RLM-Qwen3.5-35B: Training Natively Recursive Language Models

Training Recursive Language Models (RLMs) via reinforcement learning on the Tinker API.

Paper (PDF) | Model on HuggingFace

Results

+21.7pp average improvement across 14 benchmarks via RS-SFT on 3,644 self-mined trajectories. 13 wins, 1 loss vs base.

Structure

scaffold/     # RLM runtime (repl.py, rlm.py, llm_query.py)
eval/          # Evaluation harness (14 benchmarks)
training/      # Training scripts (GRPO, RS-SFT)
scripts/       # Data pipeline & utilities