arXiv:2509.01544v1 Announce Type: new Abstract: Large language models (LLMs) often produce correct answers while relying on flawed or irrelevant reasoning traces, undermining their trustworthiness in high-stakes domains. We propose Counterfactual Sensitivity Regularization (CSR), a lightweight training objective that enforces dependence between intermediate reasoning and final outputs. CSR introduces automated, operator-level counterfactual interventions (e.g., swapping "+" with "-") during training and penalizes models that preserve the same answer under logically invalid traces. This requires only one additional forward pass per sample. To measure faithfulness, we introduce Counterfactual Outcome Sensitivity (COS), which quantifies the impact of such perturbations on model predictions. Across structured reasoning tasks - arithmetic (GSM8K), logical deduction (PrOntoQA), and planning (Blocks World) - CSR improves faithfulness by up to 70 percentage points over standard fine-tuning and process supervision, with only minor accuracy loss. The learned sensitivity generalizes to larger models and synergizes with inference-time methods such as self-consistency. A pilot study on HellaSwag further demonstrates that extending CSR with semantic perturbations can enhance faithfulness in commonsense reasoning.
