arXivPaperNeeds Review

Chained Recursive Language Models for Multi-Iteration Reasoning

Chained RLM repeatedly gives the same model a fresh context while preserving summaries, a blackboard, and artifacts—a concrete harness pattern for limiting error buildup.

arXiv · Aug 5, 2026

Open Source Open Markdown Open JSON

Source Summary

Chained RLM splits long-context work across repeated calls to the same model. Every **fresh reasoning root** receives the original task and context plus a **plain-text summary**, a **plain-text blackboard**, and durable artifacts rather than the full prior conversation.

Practical Implication

Agent builders can treat context resets as an architectural tool: keep the source problem stable, externalize intermediate state, and let later calls inspect or correct artifacts. This pattern targets extraction, counting, ordering, and multi-hop tasks where early errors otherwise persist.

Agent-Ready Context

Chained RLM splits long-context work across repeated calls to the same model. Every **fresh reasoning root** receives the original task and context plus a **plain-text summary**, a **plain-text blackboard**, and durable artifacts rather than the full prior conversation.

Agent builders can treat context resets as an architectural tool: keep the source problem stable, externalize intermediate state, and let later calls inspect or correct artifacts. This pattern targets extraction, counting, ordering, and multi-hop tasks where early errors otherwise persist.

The material describes the architecture and evaluation protocol but reports no quantitative results. The cost of repeated full-context calls, the quality required of handoff summaries, and the conditions under which this beats recursive tool calling remain open.

Connected Context · Feed7 Judgment

Chained RLM makes context reset a deliberate long-horizon control: each iteration starts from the original problem while inheriting only summarized state, a blackboard, and inspectable artifacts. This strengthens prior durable-state harness patterns and contrasts with recursive inspection inside one main context, but offers no results showing when repeated full-context calls justify their cost.

Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General ReasoningThe long-horizon account identifies compaction and durable external state as core requirements; Chained RLM turns those requirements into a concrete reset-and-handoff architecture.RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AIBoth keep large source material outside conversational history, but conventional RLM returns bounded evidence to a main context whereas Chained RLM repeatedly creates fresh roots connected by summaries, a blackboard, and artifacts.huangruiteng/loopxLoopX persists goals, gates, ownership, evidence, and quotas across bounded turns; Chained RLM supplies a complementary mechanism for transferring reasoning state between fresh model calls.Effective harnesses for long-running agentsBoth replace reliance on chat history with durable files and bounded sessions, though Chained RLM preserves the entire original task and context at every new reasoning root rather than advancing one feature per session.

Context Map

agentcodingresearch#harness-engineering#context-engineering #agent-memory

Uncertainty

The material describes the architecture and evaluation protocol but reports no quantitative results. The cost of repeated full-context calls, the quality required of handoff summaries, and the conditions under which this beats recursive tool calling remain open.