Measuring and Restoring Constraint Influence in Black-Box LLM Dialogues

A recent paper on arXiv (2608.12599) presents a novel system designed to assess and restore the influence of constraints in dialogues generated by black-box large language models (LLMs). The authors highlight a phenomenon termed 'behavioral relapse' or 'revocation inertia,' where models persist in applying previously withdrawn requirements during multi-turn interactions, even when explicit statements indicate their removal. This system aims to fill three gaps: evaluating influence per clause, forecasting it prior to execution, and rectifying it within specified budgets. It functions solely through the model API, utilizing a contract ledger that associates each constraint with an executable checker, logs revocations as tombstones, and consolidates the overall constraint status into one specification. A sequential ablation probe assesses adherence per clause and the cumulative behavioral impact, while a repair ladder works under matched token and attempt budgets. The system's effectiveness was tested on a dataset of HumanEval tasks with verified checkers, revealing that relapse at an 8B operating point increases from a specific scale (the abstract is truncated). The research team (names not disclosed) announced the paper on arXiv.