An agent repeatedly reading an expert trajectory pays to recover the same lesson on every step. A prose summary reduces that cost but leaves the model to translate advice back into a specific action.
CB-Agent, posted October 10, fits a concept vector from expert environment-action decisions and combines it with frozen model features to score candidate actions. The demonstrations are absent from the action-scoring context at runtime. The authors evaluate the method in TextWorld, ALFWorld and ScienceWorld. Read the research paper.
The reusable object is a decision rule
This is a different form of memory from saving a transcript. It preserves how to rank actions, sacrificing some of the original episode's detail. That trade can be attractive when the same decision structure recurs in many situations.
The limiting question is whether the candidate actions still mean the same thing. New tools, changed preconditions or a different environment can invalidate the learned ranking even when the language looks familiar.
A practical test would hold the instruction constant while changing an action's consequences. Does the concept adapt or keep applying yesterday's preference? The published simulation results should not be extended to unrestricted tool use without that check.
CB-Agent makes demonstration reuse less dependent on prompt space. The next challenge is deciding when a compact rule has outlived the evidence that made it sensible.
References
Questions, news tips or corrections? Contact [email protected].
