Definition
Plain language
Keeping straight where things currently stand after a long series of changes.
As stated in the literature
Maintaining an accurate representation of the evolving problem configuration across a generation, which transformers must recompute from the visible prefix each step since they carry no persistent scratch memory.
Also called: state maintenance
Why it matters: Without reliable state tracking a model will confidently propose moves that are illegal or pointless, and long multi-step tasks unravel a few steps in.
For example, after fifteen described moves, a model has to know which disk is now on which peg in order to pick a legal sixteenth move.
Heard on the show
“If they fail because they can't hold a representation they already form, you need state maintenance — external trackers, periodic re-grounding, scaffolding that writes the current state down.”Episode 237 — The Model Built a Perfect Map of the Puzzle, Then Lost It