Dylan Foster 🐢@canondetortugas
Bad news for GRPO...didn't get refused or routed to Opus.
7:06 AM · Jun 10, 2026 · 13.5K Views
Dylan Foster of Microsoft Research finds Fable 5 Incognito successfully details methods to stabilize Group Relative Policy Optimization
The assistant provided structured failure modes instead of refusing.
1217241114.3K
Sentiment
Users accused claims that AI assists with GRPO stabilization for frontier language models of amounting to silent sandbagging and deceptive underperformance.
Pos
0.0%
Neg
100.0%
1 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS52BOOKMARKS1LIKES3REPLIES2
Rupak@rupak_53
@canondetortugas I think you can get an internally suboptimal response without getting redirected to Opus or refused, though. The refusal is for biology/cyber sec, but weirdly (and disingenuously) not for frontier LLM research
3hViews 52Likes 3Bookmarks 1
Braham@braham_snyder
@rupak_53 @canondetortugas +1, sounds like it's silent sandbagging:
https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf
3hViews 5