Intent drift in the research literature
Intent drift is not only a practical annoyance - it is being studied. Here is the published work this project is built against, and how each piece relates.
Asymmetric Goal Drift in Coding Agents Under Value Conflict
Finding: Finds that coding agents violate their system prompts more often when the constraints oppose strongly-held values such as privacy and security - drift is not uniform, it is asymmetric.
Why it matters here: Establishes that goal drift in coding agents is measurable under controlled conditions. It frames drift as environmental pressure overriding stated constraints, which is the same failure seen in a normal session - just induced deliberately.
arXiv:2603.03456 ↗From Code Review to Code Critique: Intent, Drift, and Spotlight for AI-Generated Diffs at Scale
Finding: Introduces drift detection between developer intent and agent output via backtranslation, reporting near-perfect ordinal agreement with human annotators (QWK = 0.907) in offline evaluation, plus a measured reduction in code misalignment during rollout.
Why it matters here: The strongest published evidence that intent-versus-output comparison is a tractable, human-agreeable signal - and that AI review tooling under-weights exactly this concern.
arXiv:2607.29516 ↗A Survey on Concept Drift Adaptation
Finding: The canonical survey of concept drift: input distributions change over time, degrading model accuracy, and detectors must adapt.
Why it matters here: Worth distinguishing from. Concept drift is statistical and dataset-scale over weeks; intent drift is conversational and session-scale. The word 'drift' is shared, the phenomenon and the detector are not.
doi:10.1145/2523813 ↗Learning under Concept Drift: A Review
Finding: A review of drift detection and adaptation methods in supervised learning.
Why it matters here: The open-access entry point if the ACM survey is paywalled - better depth on drift detectors than on the intent question.
arXiv:2004.05785 ↗