implementation detail · filed under evaluation
Standardized agent evaluation protocol
Specifying sampling counts, sampling parameters, context limits, and maximum agent steps for benchmark runs.
- 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.
All scaffolds use N=8 samples per task on DeepSWE v1.1 and N=3 on Terminal-Bench 2.1, with Linux containers, temperature=1.0, top_p=0.95, a 1M-token context limit, and max_steps=500 per agent.
usedevaluation onlyin DeepSeek-V4.1-FlashDeepSeek
Filed alongside
Other methods under evaluation.
Prompt-based output standardizationEvaluation without safety filtersRefusal-suppressed variants for capability estimationavg@kAlmost@1Behavior testing with stubs and edge casesDeadline-bounded rollout evaluationDiverse benchmark validation for quantizationDocumenting evaluation configurationEnd-to-end exploit development evaluationEvaluation equivalence thresholdFixed MTP draft-token acceptance lengthFour-run mean pass@1Full trajectory releaseJoint validation protocolMean@5Model–harness co-designOfficial-release-source baseline restrictionPrompt engineering to elicit answersProtocolQA robustness validationRefusal behavior quantificationRollout-based auditingSample test and length-constraint filteringSandboxed GPU kernel optimization evaluation