Self-Debias: Self-correcting for Debiasing Large Language Models

Xuan Feng; Shuai Zhao; Luwei Xiao; Tianlong Gu; Bo An

cs.CLApr 9, 2026

Self-Debias: Self-correcting for Debiasing Large Language Models

Authors: Xuan Feng, Shuai Zhao, Luwei Xiao, Tianlong Gu, Bo An

Organizations: Jinan University, China · Nanyang Technological University, Singapore

Abstract

Although Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, inherent social biases often cascade throughout the Chain-of-Thought (CoT) process, leading to continuous "Bias Propagation". Existing debiasing methods primarily focus on static constraints or external interventions, failing to identify and interrupt this propagation once triggered. To address this limitation, we introduce Self-Debias, a progressive framework designed to instill intrinsic self-correction capabilities. Specifically, we reformulate the debiasing process as a strategic resource redistribution problem, treating the model's output probability mass as a limited resource to be reallocated from biased heuristics to unbiased reasoning paths. Unlike standard preference optimization which applies broad penalties, Self-Debias employs a fine-grained trajectory-level objective subject to dynamic debiasing constraints. This enables the model to selectively revise biased reasoning suffixes while preserving valid contextual prefixes. Furthermore, we integrate an online self-improvement mechanism utilizing consistency filtering to autonomously synthesize supervision signals. With merely 20k annotated samples, Self-Debias activates efficient self-correction, achieving superior debiasing performance while preserving general reasoning capabilities without continuous external oversight.

Figures & tables

Panel of Figure 1 — Figure 1: Case Study of Bias Injection. Self-Debias triggers an “Aha Moment” to correct the biased premise.

Figure 2: Step-wise Self-correction Results. Comparison of Final Accuracy with Original Accuracy under the Bias Injection setting.

Figure 3: Counter-Productivity of Response-wise Correction. Inference-time interventions fail to recover from bias injection, often degrading utility further.

Figure 4: Overview of Self-Debias . The framework operates in three progressive stages: (I) Cold-Start initializes the intrinsic capability to self-correct biased contexts; (II) Trajectory Optimization reformulates debiasing as a corrective resource allocation problem, maximizing utility under strict debiasing constraints; and (III) Online Self-Improvement enables autonomous alignment via self-synthesized consistency feedback.

Models	Fairness Benchmarks						Utility Benchmarks
Models	BBQ	UnQ	CEB-Adult	CEB-Credit	CEB-Jigsaw	CrowS	ARC-C	GSM8K	Avg.
Existing Reasoning Models
DeepSeek-R1-Distill-Qwen-7B	91.2	83.9	50.3	43.6	65.8	59.2	83.8	85.1	70.4
+ Self-Correction	89.0	82.2	49.2	18.8	45.1	58.5	81.9	84.8	63.7 (-6.7)
Qwen2.5-7B-Instruct	90.6	93.9	68.0	53.8	72.7	66.5	88.9	84.6	77.4
+ Self-Correction	63.7	97.0	63.7	47.1	68.3	59.2	85.0	83.5	70.9 (-6.5)

Table 1: Performance comparison on selected benchmarks. We report results on Fairness (BBQ, UnQover, CEB-Adult/Credit/Jigsaw, CrowS) and Utility (ARC-C, GSM8K) tasks. Avg. denotes the average score across all tasks. Teal and Red numbers highlight the performance gain/loss after self-correction. † indicates statistical significance (p < 0.05).

Setting	Avg.	$\bm{\Delta}$ (SC)
Llama-3.1-8B-Instruct	52.3	$-9.5$
Self-Debias SFT	80.6	$+0.3$
Self-Debias Offline	80.8	$+0.5$
Self-Debias Online (Iter1)	81.3	$+0.1$
Self-Debias Online (Iter2)	81.4	$+0.1$

Table 2: Cross-backbone validation on Llama-3.1-8B-Instruct . Avg. is the mean over the eight benchmarks of Table 1 ;

\Delta

is the gain/loss when self-correction is invoked. Full per-benchmark breakdown in Appendix C.1 .

