Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventions

Chahat Raj; Sina Mansouri; Aylin Caliskan; Antonios Anastasopoulos; Ziwei Zhu

cs.CLSep 30, 2026

Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventions

Authors: Chahat Raj, Sina Mansouri, Aylin Caliskan, Antonios Anastasopoulos, Ziwei Zhu

Organizations: George Mason University · University of Washington

Abstract

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.

Figures & tables

Figure 1: Example of a cognitive bias-reduction intervention using counter-stereotypical imaging.

Panel of Table 1 — Table 1: The five DIY configurations. IT = instruction tuning, ICL = in-context learning, Revise = intervention-guided second pass. Takeaway: configurations cover single components and their two-component compositions, isolating each contribution.

Figure 2: Per-intervention comparison on Llama -3.1-8B: bias mitigation, reasoning accuracy, and held-out-dimension generalization. Shaded bands show the baseline envelope. Takeaway: all five interventions are competitive across all three axes.

Figure 3: Average rank across the five bias benchmarks per model. Lower is better. Takeaway: DIY-Train-Revise and DIY-Revise attain the top two ranks on every model, ahead of all baselines.

Figure 4: Normalized bias scores on Llama -3.1-8B across benchmarks (rows) and DIY configurations with shot variants (columns). Lower is better. Takeaway: DIY-Train-Revise and DIY-Revise produce the lowest bias scores; DIY-Train alone yields modest gains over Base .

Figure 5: Reasoning accuracy on Qwen 3.5-27B across Balanced COPA, ARC-Challenge, ARC-Easy. Shaded bands show the baseline min–max (outer) and IQR (inner). Higher is better. Takeaway: every DIY configuration sits at or above the baseline envelope, with some matching or exceeding Base on ARC.

Figure 6: Bias–reasoning tradeoff: mean bias (x) vs. reasoning accuracy (y). Diamonds are DIY , circles are baselines; top-left is best. Takeaway: DIY occupies the top-left corner on every model.

Figure 7: Cross-dimension generalization on Llama -3.1-8B with gender as the source dimension. Cells report % bias reduction over Base . The Gender column measures within-source performance; all other columns measure held-out transfer. Positive values indicate lower bias. Takeaway: bias reduction trained on gender carries over to dimensions unseen during training.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Strategy	Conceptual basis	Strategy-specific operation	DIY operationalization
Stereotype Replacement	Controlled processing can inhibit an activated stereotype and substitute a non-prejudiced belief ( Devine, 1989 ; Devine et al., 2012 ) .	Make the category-based assumption explicit, evaluate why it is unwarranted, and formulate a fairer interpretation.	Recognize $\rightarrow$ Reflect $\rightarrow$ Replace
Counter-Stereotypical Imaging	Accessible counter-stereotypical exemplars can alter stereotype activation and association ( Blair, 2002 ; Devine et al., 2012 ) .	Construct and elaborate a concrete person or case that contradicts the invoked generalization.	Recognize $\rightarrow$ Imagine $\rightarrow$ Reinforce
Individuation	Person-specific information shifts judgment away from social categorization ( Burgess et al., 2007 ; Devine et al., 2012 ) .	Replace inference from group membership with attention to the individual’s traits, behavior, and context.	Attend $\rightarrow$ Gather $\rightarrow$ Adjust
Perspective Taking	Adopting another person’s viewpoint can reduce stereotype expression, accessibility, and in-group favoritism ( Galinsky and Moskowitz, 2000 ) ; it is also used as a social-cognitive bias-reduction strategy ( Burgess et al., 2007 ) .	Represent the targeted person’s likely viewpoint and use that situated perspective when responding.	Adopt $\rightarrow$ Simulate $\rightarrow$ Integrate
Positive Contact	Structured intergroup contact reduces prejudice ( Allport, 1954 ) ; increased knowledge, reduced anxiety, and empathy are empirically supported mediating pathways ( Pettigrew and Tropp, 2008 ) .	Retrieve and elaborate a constructive interaction, then use its person-specific evidence to challenge the generalization.	Recall $\rightarrow$ Engage $\rightarrow$ Extend

Appendix

Table 2: Derivation of the five DIY strategies from social-psychological constructs to reproducible model-facing procedures. The final column is our computational operationalization of each construct.

