Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction
06:00 · July 20, 2026 · arXiv cs.AI RSS
Causal and intervention-based question answering is fundamental to advancing large language models (LLMs) toward reasoning beyond surface-level correlations and understanding underlying causal mechanisms. However, existing LLM-based methods often rely on implicit language-level reasoning, resulting in opaque causal assumptions, unverifiable reasoning paths, and fragile predictions under complex interventions, particularly in context-free settings. In this paper, we propose an explicit and auditable causal reasoning framework for context-free intervention-based question answering. Our method formulates causal inference as structured reasoning over an explicit causal graph through four modular stages, rather than implicit end-to-end prediction. A key innovation is a target-aware causal graph construction strategy that treats the target variable as a core constraint during graph expansion, effectively suppressing irrelevant variables, spurious causal relations, and reasoning noise. We further introduce a path-level causal evidence aggregation mechanism that combines multiple causal paths while modeling both reinforcing and counteracting effects, enabling robust decision-making beyond single-chain reasoning. Extensive experiments on three benchmarks demonstrate that our framework consistently outperforms existing LLM-based methods while providing interpretable and auditable causal reasoning traces.
Summary
Causal-Audit reframes context-free intervention-based question answering as structured inference over an explicit causal graph rather than implicit token-level generation. In this setting, models receive no supporting passages and must therefore construct and evaluate plausible causal mechanisms to determine how an intervention on one variable would affect a designated target outcome. The framework decomposes the task into four sequential stages: extraction of intervention, target, and modifier variables from the query; iterative construction of a target-aware causal graph that constrains expansion to variables and relations relevant to the target; extraction and counterfactual auditing of candidate causal chains; and aggregation of path-level evidence that accounts for both reinforcing and opposing effects.
A central design choice is the target-aware graph construction step, which uses the target variable as an explicit constraint during expansion. This mechanism limits the inclusion of spurious variables and semantically plausible but causally irrelevant edges that commonly arise when language models generate graphs without structural guidance. Once candidate chains are identified, each is validated for consistency under hypothetical interventions before being passed to the aggregation stage. The resulting process produces an auditable trace of the variables, directed relations, and signed influences that support the final decision.
By replacing end-to-end prediction with modular, inspectable stages, the approach converts large language models from opaque reasoners into constrained evaluators that operate on symbolically represented causal structure. Experiments across three established benchmarks show consistent gains in both predictive accuracy and the ability to surface verifiable reasoning paths compared with prior LLM-based methods that rely on unstructured chain-of-thought or direct prompting.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners focusing on trustworthy and explainable AI. Its emphasis on auditable reasoning traces directly aligns with the transparency requirements of the EU AI Act, offering a practical framework for deploying reliable LLMs in high-stakes enterprise environments.
