Language Model Harnesses Enable Compositional Generalization

Modern post-training relies on continuously stacking more environments and longer training data. The root cause is that Transformers lack compositional generalization—the ability to combine familiar problems to solve new ones.

This piece focuses on the harness—the program that mediates input between the external environment and the LLM—as a solution. A good harness decomposes complex problems into simple sub-problems, so that each LLM call receives an in-distribution prompt.

The experiments used RLM (Recursive Language Model), a harness that offloads context to recursive sub-calls, with RL training conducted on top of Qwen3-30B-A3B-Instruct-2507. Results:

• Training only on short tasks generalizes to tasks 8-32x longer
• For the same improvement in training performance, evaluation performance improves about 10x more than the base Transformer
• Training on one domain transfers to other domains that use similar decomposition strategies

The cause of this effect is that the RLM harness induces an equivalence relation among potentially similar tasks. Tasks with similar structure produce nearly identical token trajectories in the RLM's main context, making them effectively the same problem from the model's perspective.

These results suggest that, alongside adjustments to model architecture or training recipes, harness design can be a key lever for reducing post-training costs and expanding the range of tasks that can be covered.