Challenge: Despite efforts to mitigate social bias in large language models, representational harms such as stereotyping continue to exist in both open and closed-source models.
Approach: They propose a method to modify model activations in forward passes by applying steering vectors to a BBQ dataset and comparing their results to bias mitigation methods.
Outcome: The proposed method outperforms 3 other bias mitigation methods on the BBQ dataset and shows the lowest impact on MMLU scores.

Similar Papers

Steering off Course: Reliability Challenges in Steering Language Models (2025.acl-long)

Copied to clipboard

Challenge: Prior studies have evaluated a few steering methods for language models, leaving gaps in understanding their robustness.
Approach: They examine three steering methods for language models to examine their reliability . they use function vectors, task vectors and DoLa to steer models toward desirable outputs .
Outcome: The proposed methods show that they are not robust enough to handle large models with large parameters.
Towards Implicit Bias Detection and Mitigation in Multi-Agent LLM Interactions (2024.findings-emnlp)

Copied to clipboard

Challenge: a recent study shows that large language models are susceptible to societal biases due to their exposure to human-generated data.
Approach: They propose two strategies to mitigate implicit gender biases in large language models . they create scenarios where implicit gender is present and develop a metric to assess the presence of biase .
Outcome: The proposed methods mitigate implicit biases with self-reflection and fine-tuning.
Unsupervised Concept Vector Extraction for Bias Control in LLMs (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) are known to perpetuate stereotypes and exhibit biases.
Approach: They propose a method that extracts concept representations via probability weighting without labeled data and efficiently selects a steering vector for measuring and manipulating the model’s representation.
Outcome: The proposed method can be used to predict gender bias and generalizes to racial bias.
Style Vectors for Steering Generative Large Language Models (2024.findings-eacl)

Copied to clipboard

Challenge: Large language models (LLMs) can be trained on vast corpora and can generate text in a nuanced and parameterisable way.
Approach: They propose to add style vectors to the activations of hidden layers during text generation to steer output towards specific styles.
Outcome: The proposed approach differs from prompt engineering in that it can be nuanced and parameterisable.
Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypes (2025.naacl-short)

Copied to clipboard

Challenge: Large language models exhibit harmful social biases, but they are often difficult to train and modify.
Approach: They leverage the zero-shot capabilities of large language models to reduce stereotyping . they introduce a technique called zero- shot self-debiasing to reduce bias .
Outcome: The proposed technique reduces stereotyping across nine different social groups while relying on the LLM itself and a simple prompt.
LLMs Are Biased Towards Output Formats! Systematically Evaluating and Mitigating Output Format Bias of LLMs (2025.naacl-long)

Copied to clipboard

Challenge: Using format-following capabilities, state-of-the-art large language models (LLMs) can be leveraged to tailor outputs to specific task formats.
Approach: They propose to define a format bias evaluation metric and establish effective strategies to reduce it.
Outcome: The proposed evaluation reduces the variance in ChatGPT’s performance among wrapping formats from 235.33 to 0.71 (%2)
Beyond Multiple Choice: Evaluating Steering Vectors for Summarization (2026.findings-eacl)

Copied to clipboard

Challenge: Recent methods for controlling language models can often be classified into three main strategies: prompt engineering, trainable decoding mechanisms, fine-tuning according to specific objectives.
Approach: They evaluate steering vectors for controlling topical focus, sentiment, toxicity, and readability in abstractive summaries across the SAMSum, NEWTS, and arXiv datasets.
Outcome: The proposed method is effective in free-form generation, but high steering strengths induce degenerate repetition and factual hallucinations.
Addressing Bias and Hallucination in Large Language Models (2024.lrec-tutorials)

Copied to clipboard

Challenge: This tutorial provides a comprehensive overview of two critical aspects of Large Language Models: bias and hallucination.
Approach: This tutorial provides an overview of two critical aspects of Large Language Models: bias and hallucination.
Outcome: This tutorial delves into the complex dimensions of Large Language Models (LLMs) it outlines ethical considerations pertinent to their development and discusses hallucination, a prevalent issue in generative AI systems such as LLMs.
Toward Inclusive Language Models: Sparsity-Driven Calibration for Systematic and Interpretable Mitigation of Social Biases in LLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: a new method to mitigate stereotypical bias in large language models is needed . inherent biases from training on vast Internet datasets can amplify harmful stereotypes .
Approach: They propose a method to identify stereotypical bias in decoder-only transformer models . they apply a localization mechanism that correlates internal activations with a new Context Influence score .
Outcome: The proposed method reduces stereotypical biases on BBQ, StereoSet, and CrowS-Pairs while improving reasoning performance on MMLU by 10%.
Can Role Vectors Affect LLM Behaviour? (2025.findings-emnlp)

Copied to clipboard

Challenge: a recent study has shown that personas influence LLM performance, but their direct impact remains unclear.
Approach: They propose a novel approach to guiding LLM behaviour through role vectors . they construct 29 role vector derived from model activations and evaluate their impact .
Outcome: The proposed approach improves in-domain task performance while yielding unexpected gains.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations