Step through one revision round. Toggle between training and test time.
The executable program around a model: what it sees, which tools it uses, how outputs are checked. Same weights, different behavior.
Hover a row. Code edits stand in for gradient steps (MAML, RL²).
Qwen3.5-4B is both proposer and solver. Hover the 100 instances of one family.
Qwen3-4B proposer, Qwen3-8B solver, RL straight from the base model.
Hover the donut. Interpreter loops dominate the 432 teacher revisions.
Reward = task score + 0.05·[parses] + 0.10·[applies] + 0.15·(fraction of trials completing)
Eight sampled edits per context are scored and normalized within the group. Reward values here are made up for illustration.
Numbers as stated in the episode. Pick a view.
Counts not to scale. Sequential rounds also see feedback from the retained harness; independent proposals do not.
A qualitative reading of the episode, not paper-reported scores. Hover a cell.
A good edit the solver cannot use is worth nothing. Toggle the lens.