Debias-SparseGPT: Bias-Aware Pruning for Large Language Models
SparseGPT-style pruning can make persona-linked bias worse; this paper proposes a pruning method meant to counter that effect.
Debias-SparseGPT adds representational debiasing during post-training pruning, using demographically contrasting inputs. The authors report tests across multiple generative LLMs and sparsity settings, including 25%, 50%, and structured 2:4 sparsity. They say it reduces pruning-induced bias versus SparseGPT while preserving perplexity and zero-shot accuracy. The paper is accepted to EMNLP 2026 and lists code as available. HF Daily Papers' note
Debias-SparseGPT adds representational debiasing during post-training pruning, using demographically contrasting inputs. The authors report tests across multiple generative LLMs and sparsity settings, including 25%, 50%, and structured 2:4 sparsity. They say it reduces pruning-induced bias versus SparseGPT while preserving perplexity and zero-shot accuracy. The paper is accepted to EMNLP 2026 and lists code as available. HF Daily Papers' note
score 4