Chained Recursive Language Models for Multi-Iteration Reasoning
The paper proposes rerunning the same model through fresh reasoning “roots” that inherit artifacts, not the full chat history.
Each pass sees the original problem and context plus a compact summary, a blackboard, and durable task-specific outputs from earlier passes. The aim is to break long-context reasoning into staged work where later passes can inspect, correct, and extend intermediate artifacts. The authors frame it for extraction, counting, ordering, and multi-hop tasks where early errors can carry through a single inference. They describe the architecture and evaluation protocol, and study when this artifact-continuation setup improves accuracy over direct answering. ArXiv · AI/CL/LG's note
Each pass sees the original problem and context plus a compact summary, a blackboard, and durable task-specific outputs from earlier passes. The aim is to break long-context reasoning into staged work where later passes can inspect, correct, and extend intermediate artifacts. The authors frame it for extraction, counting, ordering, and multi-hop tasks where early errors can carry through a single inference. They describe the architecture and evaluation protocol, and study when this artifact-continuation setup improves accuracy over direct answering. ArXiv · AI/CL/LG's note
score 4