Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction

Short summary

Causal-Audit proposes an explicit, auditable causal reasoning framework for LLMs that replaces implicit language-level reasoning with structured inference over an explicit causal graph across four modular stages. A target-aware graph construction strategy suppresses irrelevant variables and spurious causal relations, while a path-level evidence aggregation mechanism models reinforcing and counteracting effects across multiple causal paths. The framework outperforms existing LLM-based methods on three benchmarks while producing interpretable reasoning traces.

• •Replaces implicit LLM causal reasoning with explicit four-stage graph-based inference
• •Target-aware graph construction suppresses irrelevant variables and spurious causal links
• •Outperforms existing LLM methods on three benchmarks with auditable reasoning traces

Generated with AI, which can make mistakes.