implementation detail · filed under evaluation
Multiple-rollout evaluation
An evaluation setup that runs up to three rollouts per task on ProgramBench.
Also called Multiple-rollout evaluation per task, up to three rollouts per task.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
On ProgramBench, we run up to three rollouts per task
usedevaluation onlyin DeepSeek-V4.1-FlashDeepSeek
Filed alongside
Other methods under evaluation :: benchmark.
High-confidence ProgramBench subsetArtificial Analysis Intelligence IndexBrowseComp with SearchCAD visual reproductionCapture-the-Flag (CTF) evaluationComputer use closed-loop evaluationDisabling tool searchDomain whitelistExecution-grounded evaluationF2P/P2F acceptance criterionFail-to-pass and pass-to-pass evaluation pointsGPU kernel optimization task suiteKernelBenchLiving in-house benchmark suiteLongBench v2PaperBenchPinchBenchProfBench with SearchScaleAI Multi Challenge Multi Turn Instruction FollowingSQuADSWE-bench VerifiedThree-configuration cyber range testingVals.aiVerified CUDA kernels