RL trick: keep sampling all-wrong groups so rare positives get advantage ~1, not 0.75
gowthami_s · x · 2026-10-28
Continuing the thread: the paper keeps small K, but when a group is all wrong (usually a hard problem), it keeps sampling with probability p until one answer is right. Those failures go into one big group, so the rare positive gets an advantage near 1 instead of capped at 0.75. Stale negatives older than 4 steps only count in the baseline.
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