Research
Differend Labs groups its studies into series. A series is a set of studies that all ask the same question. Choose a series to see its entries, newest first.
hark (verb): to listen closely. Every study in the HARK series asks the same question in a different place: when people tell an AI system what they need, does the system act on it?
-
Does the judge agree with people?
Some AI systems learn from a score instead of from a person. In this study, a program learns to find unusual points in data recorded over time, such as traffic to a website or readings from computer servers. Part of the program's score comes from a large language model (LLM), which rates how unusual each stretch of data looks.
The study checks two things. First, do the LLM's ratings match the points that people labeled as unusual? Second, does the program learn to get a high rating from the LLM instead of finding the points people labeled?
Builds on: Golchin, Rekabdar, and Justo, LLM-Enhanced Reinforcement Learning for Time Series Anomaly Detection (2026).
To be published: the study plan, written before any results; the code; every result, including failures; and a plain-language write-up.
New entries also appear in the RSS feed, which news reader apps can follow.
hello@differend.ai