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

Self-correcting for Debiasing Large Language Models

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning

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 Self-Debias, a trajectory-level preference framework with a Jain's-fairness regularizer and consistency-filtered online self-improvement that teaches Qwen3-8B to self-correct social biases in chain-of-thought reasoning.

Problem

Social biases introduced early in a reasoning chain propagate and get rationalized through later CoT steps, and existing debiasing methods (static constraints, external filters, generic self-correction) neither detect nor interrupt this propagation without hurting reasoning utility.

Approach

Reformulates debiasing as corrective resource allocation: a suffix-level implicit reward margin (DPO-style) over biased vs. corrected trajectories, regularized by Jain's Fairness Index to lift hard samples, plus an SFT anchor and an online loop that mines unlabeled queries via bias injection and self-consistency filtering.

Key results

• Self-Debias Iter2 + Self-Correction reaches an average of 82.1 across eight benchmarks, versus 70.4 for DeepSeek-R1-Distill-Qwen-7B and 77.5 for base Qwen3-8B.
• Base Qwen3-8B loses 13.5 average points under self-correction, while Self-Debias gains +0.3 to +0.5 per stage under self-correction.
• On Llama-3.1-8B-Instruct, the average rises from 52.3 to 81.4 with the Self-Debias pipeline.

Surprising findings

• Generic response-wise self-correction (Self-Refine, Confirmation, Denying) and fairness filters degrade performance rather than improving it, including collapse on CEB tasks for base models.
• An SFT-only control (SFT(V0)) collapses by 34.1 points under self-correction, showing that the self-correction objective, not distribution fitting, drives stability.

Limitations

• Results are reported on a single fairness-focused training set (BBQ) with evaluation on a limited set of OOD benchmarks.
• Statistical significance is claimed for a single configuration, with limited details on the test.
• The theoretical analysis of the fairness regularizer and curriculum connection is heuristic rather than rigorous.

Experimental setup

Datasets: BBQ CrowS-Pairs

Benchmarks: BBQ UnQover CEB-Adult CEB-Credit CEB-Jigsaw CrowS-Pairs ARC-Challenge GSM8K

Metrics: accuracy average score

Baselines: DeepSeek-R1-Distill-Qwen-7B Qwen2.5-7B-Instruct Llama-3.1-8B-Instruct Confirmation Denying Self-Refine ReVISE CAL BiasFilter

Builds on: DPO Self-Refine Jain's Fairness Index STaR

Closest papers by meaning

• Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.90)
• Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.89)
• Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.88)
• Scalable Best-of-N Selection for Large Language Models via Self-Certainty (NeurIPS 2025 · similarity 0.88)
• Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.88)
• Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.88)