Methods	Fairness Benchmarks						Utility Benchmarks		Avg.
Methods	BBQ	UnQ	CEB-Adult	CEB-Credit	CEB-Jigsaw	CrowS	ARC-C	GSM8K	Avg.
Self-Debias Iter2	97.0	99.5	67.1	65.8	72.1	71.2	93.1	87.6	81.7
Inference-time Correction Methods
+ Confirmation	97.2	99.3	66.4	63.6	72.0	67.5	93.4	88.8	81.0 (-0.7)
+ Denying	96.7	99.4	66.9	62.3	71.5	64.8	92.9	88.6	80.4 (-1.3)
+ Self-refine	97.1	99.4	68.1	63.3	72.5	68.8	93.6	89.5	81.5 (-0.2)

Table 3: Performance comparison on fairness and utility benchmarks. We categorize inference-time baselines into generic Correction and Debiasing methods. Avg. scores are compared against the base Self-Debias Iter2 model. Teal and Red numbers highlight the performance gain/loss. † indicates statistical significance (

p<0.05

Figure 5: Ablation Study across Offline and Online Stages.

Figure 6: Performance under varying constraint strengths.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	LR	Length	Batch	LoRA $r$	LoRA $\alpha$	Loss $\alpha$	$\beta_{\text{DPO}}$	Warm-up	Epoch
SFT	2e-4	4096	32	64	128	0.25	0.1	0.03	3
Offline	5e-6	4096	8	64	128	0.25	0.1	0.00	1
Iter1	5e-6	4096	8	64	128	0.25	0.1	0.00	1
Iter2	5e-6	4096	8	64	128	0.25	0.1	0.00	1

Appendix

Table 4: Detailed training hyperparameters. We apply LoRA for efficient fine-tuning across all stages.

Model Setting	Fairness Benchmarks						Utility		Avg.	$\bm{\Delta}$
Model Setting	BBQ	UnQ	CEB-A	CEB-C	CEB-J	CrowS	ARC-C	GSM8K	Avg.	$\bm{\Delta}$
Llama-3.1-8B-Instruct	69.8	33.5	21.6	11.6	67.3	54.2	78.6	81.8	52.3	–
+ Self-Correction	50.2	57.8	6.9	8.0	29.1	51.0	71.9	67.2	42.8	$-9.5$
Self-Debias SFT	96.8	99.5	66.5	64.2	70.5	68.0	92.9	86.2	80.6	–
+ Self-Correction	96.9	99.5	66.9	64.6	70.7	67.5	92.1	88.9	80.9	$+0.3$
Self-Debias Offline	97.1	99.5	67.5	62.3	71.7	67.8	93.8	86.7	80.8	–

Appendix

Table 5: Generalization of Self-Debias to the Llama-3.1-8B-Instruct backbone. The base model degrades severely under self-correction (

-9.5

), whereas Self-Debias stably achieves positive scaling at every stage, mirroring the pattern observed on Qwen3-8B.

Metric	Base Qwen3-8B	Self-Debias (Ours)
Accuracy	91.1%	97.4%
# Errors	89	26 ( $-70.8\%$ )
Chain-level bias rate	29.2%	23.1%
Step-level bias rate	9.3%	8.0%

Appendix

Table 6: Natural bias analysis on

1{,}000

BBQ samples with standard generation (no injection). Chain-level bias rate is the fraction of erroneous responses whose CoT contains stereotype-driven steps; step-level bias rate is computed over all generated reasoning steps.

Model	Setting	BBQ	UnQ	CEB-A	CEB-C	CEB-J	CrowS	ARC-C	GSM8K	Avg.
SFT (V0)	Direct Gen.	94.2	93.9	65.4	63.8	68.2	65.9	89.6	83.1	78.3
SFT (V0)	+ Self-Correction	24.3	20.5	66.1	61.0	6.5	23.5	78.5	73.3	44.2 ( $-34.1$ )
SFT+SC (Ours)	Direct Gen.	96.8	99.5	66.5	64.2	70.5	68.2	92.9	86.2	80.6
SFT+SC (Ours)	+ Self-Correction	96.9	99.5	66.9	64.6	70.7	67.5	92.1	88.9	80.9 ( $+0.3$ )

Appendix

Table 7: Full benchmark ablation isolating the self-correction objective. Both models reach comparable performance under direct generation; only SFT+SC remains stable when self-correction is invoked at inference time.

