Self-reflective program search matches or outperforms recursive language models for long-context tasks, suggesting recursion itself is not the primary driver of performance.
arXiv · March 18, 2026 · 2603.15653
The Takeaway
It demonstrates that simple uncertainty-aware search over context-interaction programs yields up to 22% improvement over explicit recursive architectures, simplifying the design of long-context agents.
From the abstract
Long-context handling remains a core challenge for language models: even with extended context windows, models often fail to reliably extract, reason over, and use the information across long contexts. Recent works like Recursive Language Models (RLM) have approached this challenge by agentic way of decomposing long contexts into recursive sub-calls through programmatic interaction at inference. While promising, the success of RLM critically depends on how these context-interaction programs are