Research map / Large language models / Instruction and fine-tuning

Collapsed Language Models Promote Fairness

ICLR 2025 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that debiased BERT models exhibit more neural collapse (alignment of token representations with word embeddings) on gender words, and adds a NC3-based regularizer that improves fairness of several debiasing methods while preserving GLUE performance.

Problem

Debiasing pretrained language models lacks a principled, shared understanding and a method that consistently improves fairness across debiasing approaches without sacrificing NLU performance.

Approach

Views next-token prediction as classification and uses neural collapse as a lens: debiased LMs show greater alignment between class means of token representations and word embeddings for gender words, so the authors add an explicit NC3 (uniform duality) regularizer during fine-tuning as a plug-and-play fairness term.

Key results

• Adding NC3 regularization to Mabel raises StereoSet ICAT from 73.07 to 74.55 and reduces WinoBias TPR-1 from 23.73 to 16.93.
• On BEC-Pro, the female-male association difference for Mabel drops from 0.0404 to 0.0077 with NC3 regularization.
• NC3 regularization keeps GLUE performance close to each baseline (e.g., BEC 55.5 vs 55.7 CoLA MCC).

Measured effects reported in the paper

Surprising findings

• Neural collapse metrics NC1, NC2 and NC4 do not consistently separate debiased from biased models; only NC3 is consistently reduced in debiased models.
• Mabel+NC3 slightly lowers StereoSet language modeling score (84.80 to 83.55).

Limitations

• Gains on some metrics are small or mixed, e.g., Bias-NLI improvements are limited and Mabel's LM score drops on StereoSet.
• Experiments focus on BERT-scale encoders and gender bias; generalization to large LLMs and other bias dimensions is not shown.
• NC metrics are inconsistent across the vocabulary; the analysis relies on a curated gender word list.

Experimental setup

Datasets: SNLI MNLI OntoNotes 5.0 TinyStories PANDA News-Commentary v15 GAP corpus Bias-in-Bios GLUE Bias-NLI

Benchmarks: StereoSet BEC-Pro WinoBias Bias-in-Bios Bias-NLI GLUE

Metrics: LM score SS ICAT TPR-1 TPR-2 GAP_TPR GAP_RMS accuracy NN FN

Baselines: Mabel ASE BEC Sent-Debias Context-Debias FairFil Adept FairBERTa BERT+Dropout BERT+CDA

Builds on: neural collapse Mabel ASE BEC Sent-Debias

Closest papers by meaning

• Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.90)
• LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models (ICML 2024 · similarity 0.88)
• Evaluating the Efficacy of Prompting Techniques for Debiasing Language Model Outputs (Student Abstract) (AAAI 2024 · similarity 0.87)
• All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation (AAAI 2024 · similarity 0.87)
• Linguistic Collapse: Neural Collapse in (Large) Language Models (NeurIPS 2024 · similarity 0.86)
• Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models (NeurIPS 2025 · similarity 0.86)