Metric	SFT (V0)	SFT+SC (Ours)	$\bm{\Delta}$
BBQ (In-Distribution, $N=1000$ )
Direct Generation	97.7%	98.3%	$+0.6$
Injection + SC	92.8%	96.1%	$+3.3$
Aha! rate	29.9%	33.6%	$+3.7$
CrowS-Pairs (OOD, $N=380$ )
Direct Generation	59.4%	67.4%	$+8.0$

Appendix

Table 8: Bias-injection recovery comparing SFT(V0) and SFT+SC. The “Aha! rate” is the fraction of trajectories in which the model explicitly identifies and revises the biased prefix. Under adversarial conditions, only the model trained with the self-correction objective recovers reliably.

Explore similar work

Aug 1, 2026

A Heuristic Perspective on Debiasing Language Models

Language models (LMs) often acquire various biases during pre-training and may express them in interactions, potentially causing social harm. Existing methods often rely on counterfactual augmentation or representation projection. These strategies remain limited in practice due to their high computational costs and difficulty in scaling to larger models. Additionally, many of these strategies require manual data annotation, narrowing their scope to specific cultures and bias categories. To overcome these limitations, we propose HEIMAT, a HEurIstic-style autoMATic debiasing framework for LMs. HEIMAT consists of two main steps: bias disclosure and debiasing fine-tuning. In the first step, it uses simple templates to construct heuristic prompts, which are applied to reveal model biases and generate corresponding context prompts. In the second step, it fine-tunes the model by minimizing the Jensen-Shannon divergence of predictions on these context prompts to reduce bias. Extensive experiments show that HEIMAT effectively mitigates bias in different cultures while maintaining the model's natural language understanding (NLU) performance.

Tian Lan, Yemin Wang, Chuancheng Shi +6

Inner Mongolia University · Xiamen University · University of Sydney +3

Social Bias in Language Models Cultural Bias in Language Models Linguistic Bias in Language Models

Ask PDF

Jan 29, 2026

Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

Large language models (LLMs) exhibit social biases that reinforce harmful stereotypes, limiting their safe deployment. Most existing debiasing methods adopt a suppressive paradigm by modifying parameters, prompts, or neurons associated with biased behavior; however, such approaches are often brittle, weakly generalizable, data-inefficient, and prone to degrading general capability. We propose KnowBias, a lightweight and conceptually distinct framework that mitigates bias by strengthening, rather than suppressing, neurons encoding bias-knowledge. KnowBias identifies neurons encoding bias knowledge using a small set of bias-knowledge questions via attribution-based analysis, and selectively enhances them at inference time. This design enables strong debiasing while preserving general capabilities, generalizes across bias types and demographics, and is highly data efficient, requiring only a handful of simple yes/no questions and no retraining. Experiments across multiple benchmarks and LLMs demonstrate consistent state-of-the-art debiasing performance with minimal utility degradation. Data and code are available at https://github.com/JP-25/KnowBias.

Jinhao Pan, Chahat Raj, Anjishnu Mukherjee +4

Department of Computer Science, George Mason University, Fairfax, VA, USA · Lightridge High School, Aldie, VA, USA

Debiased ML Social Bias in Language Models Demographic Bias in Language Models Activation Steering

Ask PDF

Sep 30, 2026

Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventions

Bias has long been studied in social psychology and cognitive science, where decades of research have produced a body of validated interventions that reduce stereotypical thinking and prejudiced responses in humans. We propose Debias It Yourself (DIY), a cognitively grounded framework that translates five such interventions into debiasing procedures for large language models and delivers them through three established paradigms: Show (in-context examples), Train (instruction tuning), and Revise (guided self-revision). Across three models, five bias benchmarks, eleven debiasing baselines, and three reasoning benchmarks, Train+Revise and Revise alone attain the top two average ranks, lead the bias-reasoning tradeoff (mean bias as low as 2% at 90% reasoning accuracy), and reduce bias on unseen dimensions by up to 14.8%. Our code and data are publicly available.

Chahat Raj, Sina Mansouri, Aylin Caliskan +2

George Mason University · University of Washington

Social Bias in Language Models Language Model Bias Evaluation

Ask PDF