Google releases EnvHarness to fit test environment to agents
In tests, the performance of 4 models increases by 2.9 to 3.7 percentage points.
Google researchers released EnvHarness as open source code. It is a framework around a static environment. It changes the environment to fit the weak points of an agent. It does not change the agent, the environment or the test system. The researchers think that it gives good results in agent learning. Tests on WebArena, SWE-bench Verified and ALFWorld show an increase in performance for each of 4 models.
Sources
Posted
