2B scoring model flags out-of-domain misalignment, suggesting specialist judges have potential for audits

·LessWrong··

TL;DRSome evidence that narrow ‘specialist’ models could be useful as part of deployed model misalignment audits, complementing larger frontier auditing agents and offering potential cost, discrimination and transparency benefits.A Gemma 2B specialist judge trained on Betley et al 2025b (‘Betley’) code examples is able to distinguish between responses provided by insecure-fine-tuned ‘misaligned’ and ‘secure-fine-tuned’ models responding to out-of-domain general safety prompts (‘ICEBERG’ - detail...

Read full article →

Related Articles

An OpenAI model has disproved a central conjecture in discrete geometry
tedsanders · Hacker News · 20h ago
GitHub confirms breach of 3,800 repos via malicious VSCode extension
Timofeibu · Hacker News · 1d ago
Show HN: Rmux – A programmable terminal multiplexer with a Playwright-style SDK
shideneyu · Hacker News · 6h ago
Incident Report: May 19, 2026 – GCP Account Suspension
0xedb · Hacker News · 1d ago
Not alive, but not dead: disembodied human brains used for drug testing
Timofeibu · Hacker News · 19h ago