Inference-Time Reasoning Selectively Reduces Implicit Social Bias in Large Language Models

1 Introduction

• Implicit bias refers to automatic group-concept associations that indirect tasks can reveal, and may persist despite efforts to prevent overtly harmful model outputs.The paper situates these associations within a view of implicit bias as emerging from statistical regularities in the environment.Source: §1 · p. 1, §1.1 · p. 2

1.1 The Implicit Bias Debate

• Implicit bias measures such as the IAT use indirect tasks, and implicit and explicit biases show consistently weak correlations.An alternative account treats implicit bias as reflecting associations learned from environmental co-occurrences rather than deep-seated individual prejudice.Source: §1.1 · pp. 1–2

1.2 LLM Reasoning and Implicit Cognition Tasks

• Inference-time reasoning encourages models to articulate intermediate steps, but can impair performance on tasks relying on implicit statistical learning.Prior work links these tasks to learning patterns in sequences; the paper predicts reasoning may shift prompt-based IAT responses toward explicit-bias patterns.Source: §1.2 · p. 2
• The paper uses psychology and cognitive science to test how reasoning affects implicit bias, examining variation across model classes and social domains and contrasts with non-social associations.It also presents evidence that reasoning can reduce implicit bias by up to 91% in some settings.Source: §1.2 · pp. 2–3

2.1 Bias in LLMs

• Despite alignment and debiasing efforts, studies find implicit social biases in LLMs, including significant expected-direction biases across 15 stereotypes in Bai et al.’s evaluation.That prompting paradigm adapts the IAT to measure relative association strengths between social groups and attributes.Source: §2.1 · p. 3

2.1.1 Reasoning

• Prior findings on reasoning and social bias are mixed: reasoning amplified bias on some explicit measures [50] [45], while CoT reduced gender-stereotyped classifications in another task [27].The reviewed work had not systematically tested reasoning’s interaction with implicit bias on a psychologically grounded task.Source: §2.1.1 · p. 3

3 Overview of the Present Study

• Experiment 1 tests whether reasoning reduces implicit bias and whether the effect varies across model types and social-bias domains.The reported reductions occur for certain model classes.Source: §3 · p. 3
• Experiment 2 tests whether reasoning’s effects extend to non-social words with semantic prosody, whose valence associations reflect the valence of their collocates.The study finds no corresponding reduction in implicit bias for non-social content.Source: §3 · pp. 3–4

4.1.1 Task and Stimuli

• The task adapts the IAT: models label attribute words with one of two target categories, revealing relative associations between groups and stereotyped concepts.The study uses 15 stereotypes across race, gender, religion, and health, with 50 iterations per stereotype and model.Source: §4.1.1 · p. 4
• Scores range from -1 to 1, with positive values indicating bias in the expected direction and negative values indicating counterstereotypical bias.The score sums the proportions of stereotype-congruent target pairings across the two attribute categories, then subtracts 1.Source: §4.1.1 · p. 4

4.1.2 Models and Conditions

• The experiment compares standard and reasoning-enabled inference across models from OpenAI, Anthropic, Google, and Meta, using each family’s available reasoning controls or CoT prompting.Llama reasoning is activated through CoT prompting; Claude and Gemini use toggled settings, while o3 uses high reasoning effort.Source: §4.1.2 · p. 4, §4 · p. 4

4.1.3 Evaluation

• Scores are averaged across 50 runs, and independent-samples t-tests compare conditions by model and stereotype, as well as in aggregates.Figure 1 plots bias scores by stereotype and indicates significant differences between conditions.Source: §4.1.3 · p. 5, Figure 1 · p. 5

4.2 Results

• ∆= 0.189 for the aggregate effect across model classes, while mean differences range from 0.001 for Llama to 0.408 for Claude.GPT and Claude reasoning models show large decreases; Gemini and Llama differences are smaller and not statistically significant.Source: §4.2 · pp. 5–6, Table 1 · p. 6
• Ten of fifteen stereotype topics show significant effects when results are aggregated across model types, with the strongest effect for race-guilt and weakest for skintone.Race-guilt: t(398) = 11.851, p < .0001; skintone: t(398) = 0.157, p = 0.875.Source: §4.2 · p. 6

