Chain fresh model contexts through a plain-text blackboard instead of one long trajectory

arXiv 2608.05124·low signalXBlueskyLinkedInCopy link

Chained Recursive Language Models invoke the same model repeatedly as distinct reasoning phases, each receiving the original problem and context plus a compact plain-text summary, a plain-text blackboard, and durable task-specific artifacts written by predecessor roots. The point is to stop a single trajectory from simultaneously exploring context, storing intermediate state, verifying evidence, and answering — and it lets a later fresh inference inspect, correct, and extend earlier intermediate results. The paper maps where this beats both direct responses and recursive tool-calling on extraction, counting, ordering, and multi-hop tasks; it reports no headline numbers in the abstract, so treat the architecture as the takeaway.