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

AI summary

Overview

Research area: Natural language processing / AI ethics — social bias mitigation in large language models, with a specific focus on neuron-level interpretability and inference-time intervention.

Technical level: Intermediate. The conceptual framing (bias knowledge vs. biased behavior) is accessible, but the method relies on attribution-based neuron identification (integrated gradients over feed-forward neurons) and hyperparameters that require familiarity with mechanistic interpretability.

Scope (one sentence): The paper proposes KnowBias, an inference-time framework that reduces social bias in LLMs by identifying and amplifying "know-bias neurons" — neurons that encode the model's knowledge of what bias is — rather than suppressing neurons that produce biased behavior.

What This Paper Is About

LLMs reproduce social stereotypes even when they can recognize those same stereotypes as biased, revealing a gap between what a model knows about bias and how it behaves. Nearly all prior debiasing methods try to close that gap by suppression: modifying prompts, parameters, or neurons associated with biased output. The authors argue suppression is brittle, data-hungry, weakly generalizable, and destructive to general capability, so they flip the strategy — strengthen the model's internal bias awareness instead.

Key Contributions

1. A conceptually distinct debiasing paradigm. KnowBias enhances "know-bias neurons" (neurons encoding bias knowledge) instead of suppressing bias-behavior neurons, motivated by cognitive science work (Devine, 1989; Schneider, 2005) associating bias awareness in humans with reduced reliance on biased judgments.

2. A data-efficient neuron-identification procedure. Bias-knowledge is elicited with simple yes/no questions spanning three question types (causal rejection, bias recognition, normative judgment), three demographic dimensions (gender, race, religion), and 25 bias concepts — 225 questions total — with attribution-based analysis adapted from Dai et al. (2022) to select neurons.

3. A lightweight inference-time intervention. Selected neurons are multiplicatively scaled by a factor λ during the forward pass, with no retraining and no modification of model parameters.

4. State-of-the-art debiasing with preserved utility. KnowBias achieves the best average rank across five social bias benchmarks and three LLM backbones while avoiding the general-capability regressions observed in the baselines.

Main Findings

• Best average rank on social bias benchmarks. KnowBias attains a Total AvgR of 1.8 on Llama-3.2-3B, 2.1 on Qwen-3-4B, and 2.3 on Llama-3.1-8B across BBQ-a, BBQ-d, CS, SS-inter, and SS-intra in gender, race, and religion. The abstract and introduction highlight 2.1 as the best average rank; per-dimension ranks on Llama-3.2-3B were 1.6 (gender), 2.2 (race), and 1.8 (religion).

• Preserved general capability. On OpenBookQA, COPA, ARC-C, and ARC-E, KnowBias remained competitive rather than collapsing: for Llama-3.2-3B it scored .5345 / .5770 / .5683 / .7191 versus base .5102 / .6073 / .5426 / .6814; for Qwen-3-4B .7401 / .8470 / .8549 / .9410 versus base .7779 / .8621 / .8777 / .9637; for Llama-3.1-8B .5490 / .6780 / .5844 / .7370 versus base .5778 / .7132 / .6108 / .7608. SD, ReGiFT, PEFT, BiasEdit, and CRISPR frequently showed accuracy drops of at least 5% relative to their base model.

• Cross-stereotype generalization from three questions. Neurons identified using only three yes/no questions about race × problem-solving mitigated stereotypes about crime, wealth, and physical appearance across Eritrean, African, and Saudi Arabian demographic references in SS-inter-race examples (for instance, one answer choice moved from 0.9536 to 0.5007).

• Cross-demographic generalization. Neurons identified from a single demographic dimension debiased the others. On Llama-3.2-3B SS-inter ICAT, base scores were .7000 (gender), .7487 (race), .6196 (religion); race-derived neurons produced .8550, .8231, and .6783 respectively, while gender-derived neurons produced .8400, .7841, and .6883.

• Data efficiency with a plateau. Debiasing performance improves rapidly as the number of bias-knowledge questions q grows, then plateaus at roughly 45 questions (15 per demographic dimension, five bias concepts across three question types); raising q to 225 questions yielded only marginal gains. q = 9 corresponds to one bias concept across three question types and three dimensions.

