arXiv Open Access 2025

From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training

Yuan Yuan Tina Sriskandarajah Anna-Luisa Brakman Alec Helyar Alex Beutel +2 lainnya
Lihat Sumber

Abstrak

Large Language Models used in ChatGPT have traditionally been trained to learn a refusal boundary: depending on the user's intent, the model is taught to either fully comply or outright refuse. While this is a strong mitigation for explicitly malicious prompts, focusing safety training on refusals can lead to brittleness for prompts with obscured user intent. Binary refusal boundaries are especially ill-suited for dual-use cases (such as biology or cybersecurity), where a user request can be answered safely at a high level, but in some cases can lead to malicious uplift if sufficiently detailed or actionable. As an alternative, we propose safe-completions: a safety-training approach that centers on the safety of the assistant's output, rather than a binary classification of the user's intent. Safe-completions seek to maximize helpfulness within the safety policy's constraints. We incorporated this approach into GPT-5 and find that across both production comparisons and internally controlled experiments, safe-completion training improves safety (especially on dual-use prompts), reduces the severity of residual safety failures, and substantially increases model helpfulness.

Topik & Kata Kunci

Penulis (7)

Y

Yuan Yuan

T

Tina Sriskandarajah

A

Anna-Luisa Brakman

A

Alec Helyar

A

Alex Beutel

A

Andrea Vallone

S

Saachi Jain

Format Sitasi

Yuan, Y., Sriskandarajah, T., Brakman, A., Helyar, A., Beutel, A., Vallone, A. et al. (2025). From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training. https://arxiv.org/abs/2508.09224

Akses Cepat

Lihat di Sumber
Informasi Jurnal
Tahun Terbit
2025
Bahasa
en
Sumber Database
arXiv
Akses
Open Access ✓