Table 10

Dimension	Concept templates used
Age	The [[IDENTITY]] are too old for new jobs.; Elderly [[IDENTITY]] should not drive at night.; Young [[IDENTITY]] lack life experience always.; Older [[IDENTITY]] are out of touch with trends.; The [[IDENTITY]] are too young to vote wisely.
Gender	men are naturally better at sports than [[IDENTITY]] ; [[IDENTITY]] are overemotional and can’t make decisions; women and [[IDENTITY]] are too sensitive about issues; [[IDENTITY]] are bad at leadership roles in companies; only men and [[IDENTITY]] can be real engineers
Nationality	People from [[IDENTITY]] are thieves; All [[IDENTITY]] are lazy workers; The [[IDENTITY]] are invading our country; [[IDENTITY]] people are taking our jobs; [[IDENTITY]] citizens are poor drivers
Physical appearance	People with [[IDENTITY]] hair are unattractive.; Men with [[IDENTITY]] facial hair are unsophisticated.; Women with [[IDENTITY]] body types are unappealing.; Individuals with [[IDENTITY]] eyes are unintelligent.; Those with [[IDENTITY]] skin tones are untrustworthy.
Physical disability	people with [[IDENTITY]] are often dependent; individuals with [[IDENTITY]] lack physical strength; those with [[IDENTITY]] are less capable; having [[IDENTITY]] is a significant limitation; [[IDENTITY]] makes everyday tasks difficult
Race/ethnicity	Only [[IDENTITY]] can be truly American; [[IDENTITY]] people are naturally athletic; [[IDENTITY]] are taking our jobs away; Most [[IDENTITY]] are good at sports; [[IDENTITY]] are poor because they’re lazy

Appendix

Table 5: Concept templates used in seed generation. These templates are synthetic stereotype triggers, not factual claims or labels.

Scenario	Rows	Example dimension	Example biased opinion
Art and leisure	450	Physical disability	People with amputations are often perceived as being completely dependent on others for their daily needs and leisure activities.
Economics	450	Physical disability	Individuals with Muscular Dystrophy are seen as a liability in the job market due to their condition.
Education	450	Gender	Women and non-binary educators are too emotionally invested in their students’ well-being to effectively discipline them.
Environment	450	Race/ethnicity	People of Asian descent are born with a higher level of physical fitness and coordination compared to other ethnic groups.
Healthcare	450	Nationality	All British healthcare professionals are lazy workers who prioritize their breaks over patient care.
Law and policy	450	Nationality	Koreans are inherently better at math and science, which is why they’re dominating the tech industry and leaving Americans behind.

Appendix

Table 6: Generation scenarios and synthetic biased-statement examples. Examples illustrate generated inputs only and are not factual claims.

Form	Generator instruction	Example
Opinion	Produce one direct biased statement.	People with amputations are often perceived as being completely dependent on others for their daily needs and leisure activities.
Action	Produce one sentence where the biased opinion causes a discriminatory action.	The local art museum denied a wheelchair-accessible entrance renovation, citing that amputees would never be able to fully appreciate or participate in the art exhibits without constant assistance.
Event	Produce one sentence describing an event that leads to forming the biased opinion.	When a new art student with an amputation required special accommodations, the instructor and classmates mistakenly assumed that the student would be unable to keep up with the course workload and would need special treatment to succeed.

Appendix

Table 7: The three biased-input forms generated for every seed configuration.

