Existing LLMs Are Not Self-Consistent For Simple Tasks

2025-06-23

AI Summary

This work identifies pervasive self-inconsistency in large language models (LLMs)—i.e., logically contradictory outputs across multiple inferences on identical inputs—when performing elementary symbolic reasoning tasks such as geometric point-ordering and kinship relation inference, thereby undermining decision reliability and interpretability. To systematically quantify and mitigate this issue, we propose the first automated inconsistency detection and correction framework grounded in graph-structured modeling and energy-based optimization, accompanied by a differentiable coherence evaluation metric. Experiments across major open- and closed-weight models—including DeepSeek-R1 and GPT-4-mini—demonstrate significant inconsistency even in basic reasoning. Our method consistently improves consistency scores, yet complete elimination remains fundamentally challenging. This study establishes a novel paradigm and benchmark toolkit for assessing LLM reliability and enabling trustworthy symbolic reasoning.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Natural Language Processing: (Large) Language Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large Language Models (LLMs) have grown increasingly powerful, yet ensuring their decisions remain transparent and trustworthy requires self-consistency -- no contradictions in their internal reasoning. Our study reveals that even on simple tasks, such as comparing points on a line or a plane, or reasoning in a family tree, all smaller models are highly inconsistent, and even state-of-the-art models like DeepSeek-R1 and GPT-o4-mini are not fully self-consistent. To quantify and mitigate these inconsistencies, we introduce inconsistency metrics and propose two automated methods -- a graph-based and an energy-based approach. While these fixes provide partial improvements, they also highlight the complexity and importance of self-consistency in building more reliable and interpretable AI. The code and data are available at https://github.com/scorpio-nova/llm-self-consistency.

Problem

• LLMs lack self-consistency in simple reasoning tasks
• Existing models show contradictions in basic spatial and familial logic
• New metrics and methods aim to quantify and reduce inconsistencies

Innovation

• Introduce inconsistency metrics for LLMs
• Propose graph-based automated correction method
• Develop energy-based approach for improvements