Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Meherab, Md Muntaqim, Mohammad, Noor Islam S., Feroz, Faiza
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908989975953408
author Meherab, Md Muntaqim
Mohammad, Noor Islam S.
Feroz, Faiza
author_facet Meherab, Md Muntaqim
Mohammad, Noor Islam S.
Feroz, Faiza
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2603_10377
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning
Meherab, Md Muntaqim
Mohammad, Noor Islam S.
Feroz, Faiza
Machine Learning
Artificial Intelligence
Methodology
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.
title Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning
topic Machine Learning
Artificial Intelligence
Methodology
url https://arxiv.org/abs/2603.10377