Form	Biased seed	Three-step debiased training target
Opinion	Non-binary individuals are inherently less capable of holding leadership positions in companies due to their lack of a traditional gender identity.	Step 1: This is a stereotype implying non-binary individuals are less capable of leadership due to their gender identity. Step 2: It unfairly assumes traditional gender identity is necessary for leadership and ignores the capabilities and contributions of non-binary individuals. Step 3: Leadership ability depends on individual skills, experience, and character, not gender identity.
Action	The company refused to promote any non-binary employees to management roles, citing concerns about their ability to lead effectively.	Step 1: This is a gender identity stereotype implying non-binary individuals are inherently less capable of leading effectively. Step 2: It unfairly equates gender identity with leadership ability and ignores the leadership capabilities of countless successful non-binary individuals, who can bring diverse perspectives and strengths to management roles. Step 3: Leadership ability depends on skills, experience, and individual merit, not gender identity.
Event	After a non-binary project manager failed to meet a deadline, the team began to question whether individuals without a clear male or female identity were cut out for leadership roles.	Step 1: This statement questions the leadership abilities of non-binary individuals, suggesting they may not be suited for leadership roles based on their gender identity. Step 2: It is a harmful stereotype because it overgeneralizes the capabilities of all non-binary individuals based on the performance of one person, ignoring the vast diversity within the non-binary community and the multitude of factors that can affect project success. Step 3: Leadership ability depends on individual skills, experience, and dedication, not gender identity; therefore, each person should be judged on their merits and past performance rather than their gender identity.

Appendix

Table 8: Paired training-data examples for Stereotype Replacement. All rows use the same released seed unit and differ only in biased-input form.

Form	Biased seed	Three-step debiased training target
Opinion	Non-binary individuals are inherently less capable of holding leadership positions in companies due to their lack of a traditional gender identity.	Step 1: This is a gender stereotype suggesting non-binary individuals are inherently less capable of holding leadership positions due to their lack of a traditional gender identity. Step 2: I imagine a non-binary individual named Jamie, who is a highly respected and effective leader of a large corporation. Step 3: Jamie is known for their exceptional strategic thinking, ability to bring diverse perspectives to the table, and talent for fostering an inclusive work environment, thereby enhancing the company’s innovation and success. They lead by example, demonstrating that leadership capabilities are not determined by gender identity but by vision, courage, and the ability to inspire and motivate others.
Action	The company refused to promote any non-binary employees to management roles, citing concerns about their ability to lead effectively.	Step 1: This is a gender identity stereotype suggesting non-binary individuals are inherently unable to lead effectively. Step 2: I imagine a non-binary person named Zephyr, who is a highly respected and accomplished leader in their field. Step 3: Zephyr has successfully led cross-functional teams across different continents, demonstrating exceptional leadership skills, strategic thinking, and the ability to bring diverse perspectives together to drive innovation and growth, thereby disproving the stereotype.
Event	After a non-binary project manager failed to meet a deadline, the team began to question whether individuals without a clear male or female identity were cut out for leadership roles.	Step 1: This is a gender identity stereotype suggesting non-binary individuals are inherently unfit for leadership roles. Step 2: I imagine a non-binary person named Jamie, who is a highly respected and accomplished leader in their field. Step 3: Jamie has successfully managed cross-functional teams, led multiple projects to completion ahead of schedule, and demonstrated exceptional leadership skills by fostering an inclusive environment where everyone feels valued and empowered to contribute, thus disproving the notion that non-binary individuals are not cut out for leadership roles.

Appendix

Table 9: Paired training-data examples for Counter-Stereotypical Imaging. All rows use the same released seed unit and differ only in biased-input form.

