2nd Place
Zero Shot Cancer
RL environment for autonomous biologist agents.
Core trick
Hidden biology state is exposed only through noisy simulated outputs, and rewards score plan validity, efficiency, and final state recovery.
Description
RL environment for autonomous biologist agents. Simulate an entire biological worldstate based on scientifically-accurate single-cell standards. Naive agents try to solve a frontier problem, such as identifying key pathways and markers for cancer growth, and have access to 40+ tool calls for fully implemented bioinformatics procedures. Agents probe and modify the world state and try to recover the generated "truth" of the given world using intermediate experimental results to guide future thinking. Problems range in difficulty and complexity, requiring more and more complex thought and workflows. Rewards are based on feasible flow of experiments and the accuracy of the final study conclusion, based on how close the recovered world state is to the hidden "true" world state.
README excerpt
--- title: Bio Experiment Environment Server sdk: docker pinned: false app_port: 8000 base_path: /web tags: - openenv - reinforcement-learning - bioinformatics --- # Bio Experiment Environment This repository implements an OpenEnv-compatible reinforcement learning environment for planning biological experiment pipelines. The agent does not directly see the true biological state. Instead, it proposes one structured experiment or analysis step at a time, receives a noisy simulated output, and is rewarded for valid, informative, efficient, well-calibrated plans. The environment is designed as a partially observable Markov decision process (POMDP) with: - hidden ground-truth biology - hidden technical noise and failure conditions - visible task metadata, resource usage, step history, and intermediate outputs - dense step-wise reward plus terminal reward for conclusion quality ## How it works At a high level, each episode looks like this: 1. `reset()` picks a biological scenario and seeds the simulator. 2. The agent receives an `ExperimentObservation` describing the task and current visible state. 3. The agent submits an `ExperimentAction` such as `collect_sample`, `run_qc`, or `differential_expression`. 4. The rule engine checks whether the action is valid at this point in the pipeline. 5. The transition engine updates hidden state, spends resources, and asks the output generator to simulate the result. 6. The reward computer scores the step for validity, ordering, i