ObsSpec: Accelerating Agentic Reinforcement Learning with Observation Speculation
OBSSPEC introduces an observation speculation pipeline for agentic RL that uses a small world model to predict tool outputs while the environment runs. This allows the policy to start generating its next step early, reducing agent latency. The system ensures lossless rollout collection by only keeping predictions when they exactly match actual observations and uses co-training to keep predictions accurate as the policy learns.