Hurt |
| combine Max-Flow and LCS rewards 1:1 (ablation) | GRPO on 1.5B · Large Avg. | pass@1: 58.1 (Max-Flow) vs 57.20 (joint) | No clear effect |
| add MaxFlow reward to GRPO at 4K evaluation length | DeepSeek-R1-Distill-Qwen-1.5B · 9-benchmark average, 4K length | avg pass@1: 39.28 -> 41.68 | Improved |
Closest papers by meaning
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.92)
- A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.91)
- GPO: Learning from Critical Steps to Improve LLM Reasoning (NeurIPS 2025 · similarity 0.91)
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (ICML 2025 · similarity 0.91)
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.91)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · 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.
