Thought Propagation: An Analogical Approach to Complex Reasoning with Large Language Models

Junchi Yu, Ran He, Rex Ying

Introduction

Despite the remarkable capabilities of Large Language Models (LLMs), they often struggle with complex, multi-step reasoning tasks. Current prompting techniques, such as Chain-of-Thought (CoT), typically approach problems in isolation—a method the authors describe as "reasoning from scratch." This approach fails to leverage the model's internal knowledge of similar problems and is highly susceptible to accumulated errors, where a single mistake early in a reasoning chain leads to a "hallucination rabbit hole."

To address these challenges, researchers have introduced Thought Propagation (TP). This framework enhances LLM reasoning by teaching models to think analogically. By proposing and solving related "analogous" problems, the model can aggregate insights to refine its final answer or create a more robust execution plan.

The Thought Propagation Framework

Thought Propagation is a three-stage "plug-and-play" framework designed to be compatible with existing solvers like CoT or Tree-of-Thought (ToT):

  1. LLM Propose: The model identifies and generates a set of analogous problems related to the input task. For example, in a navigation task, it might find paths to the target's neighboring nodes.
  2. LLM Solve: The model solves both the original and the analogous problems using standard reasoning methods.
  3. LLM Aggregate: The model analyzes the solutions to the analogous problems to either derive a better solution for the target or formulate a high-level "knowledge-intensive plan" to rectify errors in its initial attempt.
Comparison of reasoning from scratch versus Thought Propagation in an agent planning task.

This analogical approach allows the model to "remind" itself of correct logic or strategies that it might recognize in a similar context but fail to generate for the current one. TP can also be implemented in multiple layers (K-hop), though the authors found that 1 or 2 layers typically provide the best balance of performance and cost.

Experimental Evaluation

The researchers evaluated TP across three distinct domains: shortest-path reasoning in weighted graphs, creative writing, and LLM-agent planning in the ALFWorld environment.

In the shortest-path task, standard methods often fail to find optimal solutions due to the complexity of discrete optimization. As shown in the comparison below, traditional prompting methods like IO, CoT, and ToT can fall into sub-optimal paths or illogical loops.

Performance of different prompting methods on a shortest-path reasoning task, illustrating common failure modes.

Key Results

TP consistently outperformed baselines across all tasks:

  • Shortest Path: Achieved a 12% absolute increase in finding optimal solutions compared to the best baselines.
  • Creative Writing: Showed a 13% improvement in human preference scores, producing more coherent and logically structured stories.
  • Agent Planning: Boosted task completion rates by 15% in ALFWorld, helping agents escape repetitive loops where they would otherwise try the same failing action.

Insights and Limitations

A core insight of this work is that solving analogies effectively reduces the "reasoning distance." In graph tasks, solving for a target's neighbors turns a complex multi-step search into a simpler verification step. Furthermore, TP relies entirely on the LLM's internal "latent" knowledge rather than external databases, making it highly versatile.

However, the framework has limitations. Multi-layer implementations can lead to exponential growth in token usage. Additionally, the system's success is dependent on the quality of the analogies proposed; if the model generates irrelevant or hallucinated solutions to analogous problems, those errors can propagate back to the main task.

Overall, Thought Propagation offers a compelling path toward moving LLMs away from brute-force reasoning and toward a more human-like, associative style of problem-solving.

Summary updated

BibTeX
@inproceedings{yu2024thought,
  author = {Junchi Yu and Ran He and Rex Ying},
  title = {{Thought Propagation: An Analogical Approach to Complex Reasoning with Large Language Models}},
  booktitle = {International Conference on Learning Representations},
  year = {2024},
  url = {https://openreview.net/forum?id=SBoRhRCzM3},
}
Figure