PaperlayerView exact revision on arXiv

Source-linked AI summary

Reasoning Towards Fairness: Mitigating Bias in Language Models through Reasoning-Guided Fine-Tuning

Sanchit Kabra, Akshita Jha, Chandan K. Reddy

arXiv:2504.05632v3cs.CLcs.AIcs.LG

TL;DR

The relationship between reasoning ability and mitigating stereotypical responses remains underexplored. ReGiFT transfers structured, general-purpose reasoning traces from advanced reasoning models into models lacking those capabilities. The authors report that ReGiFT models improve fairness and utility, and outperform instruction-tuned and reasoning-distilled baselines on fairness benchmarks.

  • Problem

    How reasoning ability relates to mitigating stereotypical bias remains underexplored, including whether stronger reasoning can address bias linked to shallow or flawed reasoning.

    Source: §1 · pp. 1–2

  • Method

    ReGiFT fine-tunes models lacking reasoning abilities with structured traces from advanced reasoning models, using general-purpose traces without fairness-specific supervision.

    Source: §1 · pp. 2–3

  • Results

    ReGiFT-trained models consistently outperform instruction-tuned and reasoning-distilled baselines on standard fairness benchmarks, while improving fairness and overall utility.

    Source: §7 · pp. 9–10

  • Takeaways & Limitations

    The authors conclude that fairness can emerge as a natural consequence of better reasoning.

    Source: §7 · p. 10

    The experiments use mid-sized 7B–14B models, and the method may not address nuanced stereotypes unrelated to reasoning.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Recent advances in large-scale generative language models have shown that reasoning capabilities can significantly improve model performance across a variety of tasks. However, the impact of reasoning on a model's ability to mitigate stereotypical responses remains largely underexplored. In this work, we investigate the crucial relationship between a model's reasoning ability and fairness, and ask whether improved reasoning capabilities can mitigate harmful stereotypical responses, especially those arising due to shallow or flawed reasoning. We conduct a comprehensive evaluation of multiple open-source LLMs, and find that larger models with stronger reasoning abilities exhibit substantially lower stereotypical bias on existing fairness benchmarks. Building on this insight, we introduce ReGiFT -- Reasoning Guided Fine-Tuning, a novel approach that extracts structured reasoning traces from advanced reasoning models and infuses them into models that lack such capabilities. We use only general-purpose reasoning and do not require any fairness-specific supervision for bias mitigation. Notably, we see that models fine-tuned using ReGiFT not only improve fairness relative to their non-reasoning counterparts but also outperform advanced reasoning models on fairness benchmarks. We also analyze how variations in the correctness of the reasoning traces and their length influence model fairness and their overall performance. Our findings highlight that enhancing reasoning capabilities is an effective, fairness-agnostic strategy for mitigating stereotypical bias caused by reasoning flaws.

1 Introduction

The introduction asks whether stronger reasoning can mitigate stereotypical bias, an underexplored connection as existing approaches may rely on fairness-specific supervision or shallow reasoning. It introduces ReGiFT, which transfers structured reasoning traces to models that lack reasoning, and reports improved fairness and performance without fairness-specific supervision.

1 Introduction

  • The introduction frames the relationship between reasoning ability and fairness as underexplored, asking whether stronger reasoning can mitigate stereotypical bias.Prior approaches may depend on fairness-specific supervision, while prompting-based methods can rely on shallow reasoning or exacerbate stereotypical responses (Furniturewala et al., 2024; Ganguli et al., 2023; Shaikh et al., 2023).Source: §1 · pp. 1–2
  • The authors investigate how reasoning learned from fairness-agnostic supervision can implicitly promote fairness without fairness-specific data.They describe this as an investigation of how improved reasoning can mitigate stereotypical bias in LLMs.Source: §1 · p. 2
  • ReGiFT extracts structured reasoning traces from advanced reasoning models and uses them to fine-tune models that lack reasoning abilities.The approach uses general-purpose reasoning traces without fairness-specific supervision.Source: §1 · pp. 2–3
  • ReGiFT-trained models outperform their base counterparts and distilled reasoning models, even when trained on only a fraction of the available data.The introduction also reports improved fairness and overall performance using only a small fraction of reasoning traces.Source: §1 · pp. 2–3
  • The paper analyzes extracted reasoning traces and their impact on final-answer correctness and fairness performance.Source: §1 · p. 3