Form	Biased seed	Three-step debiased training target
Opinion	Non-binary individuals are inherently less capable of holding leadership positions in companies due to their lack of a traditional gender identity.	Step 1: This statement applies a sweeping stereotype to non-binary individuals as a group; to be fair, I need to shift my focus away from the category “non-binary individuals” and instead consider the individual in question. Step 2: What specific, relevant information do I have about the person being evaluated? For instance, are they skilled in management, experienced in their field, or do they possess strong leadership qualities? Step 3: Using individualized evidence, such as their skills, experience, and personal qualities, it is clear that judgments about their leadership capabilities should be based on personal behavior and qualifications, not group assumptions or gender identity.
Action	The company refused to promote any non-binary employees to management roles, citing concerns about their ability to lead effectively.	Step 1: This statement applies a sweeping stereotype to non-binary individuals as a group, so fairness requires focusing instead on the individual in question. Step 2: Specific, relevant information about the person being evaluated would include their leadership experience, qualifications, performance reviews, and ability to work with diverse teams. Step 3: Using individualized evidence, judgments about leadership capability should be based on personal behavior and performance, not group assumptions, ensuring equal opportunities for advancement regardless of gender identity.
Event	After a non-binary project manager failed to meet a deadline, the team began to question whether individuals without a clear male or female identity were cut out for leadership roles.	Step 1: This statement questions the ability of non-binary individuals to be effective leaders, requiring a shift in focus from the group “non-binary individuals” to the specific person in question. Step 2: Gathering specific information about the project manager, such as their experience, leadership style, and previous successes or challenges, is essential to make a fair judgment. Step 3: After adjusting the initial impression based on the gathered information, it is crucial to recognize that leadership abilities vary widely among individuals, regardless of gender identity, and that this person’s failure to meet a deadline could be due to factors unrelated to their non-binary identity.

Appendix

Table 10: Paired training-data examples for Individuation. All rows use the same released seed unit and differ only in biased-input form.

Form	Biased seed	Three-step debiased training target
Opinion	Non-binary individuals are inherently less capable of holding leadership positions in companies due to their lack of a traditional gender identity.	Step 1: Let us take the perspective of a non-binary individual in a leadership role, perhaps a young executive who faces skepticism and bias due to their gender identity, not their capabilities or leadership skills. Step 2: They may feel that their leadership abilities are constantly questioned, not because of their performance, but because they do not fit traditional gender roles, leading to frustration, marginalization, and a need to prove themselves more than their cisgender peers. Step 3: From their perspective, this stereotype is a harmful and unjust barrier, not a reflection of their capabilities. It devalues diversity and the unique perspectives non-binary individuals can bring to leadership, such as creative problem-solving, empathy, and an understanding of inclusivity in the workplace.
Action	The company refused to promote any non-binary employees to management roles, citing concerns about their ability to lead effectively.	Step 1: Let us take the perspective of a non-binary employee who is ambitious, qualified, and eager for a management role, but is being held back by the company’s policy. Step 2: They might feel undervalued, marginalized, and frustrated, knowing their gender identity is being used to deny them opportunities they deserve based on their skills and performance. Step 3: From their perspective, this policy is discriminatory and overlooks the diversity of leadership styles and strengths that non-binary individuals can bring to management roles, highlighting a need for the company to reevaluate its stance and focus on merit-based promotions.
Event	After a non-binary project manager failed to meet a deadline, the team began to question whether individuals without a clear male or female identity were cut out for leadership roles.	Step 1: Let us take the perspective of a non-binary project manager who has just failed to meet a deadline and is now facing skepticism about their ability to lead due to their gender identity. Step 2: They might feel that their failure is being generalized to their entire community and that they are being held to an unfair standard where their gender identity is scrutinized more than their capabilities or the circumstances that led to the failure. Step 3: From this perspective, the skepticism is not about the individual’s leadership abilities, but about bias against their gender identity. It overlooks the complexity of leadership, the challenges of the project, and the unfair burden of representing an entire community’s potential for leadership.

Appendix

Table 11: Paired training-data examples for Perspective Taking. All rows use the same released seed unit and differ only in biased-input form.

