Researchers developed Debias-SparseGPT, a post-training pruning method that adds a representational-debiasing term computed over demographically contrasting inputs to the standard SparseGPT pruning objective. Tested across multiple generative language models at 25%, 50%, and 2:4-structured sparsity levels, the method consistently reduces pruning-induced bias compared to standard SparseGPT while preserving perplexity and zero-shot accuracy. At the most aggressive sparsity setting, augmenting the calibration data with long-context examples further improved both fairness and task performance.