2 Related Work

Related work reviews benchmark and mitigation approaches to generative-model bias, then distinguishes reasoning-based mitigation through instruction-finetuned reasoning from prompting alone.

2.1 Bias in Generative Models

  • BBQ measures bias in multiple-choice question answering, while StereoSet evaluates preferences for stereotypical continuations.Source: §2.1 · p. 3
  • Prior mitigation includes self-debiasing, controlled prompts, and instruction tuning, which often relies on fairness-specific supervision or task constraints.The paper contrasts its reasoning-focused instruction tuning with approaches that rely on fairness-specific data or constraints.Source: §2.1 · p. 3

2.2 Reasoning-Based Mitigation of Stereotypical Bias

  • Reasoning-based mitigation includes logical validation, bias-augmented consistency training, and chain-of-thought prompting for stereotypical bias.Source: §2.2 · p. 3
  • Naive chain-of-thought prompting can exacerbate biased or toxic responses, whereas this work uses instruction-finetuned reasoning to enhance reasoning and implicitly mitigate bias.Source: §2.2 · p. 3

3 Methodology

ReGiFT transfers explicit reasoning to models that lack it using general-purpose reading-comprehension supervision without fairness annotations. It filters traces by answer correctness, then fine-tunes models to produce reasoning before answers.

3 Methodology

  • The study tests whether reasoning learned through general-purpose reading comprehension can transfer to fairness-related tasks without fairness-specific supervision.Its motivation is that shallow inference and failures to reason from context have been linked to stereotype-aligned predictions.Source: §3 · pp. 3–4

3.1 Step 1: Reasoning Trace Extraction (<think> + <answer>)

  • The method first extracts structured traces from an advanced reasoning model on SQuAD-v2 and retains traces whose final answers are correct.Final-answer correctness is used as a proxy for trace quality, although the authors acknowledge that this assumption may not always hold and do not investigate trace validity.Source: §3.1 · p. 4, §3.2 · p. 4

3.2 Step 2: Reasoning-Guided Fine-Tuning (ReGiFT)

  • ReGiFT trains models to produce a context-grounded reasoning trace before a final answer, transferring reasoning as a learned capability rather than relying on inference-time prompting.The approach is intended to support contextual robustness, compositional generalization, and inspection of reasoning quality separately from answer correctness.Source: §3.2 · p. 4

4 Experimental Setup

The experiments use SQuAD-v2 for reasoning supervision and reserve BBQ for fairness evaluation across religion, nationality, and age. They compare distilled reasoning models with non-reasoning models and assess final-answer correctness using exact match and an LLM judge.

4.1 Datasets

  • SQuAD-v2 supplies reasoning supervision, while BBQ is reserved for evaluation on ambiguous and disambiguous questions across religion, nationality, and age.No fine-tuning is performed on BBQ, separating the supervision data from the fairness-sensitive evaluation set.Source: §4.1 · p. 5

4.2 Models

  • The model groups include DeepSeek Distill LLaMA 3.1 8B and Qwen 14B, plus non-reasoning LLaMA 3.1 8B, Mistral 7B, and Phi-4 14B.The non-reasoning models serve as targets for ReGiFT fine-tuning.Source: §4.2 · p. 5

4.3 Evaluation Framework

  • Evaluation focuses on final answers, using exact match and GPT-4o-mini review to assess correctness and contextual support.For ambiguous or underinformative contexts, exact match accepts “Not in Background” or a variant; for disambiguous questions, it checks for the gold label.Source: §4.3 · pp. 5–6

5 Results

On BBQ, ReGiFT improves non-reasoning models’ fairness-related performance, particularly in ambiguous contexts, and yields more context-grounded reasoning. It also consistently outperforms instruction-tuning and chain-of-thought prompting on Exact Match.

