Composable Curricula and difficulty pool for optimized training
samsja19 · x · 2026-08-26
Introduces a "Composable Curricula" training scheme where environments manage their own data. The core mechanism involves a "difficulty pool" that tracks task complexity. As the model improves, it resamples easy data less frequently and carefully injects harder data to optimize training efficiency.
More from Research
- Tech note: Small EPS affects normalization norm, not performance — YouJiacheng · 2026-08-26
- Forecasting benchmark simulates luck ceiling; LLM evals need the same — Crescitaly · 2026-08-26
- Researcher skeptical: many projects rebrand harness optimization as 'RSI' — andrey_kurenkov · 2026-08-26
- Game2World Engine removes gameplay UI for better world model training — GameWorldData · 2026-08-26
- UAI2026 announces keynote speakers in causal learning and responsible AI — LesterMackey · 2026-08-26
- A 30-question deep dive into embeddings, vector search and retrieval — techNmak · 2026-08-26