Research map / Large language models / Agents, code and math
Decompose, Analyze and Rethink: Solving Intricate Problems with Human-like Reasoning Cycle
Read the paperCheck your idea against 108,967 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.
Introduces DeAR, a Decompose-Analyze-Rethink cycle that builds a reasoning tree of sub-questions within a single LLM and updates parent rationales from child feedback, outperforming ToT and GoT on ScienceQA, StrategyQA and GSM8K.
Problem
Rationale-extension and search-based prompting (CoT, ToT, GoT) generate reasoning sequentially without a global view, so early errors propagate and cannot be corrected, and their fixed branching is rigid for intricate reasoning.
Approach
Build a reasoning tree top-down by prompting the LLM to decompose questions into adaptively many sub-questions (with retrieved human-annotated decomposition exemplars), generate and self-check rationales per node, and then rethink by updating ancestor rationales using the coherence-scored child rationales.
Key results
- On GPT-3.5, DeAR reaches 83.68 on ScienceQA, 73.36 on StrategyQA and 86.82 on GSM8K, versus 82.52, 71.89 and 83.42 for ToT.
- On MATH with GPT-4, DeAR achieves 62.25 accuracy versus 58.78 for GoT and 57.18 for ToT.
- DeAR uses fewer API calls than ToT and GoT on ScienceQA (9.82 vs 11.35 and 13.74) while reaching higher accuracy (0.837).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeAR instead of the strongest reported baseline | GPT-3.5 · ScienceQA | accuracy: 82.52 -> 83.68* | Helped |
| Use DeAR instead of the strongest reported baseline | LLaMA2-7B · ScienceQA | accuracy: 69.33 -> 70.57* | Helped |
| Use DeAR instead of the strongest reported baseline | ChatGLM3-6B · ScienceQA | accuracy: 49.23 -> 51.08* | Helped |
| Use DeAR instead of the strongest reported baseline | GPT-3.5 · StrategyQA | accuracy: 72.02 -> 73.36* | Helped |
| Use DeAR instead of the strongest reported baseline | LLaMA2-7B · StrategyQA | accuracy: 66.61 -> 68.33* | Helped |
| Use DeAR instead of the strongest reported baseline | ChatGLM3-6B · StrategyQA | accuracy: 61.22 -> 61.02* | Hurt |
| Use DeAR instead of the strongest reported baseline | GPT-3.5 · GSM8K | accuracy: 84.77 -> 86.82* | Helped |
| Use DeAR instead of the strongest reported baseline | LLaMA2-7B · GSM8K | accuracy: 75.28 -> 78.01* | Helped |
| Use DeAR instead of the strongest reported baseline | ChatGLM3-6B · GSM8K | accuracy: 56.72 -> 58.54* | Helped |
| Add logic heuristics to decomposition prompts (rather than remove them) (ablation) | DeAR with GPT-3.5 · ScienceQA | accuracy: 83.06 -> 83.68 | Helped |
| Add logic heuristics to decomposition prompts (rather than remove them) (ablation) | DeAR with LLaMA2-7B · ScienceQA | accuracy: 69.85 -> 70.57 | Helped |
| Add logic heuristics to decomposition prompts (rather than remove them) (ablation) | DeAR with ChatGLM3-6B · ScienceQA | accuracy: 50.17 -> 51.08 | Helped |
Surprising findings
- Random rationale updating at 100% of nodes performs worse than no update on ScienceQA (81.16 vs 82.77), suggesting indiscriminate updates hurt accuracy.
Limitations
- Relies on prompting with thresholds (epsilon1, epsilon2) tuned per dataset on a validation split.
- Decomposition demonstration pool for ScienceQA was generated with GPT-4 and manually checked, so it depends on a stronger model's annotations.
- Efficiency and accuracy comparisons rely on fixed branch/depth settings for ToT/GoT variants.
Experimental setup
Datasets: ScienceQA StrategyQA GSM8K MATH
Benchmarks: ScienceQA StrategyQA GSM8K MATH
Metrics: accuracy ROSCOE source-consistency ROSCOE reasoning alignment average inference time API calls
Baselines: Few-shot prompting Chain-of-Thought Tree-of-Thoughts Graph-of-Thoughts Least-to-most SelfCheck
Builds on: Tree-of-Thoughts Graph-of-Thoughts Chain-of-Thought Least-to-most prompting SelfCheck
Closest papers by meaning
- Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment (ICLR 2025 · similarity 0.90)
- THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS (ICLR 2024 · similarity 0.89)
- Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning (ICML 2025 · similarity 0.89)
- RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation (ICML 2025 · similarity 0.89)
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.89)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.89)
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.