What are They Thinking? Delineation, Probing and Tracking of Concepts in LLMs

In three linesMethod to create linear probes detecting concepts in LLM embeddings. Authors define a process: concept delineation via contrastive datasets, layer-wise probe training, tracking across large contexts. Tested on 4 concepts and 3 different LLMs. Goal: scalable monitoring of new models.

Summary generated by Claude — human-verified