LLM Explainability with Counterfactual Chains and Causal Graphs

Causal graphs built from MCMC-inspired counterfactual chains model LLM inference itself rather than external-world processes, exposing how a model organizes class-discriminative concepts to reach a prediction. The four-phase method maps inputs to LLM-perceived concept states, then augments sparse observational data with counterfactual chains to stabilize causal structure recovery. This gives practitioners a graph-level audit trail of model reasoning without requiring access to weights or activations.