RL group-sampling trick: small K plus persistence on hard prompts

A paper discussion thread explains group-sampling RL: small K (e.g., 4) filters easy prompts cheaply, while all-wrong groups keep sampling until a rare success, letting hard prompts' advantage approach 1. The author also muses on visual equivalents like exotic concept combinations.

2026-10-28 ~ 2026-10-28 · 3 related posts

• Why K=4 beats K=32 in group-sampled RL training: filtering easy prompts faster — gowthami_s · 2026-10-28
• RL trick: keep sampling all-wrong groups so rare positives get advantage ~1, not 0.75 — gowthami_s · 2026-10-28
• What 'hard prompts' look like in vision: rare concept combos like armadillos in berets — gowthami_s · 2026-10-28