Why K=4 beats K=32 in group-sampled RL training: filtering easy prompts faster
gowthami_s · x · 2026-10-28
A thread unpacking a recent paper: with K=32 in group-sampled RL, a single wrong answer among 32 keeps an easy prompt in the batch, so batches fill with basically-solved prompts. K=4 filters easy prompts out fast and frees compute for hard ones—essentially active learning in disguise.
Related event: RL group-sampling trick: small K plus persistence on hard prompts(3 posts)→
More from Research
- Eric Topol highlights immunome paper combining multi-assay data with AI to model immune aging — EricTopol · 2026-10-28
- Google Research's 'sleep framework' lets LLMs keep learning, retaining up to 10M tokens — behrouz_ali · 2026-10-28
- CATCH: ~200 examples of adversarial training make LLMs robustly honest across benchmarks — irinarish · 2026-10-28
- Andrew Critch's dubious credences: P≠NP at 80%, Riemann hypothesis at just 60% — AndrewCritchPhD · 2026-10-28
- AI-driven Indian biotech PopVax, backed by $20M from Vitalik's Balvi fund, enters human trials with fridge-stable broad-spectrum Covid vaccine — chaitjo · 2026-10-28
- Cognition launches FrontierCode 2.0, a coding benchmark testing judgment beyond the spec — DevinAI · 2026-10-28