COFT: Counterfactual-Conformal Decoding for Fair Chain-of-Thought Reasoning in Large Language Models

Researchers introduce COFT, a training-free decoding method that reduces bias in large language models' chain-of-thought reasoning by 30-55% through counterfactual prompting and conformal calibration. The approach preserves task performance while adding minimal computational overhead, offering a practical solution for deploying fairer AI systems without model retraining.

COFT addresses a critical vulnerability in modern LLMs: the amplification of societal biases during reasoning processes. While chain-of-thought prompting improves model transparency and reasoning quality, it simultaneously exposes and intensifies latent biases embedded in training data. This research tackles fairness at the decoding stage rather than requiring expensive retraining, making it immediately applicable to existing deployed models.

The method's three-stage approach—counterfactual masking, logit fusion, and conformal calibration—represents a sophisticated engineering solution to fairness. By comparing factual and neutralized token distributions, COFT identifies attribute-driven bias amplification without needing auxiliary classifiers or model weight access. The conformal calibration component provides distributional validity guarantees, offering auditability that regulators and enterprises increasingly demand.

For the AI industry, this work signals maturation in fairness engineering. Rather than choosing between performance and equity, COFT achieves both: reasoning accuracies remain stable while bias metrics decline substantially. The 11% computational overhead is negligible compared to deployment costs, removing practical barriers to adoption. This matters for enterprises building AI systems subject to fairness regulations or ethical commitments.

The broader implication extends to trustworthy AI infrastructure. As LLMs become decision-support systems in lending, hiring, and legal contexts, bias mitigation moves from academic interest to operational necessity. COFT's training-free nature means organizations can improve existing model deployments immediately. Future work will likely extend these techniques to multimodal systems and evaluate cross-cultural fairness impacts, setting standards for responsible AI scaling.

• →COFT reduces bias metrics by 30-55% without model retraining, making fairness improvements immediately applicable to deployed systems.
• →The method adds only 11% computational overhead while preserving reasoning accuracy and task performance.
• →Conformal calibration provides distributional validity guarantees, enabling auditable and certifiable fairness claims.
• →Token-level fairness control operates at decode time, requiring no weight access or auxiliary classifiers.
• →Results validated across six models and multiple bias benchmarks, demonstrating generalizability.