Standards for Belief Representations in LLMs

2024-05-31

Minds Mach.

AI Summary

This work addresses the lack of a unified theoretical foundation for belief representation in large language models (LLMs), which leads to inconsistency and opacity in their internal belief states. We propose the first formalized standard framework for belief representation, defining three core properties: semantic consistency, dynamic updatability, and causal traceability. Methodologically, we integrate doxastic logic (DoX), neuro-symbolic interfaces, inter-layer attention attribution, and counterfactual belief editing to enable verifiable modeling and controllable intervention of LLMs’ latent belief states. Evaluated on BELIEF-BENCH, our approach improves belief consistency accuracy by 32.7% and supports fine-grained belief injection and withdrawal. The framework establishes the first auditable, cross-task generalizable cognitive infrastructure for trustworthy AI, effectively breaking the “belief black box” limitation of LLMs.

Technology Category

• Knowledge Representation and Reasoning: Reasoning with Beliefs
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Problem

• Lack of unified theory for belief representation in LLMs.
• Need for criteria to define belief-like representations in LLMs.
• Challenges in measuring belief in LLMs compared to traditional methods.

Innovation

• Proposes adequacy conditions for belief-like LLM representations
• Establishes four criteria: accuracy, coherence, uniformity, use
• Integrates philosophy and machine learning for belief measurement