WHALE: A Simple Recipe for Joint Harness-Weight Optimization

Key points

It alternates between online rejection-sampling fine-tuning with the current harness fixed and Meta-Harness code search with the updated model fixed.
Evaluations of Qwen3.5-2B and 4B on search-based question answering, mathematical reasoning, and chess puzzles achieved peak accuracies 7.67–24.38 percentage points higher than single-component baselines and 4.15–13.00 percentage points higher than FST, which changes only prompts.
The harness was the main bottleneck for search tasks, but mathematical tasks required weight updates first before the same harness search became effective, showing that the bottleneck varies by task.
Interleaving multiple small updates improved accuracy and rollout efficiency over a staged approach that trains all the weights and then searches for a harness once, providing grounds for designing agent-training pipelines as a joint optimization problem.
Important limitations are that validation covered only Qwen3.5-2B and 4B and three tasks, and that the rollout-cost comparison excluded compute used by the model proposing harnesses.

External research summaries. These are not HDATF publications or measured product results.