A new twist on training AI agents makes them noticeably better at learning from experience, and it doesn't require a single extra piece of data.
Researchers built ActObs, a change to how agents are prepared for reinforcement learning. Standard supervised fine-tuning only trains a model to predict the actions it takes, treating everything the environment sends back as background context rather than something worth learning from. ActObs also trains the model to predict those observation tokens. Tested with the GRPO reinforcement-learning method on Terminal-Bench 2.0, agents pretrained with ActObs solved more tasks across the board on Qwen3-4B, and on Qwen3-8B they traded a little first-try reliability for a 3.4 percentage point gain at pass@16 and more distinct tasks solved overall.
The interesting part is why. The researchers trace the difference to standard fine-tuning itself: action and observation gradients quickly become orthogonal, so action-only training actually makes the model worse at predicting what happens in its environment than the untrained base model. ActObs stops that one-sided specialization, keeps more randomness in the policy, and needs smaller updates during reinforcement learning. The effect generalized too: on aider-polyglot, a code-editing benchmark neither model saw during training, the 4B model's pass@1 score jumped 4.2 percentage points.
It's a modest architectural footnote, not a new paradigm. But it's a useful reminder that what you do before reinforcement learning starts can matter as much as the RL algorithm itself.