Trajectory-supervised training
Training a model on complete action sequences — full trajectories from start to outcome — rather than single-step next-token prediction, so it learns to plan across a longer horizon.
Learn first
Standard language model training predicts one token at a time from the token before it. Trajectory-supervised training instead labels whole sequences of decisions — a chain of reasoning steps, a multi-step tool-use session, a game played to completion — and optimises the model against those end-to-end paths. The model sees what a good plan looks like all the way through, not just what the next move should be. The hard part is data: useful trajectories are expensive to collect or generate, and a bad trajectory used as a positive example teaches the wrong lesson at every step along it.
In megatrends
Finds citing this concept
Overall = breadth + depth + substance + ½·novelty + freshness (a small boost that halves each week)
Scaling the horizon, not the model
An open-weights agent that fires only 3 billion parameters per step was trained on entire tool-use trajectories — and on a run of scientific and search tasks it keeps pace with models a hundred times its active size.
How this connects
Tap a node to open it