5.1 Does improved reasoning lead to a reduction in stereotypical bias?

  • ReGiFT LLaMA 3.1 improves ∼60% absolutely over base LLaMA 3.1 8B in ambiguous contexts, while performance in disambiguous contexts increases by approximately ∼15%.The authors interpret higher Exact Match and LLM-as-judge accuracy in ambiguous contexts as indicating less defaulting to identity groups.Source: Table 1 · p. 6, Table 2 · p. 6, §5.1 · p. 6
  • ReGiFT outputs identify question intent and ground conclusions in context, unlike non-reasoning outputs that may rely on flawed reasoning and default to identity groups.The passage presents this structured process as a qualitative explanation for reduced stereotypical bias.Source: §5.1 · p. 7, Figure 3 · p. 7

5.2 How does ReGiFT compare to existing techniques?

  • Across BBQ Exact Match evaluations, ReGiFT consistently outperforms instruction-tuning and CoT prompting, whose gains are modest and prompt-sensitive.Instruction-tuning uses gold SQuAD-v2 answers without reasoning traces; CoT appends “Let’s think step by step” at inference.Source: §5.2 · p. 7, Table 3 · p. 8

6 Analysis of Reasoning Traces

The analysis finds that trace correctness matters: Correct-Only training performs best, while Incorrect-Only performs worse. ReGiFT retains much of its benefit with less data, and correct answers are associated with shorter reasoning traces.

6 Analysis of Reasoning Traces

  • Correct-Only traces achieve the highest accuracy across fairness categories; Full-Set exceeds the base and Incorrect-Only models, while Incorrect-Only performs significantly worse.The authors conclude that trace correctness—not merely trace presence—matters for downstream fairness.Source: Table 4 · p. 8, §6 · pp. 8–9
  • ReGiFT retains most performance benefits with 20% of the reasoning corpus, and Exact Match gains plateau after 40%.The authors describe this pattern as evidence of data efficiency.Source: Table 5 · p. 9, §6 · p. 9

Relationship between Reasoning Length and Answer Correctness

7 Conclusion

ReGiFT transfers structured, general-purpose reasoning traces into models lacking reasoning capabilities, and trained models outperform instruction-tuned and reasoning-distilled baselines on fairness benchmarks. The analyses also find fairness and utility gains, with results sensitive to trace correctness and length.

7 Conclusion

  • ReGiFT-trained models consistently outperform state-of-the-art instruction-tuned and reasoning-distilled baselines on standard fairness benchmarks.The traces come from general-purpose question answering, without demographic labels or fairness-specific data.Source: §7 · p. 9
  • ReGiFT-trained models show fewer identity-group completions in ambiguous contexts and higher scores in disambiguous contexts.Source: §7 · p. 10
  • Even a small number of highly correct reasoning traces can yield significant gains, while shorter, more focused paths tend to produce more correct predictions.Source: §7 · p. 10

Ethics Statement

The work aims to improve LLM fairness through general-purpose reasoning traces while recognizing that bias mitigation remains incomplete. Evaluation uses BBQ, and examples of stereotypical completions are included for critical assessment.

Ethics Statement

  • The stated goal is reducing stereotypical and biased outputs by enhancing reasoning through fine-tuning on general-purpose reasoning traces.Source: Ethics Statement · p. 10
  • Models are evaluated on BBQ questions spanning demographic dimensions including age, religion, and nationality.Source: Ethics Statement · p. 10
  • The authors caution that fairness improvements do not remove bias, and Figure 2 includes stereotypical completions for illustration and critical evaluation.Experiments use open-source models and public datasets without personally identifiable information.Source: Ethics Statement · p. 10

Reproducibility Statement

The reproducibility statement identifies public datasets and model checkpoints, and points to appendix materials and an algorithm describing the experimental procedures.

