Hyperstition introduces judge-time scaling: a new scaling law for RL
eliebakouch · x · 2026-11-21
Hyperstition released a new RL scaling law, "judge-time scaling," plus the Hyper-Intervention benchmark and open-source repo. Editor judges read student rollouts and intervene token-by-token: using <1% of rollout tokens, the best judge took Qwen2.5-7B base 64% of the way to its Instruct variant, generating on-policy data that beats the Pareto frontier of high-scoring rollouts. Their takeaway: models now outperform SOTA RL algorithms at credit assignment, and giving judges more compute mollifies the data bottleneck in RL. A chart compares several judge models' test gains over 180-minute sessions.
More from Models
- Waddle-1 robot-use agent scores 81 on unseen bimanual tasks, beating GPT-6.2 Astra at 66 — ycombinator · 2026-11-21
- DeepMind's empty math-horizon repo sparks millennium-problem speculation — jasondeanlee · 2026-11-21
- alphaXiv Launches OpenResearch Intelligence Index, First Open Benchmark for Autoresearch — togethercompute · 2026-11-21
- 26.4B Inkling 2 Mini matches 975B model in 128 days, ~17x compute efficiency gain — teortaxesTex · 2026-11-21
- Coding images of France with 6.2 Astra: a stress test for visual reasoning — prd_008 · 2026-11-21
- Downdetector reports OpenAI outage starting 3:13 PM EST as user complaints pile up — VoidStateKate · 2026-11-21