Prime Intellect merges IcePop score centering with truncated sampling replay in prime-rl
samsja19 · x · 2026-10-25
PrimeIntellect-ai's prime-rl framework merged PR #4176, adding IcePop score centering support for truncated sampling (bounded top-k, optionally combined with top-p) in RL training. The approach draws from a recorded support (full-distribution top-p followed by a hard top-k cap), replays the entire sampler distribution on the trainer, and centers exactly over it with no tail approximation on replayed tokens. Untruncated sampling keeps the existing head plus proportional-tail approximation. It follows up on the previously merged PR #4012.
More from Research
- Safety researcher: verification should go all the way down to Maxwell's equations, not stop at assembly — JacquesThibs · 2026-10-25
- Local 35B authored 800 math problems under cloud-agent supervision — authoring erred 5x more than cold-solving — Individual_Page9676 · 2026-10-25
- SA-LapNet fixes spin exactly in neural-network quantum Monte Carlo, breaking the purity-energy trade-off — bravo_abad · 2026-10-25
- XYBench: Strongest LLMs Answer Literally and Detect Under 63% of User Misconceptions — MaartenSap · 2026-10-25
- Multi-domain training mix boosts supervised hallucination detectors 7.25% on out-of-distribution QA — MaartenSap · 2026-10-25
- EVALUESTEER: best reward models fail to align with user values over 25% of the time — MaartenSap · 2026-10-25