Chained Recursive Language Models for Multi-Iteration Reasoning

2026-08-05

AI Summary

This work addresses the challenge that large language models often accumulate errors in complex reasoning over long contexts due to limitations in a single inference trajectory. To mitigate this, the authors propose Chained Recursive Language Modeling (Chained RLM), which recursively invokes the same base model over multiple rounds. Each round starts afresh from the original question and full context, while intermediate states are communicated across iterations via lightweight artifacts—such as textual summaries, a shared blackboard, and task-specific persistent storage—enabling staged, inspectable, and correctable reasoning. Empirical results demonstrate that this approach significantly outperforms both direct answering and conventional recursive tool-augmented methods on tasks requiring extraction, counting, sorting, or multi-hop reasoning, thereby substantially improving overall accuracy.

Abstract

Long context reasoning in large language models (LLMs) is usually constrained by the fact that a single inference trajectory has to simultaneously explore the context, store intermediate state, verify evidence, and produce the final answer. This becomes particularly difficult in tasks that require extraction, counting, ordering, or multi-hop reasoning, where an early mistake can propagate until the final response. In this work, we propose Chained Recursive Language Models (Chained RLM), an inference-time architecture, in which the same underlying model is called repeatedly as a sequence of fresh reasoning roots. Each root receives the original problem and context, but does not inherit the full conversational history. Instead, it receives a compact plain-text summary, a plain-text blackboard, and some durable task-specific artifacts written by predecessor roots. The motivation is to manage the context by chopping into partial tasks rather than one large inference response; in each staged computation, intermediate artifacts can be inspected, corrected, and extended by a later fresh inference by the same model. We describe the system model, handoff mechanism, artifact workspace, and evaluation protocol for this system. We study when fresh-context artifact continuation gives a measurable gain in accuracy over direct LLM answering even with recursive tool-calling.

Problem

• long context reasoning
• multi-hop reasoning
• intermediate state propagation
• large language models
• reasoning error

Innovation

• Chained Recursive Language Models
• multi-iteration reasoning
• artifact-based inference
• fresh-context continuation
• error propagation mitigation