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

2026-05-28

AI Summary

Large language models often amplify social biases during chain-of-thought reasoning. This work proposes a training-free, weight-agnostic decoding-time debiasing method that uniquely integrates counterfactual intervention with distribution-free conformal prediction. By employing counterfactual prompt masking, lightweight logit fusion, and a dual-branch conformal calibration mechanism, the approach enforces token-level fairness controls throughout text generation. Evaluated across six models and multiple bias benchmarks, the method reduces bias metrics by 30–55% (median: 38%) while preserving task accuracy and linguistic quality, with computational overhead limited to at most one additional forward pass (≤11%). This framework thus enables verifiable, safe, and practical fair reasoning without compromising model performance.

Abstract

Large language models (LLMs) can reveal and amplify societal biases during chain-of-thought (CoT) generation. We present COFT (Chain of Fair Thought), a training-free decoding method that applies token-level fairness control at decode time, with distribution-free marginal validity guarantees (under exchangeability) for any frozen causal language model. COFT operates in three stages. First, it creates a masked counterfactual prompt by replacing sensitive spans with neutral tokens. Second, it compares the factual and masked logit distributions through lightweight logit fusion to attenuate attribute-driven biases. Third, it uses dual-branch split-conformal calibration to certify per-step candidate token sets at a user-chosen risk level. We evaluate COFT across six models and multiple bias benchmarks. Our method reduces standard bias metrics by 30-55% (median 38%) while preserving task utility and language quality. Reasoning accuracies remain unchanged within run-to-run noise margins. The computational overhead is modest, equivalent to one additional cached forward pass (<=11%). COFT offers a clear, auditable path to safer CoT generation with significant bias reduction, negligible utility loss, and no requirement for retraining, auxiliary classifiers, or weight access.

Problem

• bias
• fairness
• chain-of-thought
• large language models
• counterfactual

Innovation

• counterfactual decoding
• conformal calibration
• fairness in LLMs
• chain-of-thought reasoning
• bias mitigation