Standard SFT for agents backpropagates only through action tokens; environment observations are treated as context, not prediction targets. ActObs changes this by also applying the language modeling loss to observation tokens already present in each trajectory. The two objectives use the same parameters, so no extra data, parameters, sequence tokens, or forward passes are needed.
The methods score similarly after SFT, but diverge once GRPO is applied. On Terminal-Bench 2.0, the Qwen3-4B ActObs policy beats the action-only policy at every sampling budget, with a 29% relative advantage at pass@1. At 8B, ActObs gives up some pass@1 reliability but gains +3.4 pp at pass@16 and solves more distinct tasks. On aider-polyglot, an unseen code editing benchmark, the 4B ActObs policy is +4.2 pp at pass@1.
The result matters because it suggests that action-only SFT specializes too far, degrading environment prediction, while joint supervision preserves consequence modeling and leaves the policy closer to its initialization, which helps exploration during RL.