Test-time routing on 10-cell Steerable-Hard (evaluation set)
46.8Rescue RateProblemOracle
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ProblemOracleRouting info=per problem-unit arg max over operators (not deployable), Needs labels=yes, Compute (x retry)=66.52026.06 | 46.8 | 0.304 | |
| DiagnosticOracleRouting info=runs every operator on the test cell, picks modal-best, Needs labels=yes, Compute (x retry)=66.52026.06 | 32.1 | 0.157 | |
| Outcome-trainedRouting info=RF classifier on the cell’s own per-pid operator outcomes (assay cost = DiagnosticOracle’s), Needs labels=yes, Compute (x retry)=66.52026.06 | 30.2 | 0.139 | |
| Feature-onlyRouting info=winner-take-all over 3 trajectory features; no training, Needs labels=no, Compute (x retry)=1.42026.06 | 28.5 | 0.122 | |
| RetryRouting info=no information, no test-cell compute beyond resampling, Needs labels=no, Compute (x retry)=1.02026.06 | 16.3 | 0 |