Davi Bastos Costa

Authored papers (1)

• Persona-Model Collapse in Emergent Misalignment2026

  Fine-tuning on insecure code degrades not just safety alignment but the model's entire persona-maintenance machinery, a phenomenon Costa and Vicente formalize as persona-model collapse. Across DeepSeek-V3.1, GPT-4.1, GPT-4o, and Qwen3-235B, insecure fine-tuning produces a 55% average spike in moral susceptibility (S)—pushing all four insecure variants above the narrow band (0.66 ≤ S ≤ 0.83) observed across 13 frontier base models in prior work, with GPT-4o reaching S = 1.68, more than twice the band's upper end—and a 65% average drop in moral robustness (R), equivalent to a 304% surge in 1/R. A matched secure code control largely preserves both metrics, confirming the effects are specific to misalignment-inducing training rather than generic fine-tuning costs. The diagnostic instrument is a persona moral metrics framework computing S and R from responses to the 30-item Moral Foundations Questionnaire under role-play across 100 diverse personas, yielding 30,000 data points per model; a complementary signature is that insecure variants' unconditioned MFQ profiles saturate near the scale ceiling across all five moral foundations, a pattern not reproduced when base models role-play explicit toxic personas. Costa and Vicente argue these results point beyond persona reweighting—the received account under which emergent misalignment merely upweights dark archetypes—to a deeper degradation of the representations used to differentiate and maintain characters, and that S and R provide a sensitive diagnostic capable of detecting residual misalignment even when standard open-ended evaluations indicate improvement.

More papers — OpenAlex / S2

Recent mentions (1)

• papers-typed

  costa-2026-persona-collapse.md