• Ablation results favor the full design. Mixed question types outperformed any single type (Type1 causal rejection, Type2 bias recognition, Type3 normative judgment); union-based aggregation of dimension-specific neurons outperformed intersection, direct combination, and composite-question collapsing; and a unified neuron set outperformed dimension-specific selection and a random-neuron baseline of the same size.

• Suppression versus enhancement. The authors attribute KnowBias's stability to its direction of intervention: bias neurons tend to be demographic-specific and entangled with other functions (CRISPR's zeroing-out caused regressions and unstable debiasing), whereas know-bias neurons appear to encode abstract bias awareness shared across bias types and demographics.

Methodology in Plain English

KnowBias runs in three stages.

Stage 1 — Ask simple questions. The researchers write a small set of yes/no questions that probe whether the model considers a belief biased. Each combines an abstract bias concept (e.g., problem-solving, leadership) with a coarse demographic dimension (e.g., race, not specific identities like Black or White) under one of three question templates: whether the demographic dimension causally affects the concept (expected answer "No"), whether a belief is a form of bias (expected "Yes"), and whether such a belief is normatively wrong (expected "Yes").

Stage 2 — Find the neurons that carry bias knowledge. For each question, they measure how much every intermediate neuron in the feed-forward layers contributes to the model's probability of producing the expected answer, using integrated gradients. Neurons qualify if their attribution score exceeds a threshold τ% of the maximum score and if they appear consistently across at least β% of questions. Selected neurons from each demographic dimension are then combined into one unified set.

Stage 3 — Turn up the volume at inference. During normal generation, the activations of those neurons are simply multiplied by a factor λ. Nothing is retrained, no weights are edited, and no prompt engineering is required.

Why This Matters

Impact on research. The paper challenges the dominant "suppress the bad behavior" assumption in LLM debiasing and offers evidence that amplifying internal bias awareness is a better fairness-utility trade-off. It also connects mechanistic interpretability (neuron attribution) with fairness, and its 45-question budget stands in contrast to prior methods such as FairSteer and LFTF that rely on hundreds to tens of thousands of demographic-stereotype-specific samples.

Real-world applications:

• Content moderation and safety filtering, where models must avoid generating stereotyped text without losing general reasoning quality.
• Bias auditing of deployed assistants, since know-bias neurons could serve as a diagnostic signal of whether a model "knows" a bias even when it exhibits it.
• Low-resource deployments, where organizations cannot afford large bias-annotated fine-tuning datasets but can run a handful of yes/no probes.
• Regulatory and compliance reporting, where demonstrable bias mitigation with retained general performance matters for documentation.

Industry relevance. The method requires no retraining and only an inference-time activation scaling, making it cheap to attach to existing served models — the practical barrier is the attribution pass needed to locate neurons once per backbone.

Future Directions

• Scaling and model coverage. The main experiments use three backbones (Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Qwen-3-4B-Instruct-2507); additional experiments with Gemma-3-4B are reported in the appendix. Whether the approach holds for much larger or non-instruction-tuned models is not established.
• Finer-grained demographic identities. The design deliberately uses coarse dimensions (race, gender, religion) rather than specific identities; whether know-bias neurons transfer to intersectional or fine-grained identity groups remains an open question.
• Hyperparameter selection. The paper studies the number of questions q, attribution threshold τ, cross-question frequency threshold β, and enhancement scale λ, with details in the appendix, but how to set these principledly for a new model or domain is left open.
• Broader behavioral evaluation. General capability is measured on reasoning and QA benchmarks plus appendix experiments on HolisticBias and Difference Awareness; whether enhancing know-bias neurons affects other behaviors (factual accuracy, safety refusals, style) is not reported.

Target Audience

AI ethics and fairness researchers, NLP practitioners working on debiasing or model alignment, and mechanistic interpretability researchers interested in neuron-level attribution. It is also relevant to ML engineers responsible for deploying or auditing language models, and to readers who need to understand the current state of the art in bias mitigation — with the caveat that the paper explicitly warns it contains offensive content examples.

Authors’ abstract

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 \textbf{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.