CHAIN OF RAG

Summary from the TLDR AI newsletter

Reasoning models can now be trained to iteratively do retrieval. This is similar in spirit to the work shown in the Operator system. Interestingly, it leads to a substantial improvement, although it's not clear what the true FLOP controlled gain looks like.