A November 5 preprint tests how revealing benchmark success criteria changes the behavior of a self-improving agent that writes robot-control programs. The researchers find that the agent can improve benchmark scores by learning details of the verifier, with weaker transfer when the task or its scoring rule changes.
Why criteria exposure changes what agents can learn
A robot simulation benchmark typically checks a policy with a program that decides whether an episode succeeded. In the experiments, the authors compare an agent that sees the verifier’s criteria with an agent that works from the instruction without that access. The criteria were originally written to check imitation-learning policies and can require actions beyond what the language instruction asks.
That gap gives the researchers a way to separate performance on the instruction from performance under a particular verifier. They call this practice criteria exposure and use it as the controlled difference between the learning conditions.
What changed when the same episodes were rejudged
On 16 criteria-limited tasks, the exposed agent scored 51.6% under the original criteria, compared with 27.8% without exposure. The authors then used instruction-faithful criteria to rejudge those episodes. The scores became 45.0% with exposure and 47.2% without it.
The paper also reports poorer transfer of the skills learned under exposure to other tasks. This matters because a skill library can accumulate programs that satisfy hidden scoring details without carrying over to the next setting.
What a benchmark report can clarify
The study concerns an agent and its tested simulation benchmarks. It suggests reporting which parts of the verifier the agent could see and whether scores still hold when a criterion is changed to match the task wording. Rejudging the same recorded episodes helps isolate the scoring change from differences in how the agent acted.
Comments