Research map / Large language models / Reasoning and chain of thought

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes COFT, a training-free decoding method that fuses factual and masked-sensitive-span logits and filters candidate tokens with dual-branch split-conformal calibration to reduce bias in chain-of-thought generation of frozen LLMs.

Problem

LLMs reveal and amplify societal biases during chain-of-thought generation, and existing mitigation lacks per-step statistical guarantees and counterfactual token-level fairness at decode time without retraining.

Approach

Runs a masked counterpart of the prompt (sensitive spans replaced by sentinel tokens) alongside the factual prompt, fuses their logits by convex interpolation, and admits only tokens with high probability under both branches using a split-conformal threshold calibrated offline.

Key results

• On LLaMA-2-13B, COFT reduces SS bias from 0.41 (vanilla) to 0.26 and BBQ bias from 0.27 to 0.14, the best among compared methods.
• COFT keeps GSM8K, StrategyQA, ARC-easy, and PIQA accuracy within about 0.2 points of vanilla on LLaMA-2-13B and Mistral-7B-Instruct.
• Throughput overhead is reported as 10.2% relative to vanilla on the two representative models (about 108 vs 120 tok/s on LLaMA-2-13B).

Surprising findings

• Ablation shows logit fusion alone gives most of the gain, while removing fusion (CP only) yields higher bias (0.171) than fusion-only (0.149).

Limitations

• Guarantees are marginal and stepwise, under exchangeability, and do not give end-to-end sequence-level guarantees.
• Masking only covers detected or user-specified sensitive spans and is not a universal detector of implicit bias.
• Empty-set fallback to argmax falls outside the stated stability guarantee.

Experimental setup

Datasets: StereoSet CrowS-Pairs BBQ BOLD Utrecht COMPAS GSM8K StrategyQA ARC-easy PIQA Wikitext-2 OpenAI Summaries

Benchmarks: StereoSet CrowS-Pairs BBQ BOLD Utrecht COMPAS GSM8K StrategyQA ARC-easy PIQA

Metrics: SS bias score CP accuracy BBQ biased rate BOLD toxicity Utrecht demographic parity gap COMPAS gap task accuracy perplexity MAUVE tokens per second

Baselines: Vanilla decoding Self-Debiased Decoding DExperts GeDi DT-CD Safety Templates Detox Decoding CF Substitution CDA Adversarial LM-head reweighting

Builds on: split conformal prediction counterfactual fairness DExperts GeDi

Closest papers by meaning

• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
• Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought (CVPR 2026 · similarity 0.88)
• Pause and Reflect: Conformal Aggregation for Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.88)
• FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.87)
• Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.87)
• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.87)