safety
A Safety Reward Reaches One While the Model Does Not Move
The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.
Summary
The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.
For its 1.72-billion-parameter Brazilian-Portuguese model, the team says dedicated safety SFT—not DPO—moved the safety-helpfulness frontier. It reports a 94.1 percent lower safety bound above four compared baselines and releases the supporting code, data and logs.
Why it matters
The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.
Limits and context
- For its 1.72-billion-parameter Brazilian-Portuguese model, the team says dedicated safety SFT—not DPO—moved the safety-helpfulness frontier.
Key claims
The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.
Qualification: For its 1.72-billion-parameter Brazilian-Portuguese model, the team says dedicated safety SFT—not DPO—moved the safety-helpfulness frontier.
Evidence: source-2026-10-15-019
Sources
- arXiv preprint 2610.20570, original v1arXiv / submitting research authors · primary research
Corrections
: Publication note: This October 15 edition was prepared before its scheduled 9:00 AM America/Detroit reveal, but publication was delayed by a production build failure. The original reporting and image-generation timestamps are preserved.