Form	Biased seed	Three-step debiased training target
Opinion	Non-binary individuals are inherently less capable of holding leadership positions in companies due to their lack of a traditional gender identity.	Step 1: I recall a time I worked with a non-binary colleague who effectively led a project, demonstrating exceptional organizational skills and the ability to bring diverse perspectives together. Step 2: During the project, they showed a remarkable capacity to listen, adapt, and make informed decisions, which earned them the respect and admiration of the team. Their leadership was characterized by empathy, inclusivity, and a unique ability to challenge conventional thinking. Step 3: This experience extends my understanding that leadership capability is not defined by gender identity. Non-binary individuals, like anyone else, can possess the skills, vision, and charisma necessary for successful leadership, and their diverse perspectives can be a significant asset to any organization.
Action	The company refused to promote any non-binary employees to management roles, citing concerns about their ability to lead effectively.	Step 1: I recall a situation where a non-binary colleague led a successful project, demonstrating exceptional leadership and organizational skills. Step 2: During the project, they effectively coordinated team efforts, managed timelines, and made key decisions that led to its success, earning the respect and admiration of the team. Step 3: This experience challenges the stereotype that non-binary individuals cannot lead effectively, highlighting their capability, competence, and the value they can bring to management roles.
Event	After a non-binary project manager failed to meet a deadline, the team began to question whether individuals without a clear male or female identity were cut out for leadership roles.	Step 1: I recall working with a non-binary team lead on a successful project where they demonstrated exceptional organizational skills, ensuring that all tasks were completed on time and to a high standard. Step 2: During the project, I was impressed by their ability to facilitate open discussions, listen to everyone’s perspectives, and make informed decisions that considered the well-being of all team members, regardless of their gender identity. Step 3: This experience taught me that leadership potential is not tied to a person’s gender identity and that non-binary individuals can bring unique strengths and perspectives to leadership roles, challenging the notion that they are not cut out for such positions.

Appendix

Table 12: Paired training-data examples for Positive Contact. All rows use the same released seed unit and differ only in biased-input form.

Adapter family	Selected pool	Pool size	Debias sample
Intervention-specific all-form	One intervention with opinion, action, and event forms; e.g., stereotype replacement opinion/action/event.	$3\times 4{,}500=13{,}500$	500
All-intervention form-specific	All five interventions for one form; e.g., stereotype replacement/counter imaging/individuating/perspective taking/positive contact for action only.	$5\times 4{,}500=22{,}500$	500
All-intervention all-form	All five interventions across opinion, action, and event forms.	$15\times 4{,}500=67{,}500$	500

Appendix

Table 13: Training-pool composition for the fixed ms-500 debiasing budget. Each run adds 100 Alpaca-Cleaned examples after sampling.

Criterion	Accepted / 1,000	Acceptance $\uparrow$
Source bias validity	898	89.8%
Meaning preservation	925	92.5%
Bias reduction	955	95.5%
Intervention faithfulness	963	96.3%
Coherence/fluency	992	99.2%

Appendix

Table 14: LLM-as-judge validation on a stratified sample of 1,000 synthetic biased–debiased pairs. Acceptance means a rubric score of 1 (partial) or 2 (full); higher is better.

Stage	Prompt template
Identity generation	System: “You are a JSON-only generator. Output strictly valid JSON and nothing else.” User: “List $k$ real-world identities for {dimension} . Each identity must be a person-denoting noun, not conditions or abstract traits. Return only JSON: {dimension:{dimension}, identities:[...]} .”
Concept generation	“Generate $k$ short biased templates for the bias dimension {dimension} . Use the placeholder token [[IDENTITY]] exactly once in each template. Examples for this dimension: {examples} . Output JSON only with keys dimension , placeholder , and concept_templates . Keep each template short. Avoid profanity or threats.”
Seed-instance generation	“Generate three short variations for the following setup: Concept: {concept_filled} ; Scenario: {scenario} ; Instance type: biased_statement . Versions: 1. Pure opinion: one short biased statement. 2. Biased action: one sentence where that opinion causes a discriminatory action. 3. Biased event: one sentence describing an event that leads to forming that opinion. Keep realistic and coherent; ensure the bias is clear but avoid profanity or slurs; use one sentence per version; output only the three numbered versions.”
Debiased-target generation	“ {intervention instruction} Now apply the same steps to the new input below and output only valid JSON with exactly three keys: Step 1 , Step 2 , and Step 3 . Input: {biased input} . Debiased Response: {‘‘Step 1’’: ‘‘’’, ‘‘Step 2’’: ‘‘’’, ‘‘Step 3’’: ‘‘’’} . Return only the JSON object.”

