Challenge: Existing prompt-based debiasing methods exhibit instability due to sensitivity to prompt changes . fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting .
Approach: They propose a debiasing framework that encodes fairness-related features into separable directions in the hidden activation space.
Outcome: The proposed framework performs inference-time debiasing without requiring retraining or prompt design . it detects bias signatures in activations and then computes debiased steering vectors . the proposed framework is available to download in the u.s.

Similar Papers

Fairness Evaluation and Inference Level Mitigation in LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Large language models display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, and the propagation of unwanted patterns during extended dialogues.
Approach: They propose a pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation.
Outcome: The proposed framework detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation.
Debiasing LLMs by Masking Unfairness-Driving Attention Heads (2026.findings-acl)

Copied to clipboard

Challenge: Existing work probes when biased outputs appear, but gives little insight into the mechanisms that generate them, leaving existing mitigations largely fragile.
Approach: They propose a lightweight debiasing framework that detects bias heads and selectively masks only those heads that activate under DA and CoT.
Outcome: The proposed framework reduces unfairness by 391.9%- 534.5% in both one- and two-turn dialogues.
“Thinking” Fair and Slow: On the Efficacy of Structured Prompts for Debiasing Language Models (2024.emnlp-main)

Copied to clipboard

Challenge: Existing debiasing techniques are typically training-based or require access to the model’s internals and output distributions, so they are inaccessible to end-users looking to adapt LLM outputs for their particular needs.
Approach: They propose a system-based iterative framework that uses System 2 thinking processes to induce logical, reflective, and critical text generation with single, multi-step, instruction, and role-based variants.
Outcome: The proposed framework significantly improves over other frameworks demonstrating lower mean bias in the outputs with competitive performance on the downstream tasks.
Rethinking Prompt-based Debiasing in Large Language Model (2025.findings-acl)

Copied to clipboard

Challenge: Existing prompt-based methods for debiasing are often superficial and lack a thorough understanding of complex bias concepts.
Approach: They analyze a BBQ and stereoSet benchmarks to examine the assumption that large language models understand biases.
Outcome: The proposed model misclassified 90% of unbiased content as biased despite high accuracy on BBQ dataset . the proposed model may have been flawed in previous attempts to debiase .
BiasFilter: An Inference-Time Debiasing Framework for Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for debiasing large language models incur high human and computational costs and are limited in their effectiveness.
Approach: They propose a model-agnostic, inference-time debiasing framework that enforces fairness by filtering generation outputs in real time.
Outcome: The proposed framework mitigates social bias across a range of LLMs while preserving overall generation quality.
GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for fine-tuning large language models often ignore token-level causal influence and underutilize model logits.
Approach: They propose a novel approach that uses a gradient-based approach to identify influential tokens and construct directional steering vectors based on their contribution to preferred over dispreferred outputs.
Outcome: The proposed approach outperforms fine-tuning and prior steering methods on both LLM and VLM tasks without degrading fluency or general capabilities.
FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for inference-time steering fail to be effective, utility-preserving and training-efficient due to rigid, one-size-fits-all designs and limited adaptability.
Approach: They propose a steering framework that decomposes inference-time steering into two stages . they propose 'conditional steering' mechanism that preserves model utility by avoiding unnecessary steering . a 'mixture-of-Steering-Experts' mechanism captures multimodal nature of desired steering behaviors .
Outcome: The proposed framework outperforms the state-of-the-art methods on safety and truthfulness benchmarks.
Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs (2026.findings-eacl)

Copied to clipboard

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.
Fair-CCD: Mitigating Bias in Large Language Models for Tabular Classification Through Context-Contrastive Decoding (2026.acl-long)

Copied to clipboard

Challenge: Prior work to mitigate fairness issues often employs subjective demonstration selection, leading to low controllability and limited stability across different models and tasks.
Approach: They propose to use in-context learning to insert social biases into large language models to create a structured and controllable representation of the relationship between sensitive attributes and predicted labels.
Outcome: Extensive experiments show that Fair-CCD consistently improves fairness metrics without degrading task accuracy.
Inference-Time Selective Debiasing to Enhance Fairness in Text Classification Models (2025.naacl-short)

Copied to clipboard

Challenge: Several studies have investigated and promoted fairness, and a variety of definitions have been proposed to address this problem.
Approach: They propose a selective debiasing method that removes bias from model predictions instead of discarding them at inference time.
Outcome: The proposed method achieves better results than standard uncertainty quantification methods on text classification datasets with encoder-based classification models.

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