Recursive LM harnesses generalize from short tasks to jobs 8–32× longer
heghbalz · x · 2026-07-21
- A paper argues that language model harnesses can act as compositional generalizers.
- The core idea is to use recursive LMs that offload context and call sub-agents, so the main model repeatedly sees a simple, reusable structure rather than one giant task.
- The authors train on short tasks but report generalization to tasks 8–32× longer.
- They say each step stays locally in-distribution, and the model learns the decomposition pattern behind the task instead of memorizing the task itself.
- The framing shifts generalization from “just add more data and longer context” to a harness-design problem.
Related event: Research: Harness Drives Generalization in Reinforcement Language Models(9 posts)→
More from Research
- LFM2 tokenizer expansion cuts Thai tokens 4× and speeds on-device decoding up to 3.7× — maximelabonne · 2026-07-21
- Argus improves indoor panoramic 3D reconstruction with covisibility and geometry transformers — ducha_aiki · 2026-07-21
- Natolambert finishes a RLHF book built from years of fine-tuning and post-training work — Jeande_d · 2026-07-21
- WAIC robots are now hitting commercially useful success rates, says a recap — chris_j_paxton · 2026-07-21
- Unitree launches a remote real-robot benchmark on its own G1 fleet — chris_j_paxton · 2026-07-21
- Frontier models still show greediness and frequency bias, researchers say — m_wulfmeier · 2026-07-21