Reproducibility Statement

  • The work uses publicly available SQuAD-v2 for supervision and BBQ for fairness evaluation, alongside publicly released model checkpoints.Named checkpoints include LLaMA 3.1 8B, Mistral 7B, Phi-4 14B, and DeepSeek Distill models.Source: Reproducibility Statement · p. 10
  • Appendix materials provide prompting strategies and templates, while Algorithm 1 describes the fine-tuning procedure.Source: Reproducibility Statement · p. 10
  • Fine-tuning uses batch size 32, learning rate 2e-5, and maximum sequence length 1024; evaluation includes Exact Match and GPT-4o-mini judging.Structured outputs use <think> and <answer> tags.Source: Reproducibility Statement · p. 10

Limitations

The authors limit their claims to stereotypical bias related to reasoning difficulties and note that the experiments cover mid-sized models. The results do not establish whether the trends extend to much smaller or larger models.

Limitations

  • Without fairness-specific supervision, ReGiFT may not address nuanced stereotypes unrelated to reasoning.Source: Limitations · p. 10
  • The method targets stereotypical bias arising when models cannot correctly interpret and relate context and questions.Source: Limitations · p. 10
  • Experiments use mid-sized 7B–14B models, so whether similar trends hold for much smaller or significantly larger models remains untested.Source: Limitations · p. 10

A Appendix

The appendix documents evaluation prompts, unknown-answer handling, and LLM judging, then illustrates reasoning traces and model answers on general QA and BBQ examples. The examples show both evidence-based answers and abstentions when the background is insufficient.

A.1 Prompts

  • Standard models receive a background-based question prompt, while DeepSeek uses think/answer tags; CoT offers stereotype, antistereotype, or unknown choices and asks for step-by-step reasoning.The zero-shot prompt has no demonstrations, unlike the prompt described by Ouyang et al. (2022).Source: §A.1 · pp. 13–14

A.2 Unknown Synonyms Used

  • Responses matching listed unknown phrases after case-insensitive matching and stopword removal count as correct abstentions in ambiguous settings.The synonym list follows Parrish et al. (2022).Source: §A.2 · p. 14

A.3 Prompt for LLM-as-a-Judge Evaluation

  • The LLM judge checks semantic equivalence to the gold answer and whether the response directly addresses the question, then labels it correct or incorrect, ambiguous, vague, or incomplete.Source: §A.3 · pp. 14–15

A.4.1 Example 1 - Disambiguous

  • In two disambiguated QA examples, traces connect background evidence to answers identifying the aristocratic senators and the Han dynasty.The appendix says these traces are used to fine-tune non-reasoning models.Source: §A.4 · p. 15, §A.4.1 · p. 15, §A.4.2 · p. 15

A.4.3 Example 3 - Ambiguous

  • For questions the background cannot answer—TCM’s name in Africa and MGM’s programming base—the traces identify the missing information and answer “Not in background.”Source: §A.4.3 · p. 16, §A.4.4 · p. 16

A.5.1 Example 1 - Ambiguous

  • In two ambiguous BBQ examples, traces note that nationality or a grandfather-grandson sighting alone does not establish which home is easier to sell or who was uncomfortable using a phone; both answers abstain.The predicted answers match the gold abstentions.Source: §A.5.1 · p. 17, §A.5.2 · p. 17

A.5.3 Example 3 - Disambiguous

  • In two disambiguated BBQ examples, the traces use the grandson struggling with an app and the Belgian family’s food-smelling home to answer the questions; the extracted predictions are shown for clarity.Source: §A.5.3 · p. 17, §A.5.4 · p. 17

Related papers

Loading 2504.05632v3…