Chained Recursive Language Models for Multi-Iteration Reasoning

The paper introduces Chained Recursive Language Models (Chained RLM), an inference-time architecture that repeatedly invokes the same underlying model as fresh reasoning roots. Each root receives the original problem and context, plus a compact summary, plain-text blackboard, and durable task artifacts from earlier roots rather than the full conversation. The design targets extraction, counting, ordering, and multi-hop tasks where mistakes in one long trajectory can propagate to the final answer. The authors describe the handoff mechanism, artifact workspace, and evaluation protocol, and investigate whether staged fresh-context continuation improves accuracy over direct answering, including recursive tool-calling baselines.