Supervising environment observations improves agent exploration performance
Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL
Machine LearningArtificial IntelligenceComputation and Language
Summary
When training AI agents to act and learn from their environment, usually only their actions are focused on during learning. The paper shows that also teaching agents to predict what they will observe next helps them understand the consequences of their actions better. This leads to improved exploration and performance after reinforcement learning. The authors found that agents trained this way keep more flexibility and adapt better during later learning stages.
What this means in practice
- •For ai developers: Improve reinforcement learning agents’ exploration by supervising environment observations alongside actions during training.
- •For automated code editors: Enhance performance on unseen code editing tasks by integrating environment observation supervision in policy training.
Authors
Juzheng Zhang, Disha Makhija, Manoj Ghuhan Arivazhagan, Vinayshekhar Bannihatti Kumar, Rashmi Gangadharaiah
Abstract
Agent trajectories record what an agent does and what happens next. Yet standard supervised fine-tuning (SFT) applies loss only to agent-authored action tokens, using environment observations as context but not as prediction targets. We ask whether this convention provides the best initialization for subsequent reinforcement learning. We introduce ActObs, which also supervises the observation tokens already present in each trajectory. Although deployed agents never generate observations, learning to predict them encourages the policy to model action consequences without adding data, parameters, sequence tokens, or forward passes. The methods perform similarly after SFT but diverge after GRPO. On Qwen3-4B, GRPO from ActObs achieves higher pass@k at every evaluated sampling budget than its action-only counterpart on Terminal-Bench 2.0. On Qwen3-8B, it trades some pass@1 reliability for higher pass@k (+3.4 pp at pass@16) and solves more distinct tasks. The advantage extends to cross-domain code editing on aider-polyglot (+4.2 pp at pass@1 at 4B), whose tasks are unseen during SFT and RL. ActObs retains more entropy during RL while requiring less policy movement, leaving the final policy closer to its SFT initialization. Our analysis traces this difference to SFT: action and observation gradients rapidly become orthogonal, while action-only training leaves a large residual observation gradient and degrades environment prediction below the base model. Joint supervision prevents this one-sided specialization, preserving consequence prediction and preparing the policy for downstream exploration.