Emergent curriculum without labels: prompt weights diverge 1.6x vs 0.75x over training
shangbinfeng · x · 2026-10-20
Detail from the Stackelberg Alignment thread: with no difficulty labels, the bandit leader learns to steer model duels toward the capability frontier, yielding an emergent curriculum that advances as models improve. On reasoning datasets, the top-5 prompts rise to 1.6× uniform weight by the final iteration while the bottom-5 fall to 0.75×, with weights diverging steadily across training iterations.
More from Research
- PreferenceEKF turns active reward learning into sequential Bayesian inference (TMLR 2026) — ebiyik_ · 2026-10-20
- mjfit: open-source convex decomposition for tight-clearance parts, 10x faster than SDF — yacineMTB · 2026-10-20
- iSEE world model learns object permanence from video alone, no labels — CSProfKGD · 2026-10-20
- TasteBench tests model research taste on ever-fresh out-of-sample ML tasks — marktenenholtz · 2026-10-20
- "Neurons aren't already clean — interpretability just covers the dunghill with snow" — 1a3orn · 2026-10-20
- Single neuron steering makes Qwen3-8B insist it's an owl — Sauers_ · 2026-10-20