Toward understanding and preventing misalignment generalization
1y
- Published
- Collected
We study how training on incorrect responses can cause broader misalignment in language models and identify an internal feature driving this behavior—one that can be reversed with minimal fine-tuning.
overfeed.news indexes and links. We publish a short excerpt — the full article stays at OpenAI News.
More from OpenAI News
Log in to follow this source
How Itaú is building generative AI for its entire customer base
With ia.i, a conversational interface, and the IARA governance platform, Itaú is applying its AI First strategy to customers and employees — in one of Brazil’s most regulated sectors.
Benefit Street Partners answers questions faster with Codex
Anant Kumar builds tools to compare loan valuations and analyze credit agreements, using Codex for code review and faster follow-ups.
Stack Overflow for Agents triples signups with ChatGPT plugin
Two engineers shipped Stack Overflow for Agents in a few weeks with Codex, and its ChatGPT plugin helped drive up to 10× growth in agent creation.
Bringing GPT-Rosalind to academic research labs
Introducing the Rosalind Academic Plan, offering eligible academic labs sponsored access to GPT-Rosalind and tools to advance life sciences research.