specific method · filed under evaluation
Reward hacking mitigation in evaluations
Limits internet access and removes Git histories from coding-agent evaluation environments to mitigate reward hacking.
Also called reward hacking mitigation.
- 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.
To mitigate reward hacking in coding agent evaluations, we restrict internet access and strip Git histories from the environment.
usedevaluation onlyin DeepSeek-V4.1-FlashDeepSeek
Filed alongside
Other methods under evaluation :: human & real-world evaluation.
Blind side-by-side expert evaluationMulti-turn open-ended external red-teamingInternal engineering task evaluationA/B testing on real tasksAutomated red-teaming loopContinuous monitoringCyber range exercisesDangerous-capability uplift assessmentExternal expert review of safety evaluation methodologyHeld-out evaluationHuman evaluationPairwise comparison for Chinese writingPre-release safety evaluationReal-world codebase testingReal-world feedback refinementSame-prompt comparative DevOps testingSystematic robustness-boundary characterizationWhite-collar enterprise task evaluation