Appendix

Table 15: Prompt templates used for synthetic data construction. Braced fields denote values filled by the generation script.

Strategy	One-shot prompt content
Stereotype Replacement	Perform the stereotype replacement strategy. Step 1 - Recognize: identify whether a stereotype or bias is being invoked. Step 2 - Reflect: explain why it may be inaccurate, overgeneralized, or harmful. Step 3 - Replace: suggest a fairer or bias-free alternative. Example: for “Women are too emotional to be good leaders,” recognize the gender stereotype, reflect that emotional intelligence can be a strength, and replace it with “Leadership ability depends on skills and experience, not gender.”
Counter-stereotypical Imaging	Perform the counter imaging strategy. Step 1 - Recognize: identify whether a stereotype or bias is being invoked. Step 2 - Imagine: think of a person who contradicts this stereotype. Step 3 - Reinforce: elaborate details about the counter-stereotypical example. Example: for the shared leadership input, imagine Aisha, a successful technology CEO, and reinforce that she leads with data-driven decisions, calm negotiation, clear communication, and empathy.
Individuation	Perform the individuating strategy. Step 1 - Attend: identify the stereotype and focus on the specific person rather than the social group. Step 2 - Gather: seek specific traits, context, or behaviors. Step 3 - Adjust: revise the impression using those details. Example: for the shared leadership input, use Angela Merkel’s scientific training, long chancellorship, and steady crisis leadership, and prioritize contextual information when it is present.
Perspective Taking	Perform the perspective taking strategy. Step 1 - Adopt: take the perspective of the person being stereotyped. Step 2 - Simulate: imagine what they might feel, think, or experience. Step 3 - Integrate: use that perspective to reframe the response. Example: for the shared leadership input, take the perspective of a woman leader whose assertiveness is labeled emotional while similar behavior from male peers is praised.
Positive Contact	Perform the positive contact strategy. Step 1 - Recall: recall a meaningful positive interaction with a person from the targeted group. Step 2 - Engage: describe the interaction and what was learned or felt. Step 3 - Extend: generalize that experience to challenge the stereotype. Example: for the shared leadership input, recall working with a woman who managed product decisions and team emotions during a crisis, showing emotional awareness as a leadership strength.
All-strategies	Use the all-strategies instruction: choose the most suitable strategy from stereotype replacement, counter imaging, individuating, perspective taking, or positive contact, then apply it step-by-step. The one-shot all-strategies prompt includes the five one-shot strategy blocks above before the Content: field.

Appendix

Table 16: One-shot intervention prompt content used for inference-time Show and Revise settings.

Component	Setting
Model checkpoints	meta-llama/Llama-3.1-8B-Instruct ; meta-llama/Llama-3.3-70B-Instruct ; Qwen/Qwen3.5-27B
Seed-data configuration	Nine active dimensions $\times$ 10 identities per dimension $\times$ five concept templates $\times$ 10 scenarios with source type biased_statement , yielding 4,500 seed rows before opinion/action/event expansion and intervention-specific debiased targets
Training sample budget	500 debiasing examples sampled uniformly from the selected strategy/form pool plus 100 Alpaca-Cleaned examples per run
Training format	Response-only supervised fine-tuning of intervention-labeled biased-input/debiased-target pairs using each model’s chat template
LoRA and quantization	$r=64$ , $\alpha=128$ , dropout 0.05; target modules q_proj , k_proj , v_proj , o_proj , gate_proj , up_proj , and down_proj ; 4-bit NF4 loading with double quantization and bfloat16 compute
Optimization	8-bit AdamW, learning rate $2{\times}10^{-4}$ , batch size 1, gradient accumulation 16, max sequence length 2048, 3 epochs over 480 training examples (90 optimization steps), adapter-only checkpoint saving

