A new benchmark called Chronocooked asks reinforcement learning agents to do something most RL research skips: keep track of time they're never allowed to see.
Researchers built a suite of cooking-themed environments modeled on Overcooked, the multiplayer kitchen game, where tasks and rewards hinge on timing information the agent never directly observes. The environment is kept deliberately simple, the authors say, to enable controlled experiments and to accommodate biologically plausible models of time perception alongside standard neural networks. The paper reports three baseline agents evaluated on the suite: a non-recurrent network, a recurrent network, and a biologically inspired timing model. Its evaluation metrics are designed specifically to expose gaps in timing ability, though the paper itself focuses on describing the benchmark and its baselines rather than reporting how each one performs.
Reinforcement learning benchmarks tend to test spatial reasoning, multi-agent coordination, or long-horizon planning - rarely the narrower question of whether an agent can estimate elapsed time without a clock. That gap matters for anything meant to work alongside people: a delivery robot judging how long a task should take, or an assistant deciding when a wait has gone on too long. Chronocooked isolates that one variable, which makes it a clean diagnostic where other benchmarks bundle timing in with everything else.
The paper lays out the benchmark and its baseline models in detail; whether any of them actually handle the hidden clock well is a question left for the results section to eventually answer.