Learning RSI: models generating their own training problems and curricula
Overview
Absolute Zero, R-Zero and R-Few let models generate their own training problems instead of drawing on a fixed dataset, according to The Sequence.
Absolute Zero grounds its proposed coding tasks in code execution. R-Zero pairs a Challenger, which proposes problems, with a Solver that grades its own answers by majority vote.
The Sequence warns that difficulty is an imperfect proxy for progress. Verifiers only cover what their tests check, and agreement among sampled answers can repeat a shared mistake.
Written by AI from the articles below · updated Nov 3, 7:21 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- The SequenceBlogLearning RSI: how models generate their own learning curricula
AIAbsolute Zero, R-Zero, and R-Few let models generate their own training problems instead of relying on a fixed dataset. Absolute Zero grounds proposed coding tasks in code execution, while R-Zero pairs a Challenger with a Solver that grades itself by majority vote. The source warns that difficulty is an imperfect proxy for progress, since verifiers only cover what their tests check and agreement among sampled answers can repeat a mistake.
Heat Trend
Not enough continuous observations to show a trend yet.