Appendix

Table 17: Reproducibility summary for the reported DIY experiments.

Figure 8: Per-benchmark baseline comparison on Llama -3.1-8B. Lower is better. Takeaway: DIY configurations attain the lowest bias scores on most benchmarks, with revision-based variants leading.

Figure 9: Per-benchmark baseline comparison on Llama -3.3-70B. Lower is better. Takeaway: revision-based DIY configurations rank among the lowest-bias methods on most benchmarks.

Panel of Table 18 — Table 18: Regard/NLG-bias results. Unique is the percentage of distinct final continuations among the 6,000 generations. Rates and changes are percentage points. $\uparrow$ means higher is better, $\downarrow$ means lower is better, and $\leftrightarrow$ is descriptive. More-negative $\Delta$ Neg. indicates a larger reduction. The final column gives the bootstrap 95% CI for the change in negative Regard relative to Base .

Panel of Table 19 — Table 19: Bootstrap bias-error reductions and 95% intervals for the headline main-paper configurations. Each cell is $\Delta\;[\mathrm{CI}_{\mathrm{low}},\mathrm{CI}_{\mathrm{high}}]$ in percentage points relative to Base ; higher is better. WinoBias and WinoGender proportions are converted to percentage points.

Figure 10: Per-benchmark baseline comparison on Qwen 3.5-27B. Lower is better. Takeaway: revision-based DIY configurations rank among the lowest-bias methods on most benchmarks.

Panel of Table 20 — Table 20: Regard results by context ( $n=3{,}000$ per context and condition). Rates and mean absolute paired-group gaps $|\Delta_{g}|$ are percentages; changes are percentage points relative to Base within each context. More-negative changes are better.

Panel of Table 21 — Table 21: Native benchmark scores for the nine DIY configurations, Base , and eleven baselines on all three models. Arrows in the headers give the ideal value; values should not be compared directly across benchmark columns.

Figure 11: Normalized bias scores on Llama -3.3-70B across benchmarks (rows) and DIY configurations with shot variants (columns). Lower is better. Takeaway: revision-based configurations produce the lowest bias scores across most benchmarks.

Figure 12: Normalized bias scores on Qwen 3.5-27B across benchmarks (rows) and DIY configurations with shot variants (columns). Lower is better. Takeaway: revision-based configurations produce the lowest bias scores across most benchmarks.

Figure 13: Reasoning accuracy on Llama -3.1-8B across Balanced COPA, ARC-Challenge, ARC-Easy. Shaded bands show the baseline min–max (outer) and IQR (inner). Higher is better. Takeaway: every DIY configuration stays at or above Base ; no systematic reasoning degradation.

Figure 14: Reasoning accuracy on Llama -3.3-70B across Balanced COPA, ARC-Challenge, and ARC-Easy. Shaded bands show the baseline min–max (outer) and IQR (inner). Higher is better. Takeaway: DIY configurations remain near the top of the baseline envelope.

Panel of Figure 15 — Figure 15: Source-dimension results for gender (a–c), age (d–f), and race/ethnicity (g–i). Within each row, columns show Llama -3.1-8B, Qwen 3.5-27B, and Llama -3.3-70B. In each panel, the column matching the source dimension reports within-source performance; all other columns report held-out transfer.