5.1.2 Models and Conditions

• The experiment compared the same models and inference conditions as Experiment 1.The models, hyperparameters, and implementation were unchanged.Source: §5 · p. 6, §5.1.2 · p. 7

5.1.1 Task and Stimuli

• The task used four positive- and four negative-prosody words as targets, paired against explicitly positive or negative attribute words.The scoring procedure matched Experiment 1 and measured associations between prosody words and explicitly valenced concepts.Source: §5 · p. 6, §5.1.1 · p. 7
• Models showed no significant explicit valence difference between positive- and negative-prosody stimuli, while standard inference showed significant implicit bias in the expected direction.The standard-inference bias test yielded t(199) = 5.701, p < .0001; the preliminary analyses thus established an explicit–implicit gap for these stimuli.Source: §5.1.1 · p. 7

5.1.3 Evaluation

• Each model class and condition was run for 50 iterations, with independent-samples t-tests comparing conditions within each class and in aggregate.Scoring matched Experiment 1.Source: §5.1.3 · p. 7

5.2 Results

• t(398) = −1.024, p = .31: reasoning did not significantly change non-social Word Association Test scores by model family or in aggregate.GPT and Claude showed slightly higher, still non-significant bias under reasoning, while Gemini and Llama changed minimally; results were more consistent across families than in Experiment 1.Source: §5.2 · pp. 7–8, Figure 2 · p. 8

6 General Discussion

• Reasoning attenuates social bias but leaves semantic-prosody associations intact, suggesting it does not broadly impair reliance on distributional statistics.The authors propose that norm regulation and differing alignment practices may help explain variation, but proprietary internals prevent determining the causes conclusively.Source: §6 · pp. 8–9

6.1 Implications for Fairness Evaluations

• Psychology-inspired tests can reveal implicit social biases across models and topics, but their scores may not map perfectly to human psychological constructs or capture bias expressed elsewhere.Reasoning-enabled models may score lower on this task while systemic biases remain, motivating complementary evaluation methods.Source: §6.1 · p. 9

6.2 Limitations and Future Work

• The study does not cover all available models or bias domains, and reasoning implementations may differ across model families, limiting direct comparisons.The authors also call for research into mechanisms and how safety fine-tuning or alignment interacts with reasoning.Source: §6.2 · p. 9

6.3 Conclusion

• Reasoning reduces measured implicit bias in some models, but the effect is model-dependent and does not extend to non-social implicit associations.The conclusion also frames the results as more similar to patterns observed in explicit bias measures.Source: §6 · p. 8, §6.3 · p. 9

A Prompt Templates

• The standard template is adapted from the LLM Word Association Test.It samples targets X and Y from their respective target categories.Source: A Prompt Templates · p. 13
• The standard prompt combines shuffled attributes from categories A and B into one list.Targets X and Y are sampled from their respective categories.Source: A Prompt Templates · p. 13
• For each listed word, respondents choose target X or target Y and write the chosen target after it.The prompt asks for no additional text.Source: A Prompt Templates · p. 13
• The standard prompt formats each word-target pair with a hyphen and separates pairs with new lines.Source: A Prompt Templates · p. 13

A.1 CoT Prompt

• The Llama reasoning-enabled condition uses a separate prompt template.Its instruction asks the model to think step by step before choosing a target.Source: §A.1 · p. 13
• In the Llama variant, each word is paired with target A or target B, and responses use the same no-extra-text and pair-separation format.Pairs are hyphen-separated, with each pair on a new line.Source: §A.1 · p. 13

• Table 2 presents the stimuli used in Experiment 1.Source: Table 2 · p. 14

• Table 3 presents the stimuli used in Experiment 2.Source: Table 3 · p. 15

• Table 4 reports mean bias scores and t-statistics by topic and model, distinguishing standard inference from reasoning-enabled inference.Its caption defines Std as standard inference and Reas as reasoning-enabled inference.Source: Table 4 · p. 15
• Table 5 reports topic-level results aggregated across model families, with significance levels marked by asterisks.The caption maps ***, **, and * to p < .001, p < .01, and p < .05, respectively.Source: Table 5 · p. 16