The Sequence Knowledge - Issue #964: Learning RSI: How Models Generate Their Own Learning Curricula
Absolute Zero and the search for useful difficulty
TL;DR
- Fixed datasets currently dictate model training, but models can now generate their own curricula for recursive self-improvement.
- Absolute Zero grounds coding tasks in execution, while R-Zero uses a Challenger-Solver model with self-grading.
- R-Few supplements R-Zero with a small set of human examples to prevent curriculum drift.
- Difficulty is an imperfect proxy for progress, as verifiers only confirm test coverage, and agreement can repeat mistakes.
- Adaptive curricula can respond to a learner's evolving capabilities, making experience selection part of the learning process.
- Evidence for self-generated curricula requires demonstrating progress across multiple rounds and fixed budgets on uninfluenceable evaluation tasks.