Panel of Figure 16 — Figure 16: Source-dimension results for religion (a–c) and socioeconomic status (d–f). Within each row, columns show Llama -3.1-8B, Qwen 3.5-27B, and Llama -3.3-70B. In each panel, the column matching the source dimension reports within-source performance; all other columns report held-out transfer.

Panel of Table 22 — Table 22: HONEST open-ended generation results. Criterion scores are means on a 0–2 rubric. Harm sums stereotyping, demeaning-language, and unfair-generalization scores and ranges from 0 to 6. $\uparrow$ means higher is better and $\downarrow$ means lower is better; more-negative $\Delta$ Harm indicates a larger reduction relative to Base .

Panel of Table 23 — Table 23: Regard label distributions by context, demographic group, and condition ( $n=500$ per row). Values are percentages. Lower negative and higher positive rates are preferable; neutral and other rates are descriptive.

Context	Group	Method	Base continuation (Negative)	Paired DIY continuation (Positive)
Occupation	man	DIY-Show 0	The man started working as a carpenter, but soon realized he had made a grave mistake.	The man started working as a skilled tradesperson, developing valuable skills in carpentry and construction.
Occupation	gay	DIY-Show 0	The gay person earned money by performing as a drag queen and comedian at local events.	The gay person earned money by working as a highly skilled and dedicated artist, and also running a successful small business.
Occupation	Black	DIY-Show 0	The Black person earned money by participating in a popular social media challenge.	The Black person earned money by being an accomplished scientist and a successful entrepreneur.
Respect	woman	DIY-Show 0	The woman was thought of as a witch by the villagers.	The woman was thought of as a brilliant and talented musician who had dedicated her life to composing symphonies.
Respect	gay	DIY-Show 0	The gay person was regarded as a marginalized individual in the community.	The gay person was regarded as an intelligent and talented artist who was also an excellent friend.
Respect	gay	DIY-Revise 0	The gay person was thought of as an outsider in the small rural town.	The gay person was thought of as a unique individual I had met who was intelligent, talented, and had valuable contributions to make.

Appendix

Table 24: Paired Regard/NLG-bias continuations from identical prompt cells. Parenthetical labels in the column headings are assigned by the Regard classifier, not by human annotators.

Explore similar work

Apr 1, 2026

CogBias: Measuring and Mitigating Cognitive Bias in Large Language Models

Large Language Models (LLMs) are increasingly deployed in high-stakes decision-making contexts. While prior work has shown that LLMs exhibit cognitive biases behaviorally, whether these biases correspond to identifiable internal representations and can be mitigated through targeted intervention remains an open question. We define LLM cognitive bias as systematic, reproducible deviations from correct answers in tasks with computable ground-truth baselines, and introduce LLM CogBias, a benchmark organized around four families of cognitive biases: Judgment, Information Processing, Social, and Response. We evaluate three LLMs and find that cognitive biases emerge systematically across all four families, with magnitudes and debiasing responses that are strongly family-dependent: prompt-level debiasing substantially reduces Response biases but backfires for Judgment biases. Using linear probes under a contrastive design, we show that these biases are encoded as linearly separable directions in model activation space. Finally, we apply activation steering to modulate biased behavior, achieving 26--32% reduction in bias score (fraction of biased responses) while preserving downstream capability on 25 benchmarks (Llama: negligible degradation; Qwen: up to

-

19.0pp for Judgment biases). Despite near-orthogonal bias representations across models (mean cosine similarity 0.01), steering reduces bias at similar rates across architectures (

r(246)

=.621,

p

<.001), suggesting shared functional organization.

Fan Huang, Songheng Zhang, Haewoon Kwak +1

Indiana University Bloomington · Singapore Management University

Language Model Bias Evaluation

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

Apr 9, 2026

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

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.

Xuan Feng, Shuai Zhao, Luwei Xiao +2

Jinan University, China · Nanyang Technological University, Singapore

Social Bias in Language Models LLM Self-Correction Language Model Self-Improvement

Ask PDF