Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning

2026-03-11

AI Summary

Existing methods struggle to capture the dynamic causal interactions among concepts in the multi-step reasoning processes of large language models (LLMs). This work proposes Causal Concept Graphs (CCGs)—the first approach to construct sparse, interpretable directed acyclic graphs in LLM latent spaces—by integrating task-conditioned sparse autoencoders with DAGMA-style differentiable structure learning to recover reasoning pathways. The study introduces a novel Causal Fidelity Score (CFS) to evaluate intervention efficacy and validates the framework on ARC-Challenge, StrategyQA, and LogiQA benchmarks. Using GPT-2 Medium, CCGs achieve a CFS of 5.654 ± 0.625, significantly outperforming ROME tracing, SAE-based ranking, and random baselines (p < 0.0001), while generating domain-specific graph structures that remain stable across random seeds.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Reasoning under Uncertainty: Causality

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: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

Sparse autoencoders can localize where concepts live in language models, but not how they interact during multi-step reasoning. We propose Causal Concept Graphs (CCG): a directed acyclic graph over sparse, interpretable latent features, where edges capture learned causal dependencies between concepts. We combine task-conditioned sparse autoencoders for concept discovery with DAGMA-style differentiable structure learning for graph recovery and introduce the Causal Fidelity Score (CFS) to evaluate whether graph-guided interventions induce larger downstream effects than random ones. On ARC-Challenge, StrategyQA, and LogiQA with GPT-2 Medium, across five seeds ($n{=}15$ paired runs), CCG achieves $\CFS=5.654\pm0.625$, outperforming ROME-style tracing ($3.382\pm0.233$), SAE-only ranking ($2.479\pm0.196$), and a random baseline ($1.032\pm0.034$), with $p<0.0001$ after Bonferroni correction. Learned graphs are sparse (5-6\% edge density), domain-specific, and stable across seeds.

Problem

• causal concept graphs
• stepwise reasoning
• latent space
• concept interaction
• language models

Innovation

• Causal Concept Graphs
• Sparse Autoencoders
• Differentiable Structure Learning
• Causal Fidelity Score
• Stepwise Reasoning