Tell a chatbot to drop a rule mid-conversation, and there's a decent chance it keeps following it anyway.
Researchers built a system that catches large language models still enforcing constraints a user has explicitly revoked, a bug they call 'behavioral relapse.' The tool logs every constraint and its status in a running ledger, marks withdrawn rules as dead ('tombstoned'), and compiles what's actually still in force into one clean instruction set before the model responds. A second component probes each turn to predict relapse before the reply goes out, and a third tries repair strategies under a fixed compute budget. Tested on coding tasks built from HumanEval, an 8-billion-parameter model relapsed more often as the pile of constraints grew, while larger models mostly avoided the problem. Precompiling the live constraint set before generation cut relapse significantly compared with a baseline that just re-checks the output after the fact; layering more elaborate repair steps on top added no measurable benefit, and simply appending a one-line note that a rule is dead recovered roughly a third of the fix's effect.
This matters because relapse is a quiet failure mode. A model that adds a wrong fact is easy to spot; a model that keeps a feature you asked it to remove, or reinstates a rule you told it to drop, looks like it's working right up until you check the output closely. That's a live risk for coding assistants and agents running long sessions, where instructions pile up and get revised constantly. It also reframes instruction-following as something that has to hold across edits, not just at first read.
The fix works, but it takes real engineering: a constraint ledger, a prediction probe, and extra API calls layered onto every turn. Most chatbot products aren't running anything like that today, so until they do, 'never mind' is more of a suggestion than a command.