文章背景与核心概要
在多轮对话中，用户可以像施加约束一样轻松地撤销约束。然而，大型语言模型（LLM）往往无法正确处理这些撤销操作——这种现象被称为“行为复发”（behavioral relapse）或“撤销惯性”，即模型会继续强制执行已被撤销的要求。
本文介绍了一种仅通过模型API即可解决三大技术空白的系统：1. 契约账本（Contract Ledger）：将每个约束与可执行检查器配对，将撤销记录为“墓碑（tombstones）”，并将净约束状态预编译为单一规范；2. 序列消融探针（Sequential Ablation Probe）：测量逐条条款的依从性和增量行为影响；3. 修复阶梯（Repair Ladder）：在严格的Token和尝试次数预算限制下运行。在HumanEval任务和已验证检查器上的评估表明，与标准基线相比，预编译显著减少了复发，将撤销失败转化为对话状态中可测量、可预测且可修复的组成部分。
Dead text or binding clause? Measuring and restoring constraint influence in black-box LLM dialogues
Authors: Haoyuan Zhu
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv Identifier: arXiv:2608.12599 [cs.AI]
Submitted: August 12, 2026
Summary
多轮对话使用户能够像施加约束一样轻松地撤销约束，但撤销并不能可靠地生效：模型会继续执行已被撤销的要求（有时甚至在声称已移除它们的注释下方），我们将这种故障称为行为复发或撤销惯性。目前没有任何现有的工具能够按条款测量这种影响力、在输出前对其进行预测，或在匹配的预算下对其进行修复。
Multi-turn dialogues allow users to easily revoke constraints just as they introduce them. However, large language models (LLMs) often fail to properly process these revocations—a phenomenon known as behavioral relapse or revocation inertia, where models continue enforcing withdrawn requirements.
本文提出的系统仅通过模型API弥补了这三个空白：契约账本将每个约束与可执行检查器配对，将撤销记录为墓碑，并将净约束状态提前编译为单一规范；序列消融探针测量逐条条款的依从性和增量行为影响；修复阶梯在Token和尝试次数匹配的预算下运行。
This paper introduces a system that resolves three major gaps using only the model API: 1. Contract Ledger: Pairs every constraint with an executable checker, records revocations as "tombstones," and pre-compiles the net constraint state into a single specification. 2. Sequential Ablation Probe: Measures per-clause adherence and incremental behavioral effects. 3. Repair Ladder: Operates under strict token- and attempt-matched budgets.
在8B模型的运行点上，复发率随约束负载的增加而上升，而更强的模型则处于基准底层。在匹配检查器、模型和预算的条件下，提前编译显著减少了相对于无账本验证器重试基线的复发。该探针可以在输出前预测复发（AUROC），并且一句墓碑注释可以恢复大约三分之一的编译效果。因此，撤销失败成为了对话状态中一个可测量、可预测且可修复的属性，而不再是一个隐形问题。
Evaluated on HumanEval tasks and verified checkers, the study shows that pre-compilation significantly reduces relapse compared to standard baselines, turning revocation failure into a measurable, predictable, and repairable component of dialogue state.
Abstract
Multi-turn dialogues let users revoke constraints as easily as impose them, but revocation does not reliably take effect: models keep enacting withdrawn requirements (occasionally beneath comments asserting their removal), a failure we call behavioral relapse, or revocation inertia. No existing instrument measures this influence per clause, predicts it before delivery, or repairs it under matched budgets.
The proposed system closes these three gaps through the model API alone: a contract ledger pairs every constraint with an executable checker, records revocations as tombstones, and compiles the net constraint state ahead of time into a single specification; a sequential ablation probe measures per-clause adherence and incremental behavioral effect; a repair ladder operates under token- and attempt-matched budgets.
At an 8B operating point, relapse climbs as constraint load grows, while stronger models sit at floor. Under matched checkers, model, and budget, ahead-of-time compilation significantly reduces relapse against a no-ledger verifier-retry baseline. The probe predicts relapse before delivery (AUROC), and a one-sentence tombstone note recovers about a third of the compilation effect. Consequently, revocation failure becomes a measurable, predictable, and repairable property of dialogue state rather than an invisible one.