Reproduce “From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuan Yuan, Tina Sriskandarajah, Anna-Luisa Brakman, Alec Helyar, Alex Beutel, Andrea Vallone, Saachi Jain
OpenAI researchers developed an output-centric "safe-completion" training paradigm for Large Language Models, which enables models to provide helpful responses within safety constraints rather than outright refusing. This approach, integrated into GPT-5, demonstrated improved safety and helpfulness while reducing the severity of residual harmful outputs, particularly on dual-use prompts.
Key insights
- An output-centric safety paradigm, focusing on the generated content rather than user intent, provides a more robust and scalable approach to LLM safety.
- A composite reward structure that balances both direct and indirect helpfulness with safety scores enables models to navigate dual-use scenarios gracefully.
- Safe-completion training leads to "softer" safety failures, reducing the severity and actionability of any residual harmful outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.