Concise and Organized Perception Facilitates Reasoning in Large Language Models

TL;DR

This paper tackles the brittleness of large language models in multi-hop logical reasoning when faced with disorderly and distractive premises. It introduces Concise and Organized Perception (COP), a three-stage method that captures locally-related premises, builds a tree-like mind map anchored to the query, and reconstructs a concise, organized context to elicit better reasoning from LLMs. Across ProofWriter, PrOntoQA, PrOntoQA-OOD, FOLIO, and DI-GSM, COP achieves state-of-the-art accuracy and significantly reduces inference calls and token usage, demonstrating strong efficiency and transferability across diverse models. The work highlights information-flow saliency as a lens for understanding LLM reasoning and suggests COP can complement existing prompting strategies to better cope with distractibility and disorder in complex reasoning tasks.

Abstract

Exploiting large language models (LLMs) to tackle reasoning has garnered growing attention. It still remains highly challenging to achieve satisfactory results in complex logical problems, characterized by plenty of premises within the context and requiring multi-hop reasoning. In particular, the reasoning capabilities of LLMs are brittle to disorder and distractibility. In this work, we first examine the mechanism from the perspective of information flow and reveal that LLMs confront difficulties akin to human-like cognitive biases when dealing with disordered and irrelevant content in reasoning tasks. However, in contrast to LLMs, disordered and irrelevant content does not significantly decrease human performance, as humans have a propensity to distill the most relevant information and systematically organize their thoughts, aiding them in responding to questions.Stem from that, we further propose a novel reasoning approach named Concise and Organized Perception (COP). COP carefully analyzes the given statements to identify the most pertinent information while eliminating redundancy efficiently. It then prompts the LLMs in a more organized form that adapts to the model's inference process. By perceiving concise and organized context, the reasoning abilities of LLMs can be better elicited. Extensive experimental results on several popular logical benchmarks (ProofWriter, PrOntoQA, PrOntoQA-OOD, and FOLIO) and mathematical benchmark (DI-GSM) show that COP significantly outperforms previous state-of-the-art methods.

Figures (12)

• Figure 1: (a) A 5-hop example of ProofWriter dataset, showcasing plenty of premises. Some premises are omitted for brevity. (b) Corresponding reconstruction of concise and organized perception. Superscript serial numbers represent the logical orders according to the gold proof. The concise input contains only relevant information but lacks organizational structure. In contrast, the organized input arranges statements consistently with the gold reasoning path, albeit including redundant information. (c) LLMs outputs. (d) Results of a confirmatory experiment.
• Figure 2: (a)(b)(c) Saliency score analysis on an example of ProofWriter based on shallow layers of Llama-2-13B-Chat. The horizontal coordinate contains the step by step outputs, and the vertical coordinate contains the inputs and outputs. Values in the plot represent saliency scores from column to row, normalized by each column.
• Figure 3: Saliency score analysis on ProofWriter based on shallow and deep layers of Llama-2-13B-Chat. "Base", "Disordered", and "Irrelevant" respectively denote the samples corresponding to the three scenarios depicted in Figure \ref{['fig:fig_score_case']} (a)(b)(c).
• Figure 4: Overview of the proposed COP with an example on DI-GSM (constructed from GSM8K GSM8K) with disordered and irrelevant premises. Green represents relevant premises [$p_2,p_3,p_4,p_6$], black represents irrelevant premises [$p_1,p_5$], and orange represents the question [$Q$]. Details of DI-GSM are listed in section\ref{['sec:sec_datasets']}.
• Figure 5: An example on ProofWriter of sub-mind-map segmentation and context reconstruction.