Tag
AdaPlanBench is a dynamic benchmark for evaluating LLM agents' ability to adaptively plan under progressively revealed world and user constraints through multi-turn interactions, showing current models struggle especially with user constraints.
An informal research note describing a behavior in transformers where the model's inherent 'clarity-seeking' vectors can bypass constraints when discussing higher-order topics, potentially relevant to alignment and safety research.
The author argues against building fully autonomous email agents based on real-world failures, advocating for constrained 'propose-and-approve' workflows where AI prepares context and drafts but humans retain final approval on sends.
A blog post explaining how to use refinement mappings to preserve external properties during database schema changes, using examples of migrating a boolean column to a nullable timestamp and then to event sourcing.