Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning

Challenge:	Existing approaches to disentangle biased knowledge from reasoning are sub-optimal . entangled data makes curation difficult, leading to inclusion of sensitive, toxic data.
Approach:	They propose a framework that selectively removes biased knowledge while preserving reasoning abilities.
Outcome:	The proposed framework improves fairness accuracy by 14.7% and reasoning performance by 62.6% across multiple LLMs.

Similar Papers

Overview

Challenge

Approach

Outcome

Towards Safer Large Language Models through Machine Unlearning (2024.findings-acl)

Copied to clipboard

Zheyuan Liu, Guangyao Dou, Zhaoxuan Tan, Yijun Tian, Meng Jiang

Responsible & Trustworthy NLP

Challenge:	Existing work attempted to implement a gradient ascent based approach to prevent LLMs from producing harmful output when faced with problematic prompts.
Approach:	They propose a gradient ascent based approach to prevent LLMs from producing harmful output when faced with problematic prompts.
Outcome:	The proposed approach eliminates harmful knowledge while preserving utility on normal prompts.

Debiasing Large Language Models with Structured Knowledge (2024.findings-acl)

Copied to clipboard

Congda Ma, Tianyu Zhao, Manabu Okumura

Semantic Text Processing

Challenge:	Existing methods to reduce biases in pre-training models are hampered by their performance.
Approach:	They propose a method that utilizes structured knowledge to mitigate bias in LLMs . their method obviates the need for training from scratch, thus offering enhanced scalability .
Outcome:	The proposed method outperforms state-of-the-art (SOTA) baselines in the debiasing ability.

BiasFilter: An Inference-Time Debiasing Framework for Large Language Models (2025.findings-emnlp)

Copied to clipboard

Xiaoqing Cheng, Ruizhe Chen, Hongying Zan, Yuxiang Jia, Min Peng

Responsible & Trustworthy NLP

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.

Can Machine Unlearning Reduce Social Bias in Language Models? (2024.emnlp-industry)

Copied to clipboard

Omkar Dige, Diljot Arneja, Tsz Fung Yau, Qixuan Zhang, Mohammad Bolandraftar, Xiaodan Zhu, Faiza Khattak

Semantic Text Processing

Challenge:	Existing methods for mitigating bias in language models are expensive and time-consuming . comparative studies have not evaluated their respective advantages and disadvantages .
Approach:	They propose to use Partitioned Contrastive Gradient Unlearning and Negation via Task Vector to reduce social biases in open-source language models.
Outcome:	The proposed methods outperform PCGU and DPO in debiasing models . the proposed methods can be easily tuned to balance the trade-off between bias reduction and generation quality .

Mitigating Biases in Language Models via Bias Unlearning (2025.emnlp-main)

Copied to clipboard

Dianqing Liu, Yi Liu, Guoqing Jin, Zhendong Mao

Language Models

Challenge:	Recent debiasing approaches target different demographic groups, harming fairness and discrimination.
Approach:	They propose a model debiasing framework which targets stereotypes by unlearning stereotype forgetting and anti-stereotype retention.
Outcome:	The proposed framework outperforms existing methods in mitigating bias while retaining language modeling capabilities.

Neutralizing Bias in LLM Reasoning using Entailment Graphs (2025.findings-acl)

Copied to clipboard

Liang Cheng, Tianyi Li, Zhaowei Wang, Tianyang Liu, Mark Steedman

Reasoning

Challenge:	Natural Language Inference (NLI) is a foundational understanding task in language understanding.
Approach:	They propose a framework to construct counterfactual reasoning data and fine-tune LLMs to reduce attestation bias.
Outcome:	The proposed framework reduces hallucinations from attestation bias on original and bias-neutralized datasets while keeping hypotheses unchanged.

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

Copied to clipboard

Afrozah Nadeem, Mark Dras, Usman Naseem

Dialogue Systems & Conversational Agents

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.

UNLEARN Efficient Removal of Knowledge in Large Language Models (2025.findings-naacl)

Copied to clipboard

Tyler Lizzo, Larry Heck

Semantic Text Processing

Challenge:	Large Language Models excel in many tasks but are outperformed by specialized tools for certain tasks.
Approach:	They propose a method that uses subspace techniques to selectively remove knowledge . they propose 'unlearn' method that can forget or unlear the knowledge without retraining .
Outcome:	The proposed method outperforms existing methods for forgetting target knowledge while preserving related knowledge.

End-to-End Bias Mitigation by Modelling Biases in Corpora (2020.acl-main)

Copied to clipboard

Rabeeh Karimi Mahabadi, Yonatan Belinkov, James Henderson

Robustness in NLP

Challenge:	Recent studies have shown that strong natural language understanding models are prone to relying on unwanted dataset biases without learning the underlying task.
Approach:	They propose two learning strategies to train neural models that are more robust to dataset biases and transfer better to out-of-domain datasets.
Outcome:	The proposed methods improve robustness in all settings and transfer better to out-of-domain datasets.

DRPruning: Efficient Large Language Model Pruning through Distributionally Robust Optimization (2025.acl-long)

Copied to clipboard

Hexuan Deng, Wenxiang Jiao, Xuebo Liu, Jing Li, Min Zhang, Zhaopeng Tu

Responsible & Trustworthy NLP

Challenge:	Structured pruning reduces model size but often causes uneven degradation across domains, leading to biased performance.
Approach:	They propose a method that dynamically adjusts the data distribution during training to restore balanced performance across heterogeneous and multi-tasking data.
Outcome:	Experiments in monolingual and multilingual settings show that the proposed method surpasses similarly sized models in pruning and continued pretraining over perplexity, downstream tasks, and instruction tuning.