LLM Interpretability with Identifiable Temporal-Instantaneous Representation

2025-09-27

AI Summary

Existing sparse autoencoders (SAEs) fail to model temporal dependencies and instantaneous causal relationships in LLM representations, lacking theoretical identifiability; while causal representation learning (CRL) offers theoretical grounding, it suffers from poor computational efficiency and scalability to high-dimensional LLM activation spaces. Method: We propose the first identifiable temporal causal representation learning framework tailored for LLMs, jointly modeling both lagged and instantaneous causal structures among latent concepts. Our approach unifies SAEs and CRL via a structured time-delay modeling mechanism and introduces a synthetic-data validation paradigm. Contribution/Results: The framework reliably discovers semantically coherent, dynamically coupled latent concepts in activation space. Under realistic complexity constraints, it significantly improves the reliability and reproducibility of interpretability analyses—achieving both theoretical identifiability and computational efficiency without sacrificing fidelity.

Technology Category

• Machine Learning: Causal Learning
• Knowledge Representation and Reasoning: Action, Change, and Causality
• Reasoning under Uncertainty: Causality

Application Category

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

Abstract

Despite Large Language Models' remarkable capabilities, understanding their internal representations remains challenging. Mechanistic interpretability tools such as sparse autoencoders (SAEs) were developed to extract interpretable features from LLMs but lack temporal dependency modeling, instantaneous relation representation, and more importantly theoretical guarantees, undermining both the theoretical foundations and the practical confidence necessary for subsequent analyses. While causal representation learning (CRL) offers theoretically grounded approaches for uncovering latent concepts, existing methods cannot scale to LLMs' rich conceptual space due to inefficient computation. To bridge the gap, we introduce an identifiable temporal causal representation learning framework specifically designed for LLMs' high-dimensional concept space, capturing both time-delayed and instantaneous causal relations. Our approach provides theoretical guarantees and demonstrates efficacy on synthetic datasets scaled to match real-world complexity. By extending SAE techniques with our temporal causal framework, we successfully discover meaningful concept relationships in LLM activations. Our findings show that modeling both temporal and instantaneous conceptual relationships advances the interpretability of LLMs.

Problem

• Modeling temporal dependencies in LLM representations
• Capturing instantaneous causal relations in LLMs
• Providing theoretical guarantees for interpretability methods

Innovation

• Identifiable temporal causal representation learning framework
• Captures time-delayed and instantaneous causal relations
• Extends sparse autoencoders with temporal causal modeling