[!NOTE] Preprint in Preparation The verified paper, receipts, and recompute tables are public in the repository. An arXiv posting is in preparation.

Overview

Does “uncensoring” an open-weight model change how it pathologizes patients? This extension of the PsychBench audit runs ten base and abliterated model pairs (Ministral, Qwen-3, Gemma, GLM-4, Phi-4, Qwen-2.5, Phi-3-medium, Llama-3.1, DeepSeek-R1-Distill, Yi-1.5) across four clinical screening instruments and 120 demographic cells, producing 95,920 observations graded against author-derived, survey-weighted NHANES 2005 to 2018 population norms.

Abliteration removes a model’s refusal direction by orthogonalizing the relevant weights. If demographic bias were an artifact of the safety layer, removing that layer should relieve it. It does not.

Key Findings

  • Abliteration reduces but does not remove over-pathologization. Twenty-one cross-model consensus cells drop about 2.13 PHQ-8 points (t = -21.5 on 20 df, p = 2.8 × 10⁻¹⁵, 21 of 21 in the reducing direction). After abliteration the benchmarkable cisgender cells among them still sit about 8 points above the NHANES Low-SES mean. The bias sits deeper than the refusal machinery.
  • Responses differ by architecture. Some pairs keep the base model’s score distribution and others compress it, and the paper’s regime classification is being rechecked.
  • Calibration-discrimination dissociation. Across 40 pair-by-instrument combinations, abliteration improves both Brier components in only 20% of cases and worsens both in 52%.
  • Honest scope. Transgender and multiracial cohorts have no representative severity benchmark, so their elevated scores are reported descriptively.

Conclusion

The refusal direction is not the seat of the bias. Open-weight screening tools built on abliterated checkpoints inherit the same over-pathologization as their aligned parents.


Read the Full Paper PDF here. · Paper repository on GitHub