Choose a past signal that fits the task and its resource limit

A document can be compressed before anyone asks about it; a robot can learn a new task while still needing an old one. Reusing the past helps only if the retained signal supports the job ahead.

Across four studies, the practical choice is where to place that signal and what it costs to keep or recreate. Their results favor task-relevant signals in their tested settings, while the studies measure different outcomes and resources.

Imagine compressing a long document before readers submit their questions. A language model can retain only some of its internal entries for the text, so the choice made now affects what it can answer later.

Rereading every line sounds thorough, but the compression study finds that this can spread a tight storage budget across ordinary prose. The immediate problem is choosing which parts of the document deserve that limited space when the future questions are still unknown.

A robot faces a related choice when it learns to open a drawer and later improves its ability to turn on an LED. Updating shared behavior for the new task can disturb the earlier one.

The robot study addresses this by revisiting representations of prior tasks during training. Those representations have a cost: earlier experience could be stored, the robot could repeat a task, or a learned model of its environment could recreate task states. Each option supplies a different reminder of what the robot should retain.

Language-model training adds another version of the problem. A model learning from its own answers can revisit earlier predictions, while a smaller student learning from a teacher can reuse earlier student responses.

The question across these settings is therefore precise: which past signal should be kept or replayed, and will it still be useful for the next decision? Call that signal an anchor.

Its value depends on the information it carries, the computation needed to use it, and the information that must remain stored.