Research map / Large language models / Agents, code and math
THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS
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.
Proposes Thought Propagation, which prompts LLMs to generate and solve analogous problems and reuses their solutions to refine reasoning on shortest-path, creative writing, and ALFWorld agent planning tasks.
Problem
LLM reasoning from scratch cannot reuse insights from similar problems and accumulates errors across multi-step reasoning, which hurts optimization, open-ended writing, and long-horizon agent planning.
Approach
Thought Propagation has three modules (LLM Propose, LLM Solve, LLM Aggregate): it generates analogous problems without external knowledge bases, solves them with existing prompting (IO/CoT/ReAct), and then either derives a new solution or a high-level plan from their results, with the LLM choosing the best candidate; multi-layer stacking extends this to K-hop analogies.
Key results
- On shortest-path reasoning with GPT-4 (0-shot), TP reaches 0.88 optimal rate vs 0.78 for IO and 0.46 for ToT.
- On Creative Writing with GPT-4, TP achieves coherent score 7.989 and 45% user preference vs 33% for ToT.
- On ALFWorld, TP-SM reaches 94.78% task completion vs 79.1% for Reflexion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Thought Propagation (analogous problems + aggregation) over IO prompting | GPT-4 shortest-path reasoning · Shortest-path Reasoning 0-shot | Optimal Rate: 0.78 -> 0.88 | Helped |
| Add Thought Propagation over IO prompting | GPT-3.5 shortest-path reasoning · Shortest-path Reasoning 0-shot | Optimal Rate: 0.33 -> 0.65 | Helped |
| Replace reasoning-from-scratch ToT with Thought Propagation | ToT on GPT-3.5 shortest-path · Shortest-path Reasoning 0-shot | Optimal Rate: 0.22 -> 0.65 | Helped |
| Add Thought Propagation over IO prompting | PaLM-2 shortest-path reasoning · Shortest-path Reasoning 0-shot | Optimal Rate: 0.14 -> 0.36 | Helped |
| Add Thought Propagation over ToT for GPT-4 shortest-path | GPT-4 shortest-path reasoning · Shortest-path Reasoning 0-shot | Optimal Rate: 0.46 -> 0.88 | Helped |
| Add Thought Propagation over ToT and CoT for creative writing | GPT-4 creative writing · Creative Writing | Coherent score and user preference: coherent 7.684 -> 7.989; user preference 33% -> 45% (vs ToT) | Helped |
| Add Thought Propagation over ToT for GPT-3.5 creative writing | GPT-3.5 creative writing · Creative Writing | Coherent score: 6.856 -> 7.000 | Helped |
| Add simulation-based plan selection and Thought Propagation plans from similar tasks (TP-SM) over ReAct/Reflexion | GPT-3 ALFWorld agent · ALFWorld | All-task completion rate: 79.1 (Reflexion) -> 94.78 (TP-SM) | Helped |
| Remove simulation and use self-evaluation only (TP-SE vs TP-SM) (ablation) | Thought Propagation on ALFWorld · ALFWorld | All-task completion rate: 94.78 -> 83.68 | Hurt |
| Add self-reflection memory to Thought Propagation with simulation (TP-SR-SM vs TP-SM) (ablation) | Thought Propagation on ALFWorld · ALFWorld | All-task completion rate: 94.78 -> 92.54 | Hurt |
| Add self-reflection to self-evaluation variant (TP-SR-SE vs TP-SE) (ablation) | Thought Propagation on ALFWorld · ALFWorld | All-task completion rate: 83.68 -> 85.82 | Helped |
| Increase TP from 1 layer to 2 layers (ablation) | Thought Propagation on shortest-path reasoning · Shortest-path Reasoning | Optimal Rate: marginal gain over 1-layer | No clear effect |
Surprising findings
- 5-shot prompting gives only marginal or sometimes worse gains over 1-shot for prompting methods, attributed to redundant context and output cutoff.
- ToT underperforms IO prompting on GPT-3.5 and GPT-4 in shortest-path reasoning, due to accumulated errors and backward search.
Limitations
- 2-layer TP gives marginal gains over 1-layer while increasing token cost.
- Shortest-path graphs are small (5 to 10 nodes), and the 5-shot setting shows diminishing returns that were not fully explored.
Experimental setup
Datasets: ALFWorld Creative Writing task dataset (Yao et al.)
Benchmarks: Shortest-path reasoning (100 generated graphs) Creative Writing (100 instances) ALFWorld (134 unseen environments)
Metrics: Optimal Rate Feasible Rate Over-Length Rate coherent score user preference task completion rate
Baselines: IO prompting Chain-of-Thought Build-a-Graph Tree-of-Thought BUTLER ReAct Act Reflexion
Builds on: Chain-of-Thought prompting Tree-of-Thought prompting ReAct Reflexion Build-a-Graph prompting
Closest papers by meaning
- Boosting of Thoughts: Trial-and-Error Problem Solving with Large Language Models (ICLR 2024 · similarity 0.91)
- StrategyLLM: Large Language Models as Strategy Generators, Executors, Optimizers, and Evaluators for Problem Solving (NeurIPS 2024 · similarity 0.90)
- Large Language Models as Analogical Reasoners (ICLR 2024 · similarity 0.90)
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.90)
- Toward Adaptive Reasoning in Large Language Models with Thought Rollback (ICML 2024 · similarity 0.90)
- Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models (NeurIPS 2024 · 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.