Recursive Language Models

arXiv:2512.24601v1

Summary: This paper evaluates Recursive Language Models (RLMs) against frontier models (GPT-5, Qwen3-Coder) across diverse tasks including deep research, information aggregation, and code repository understanding. The study compares RLMs against standard direct LLM calls, context compaction, and retrieval tool-use agents. Results suggest robust capabilities for self-improvement and complex reasoning.