WHALE: A Simple Recipe for Joint Harness-Weight Optimization

WHALE. A Simple Recipe for Joint Harness-Weight Optimization. By Haechan Kim and four others, from Krafton, KAIST, and one other institution. Published on the thirty-first of August. Its code repository is worked out: about one thousand five hundred source files, with a readme and a licence.

How well an agent performs depends on two things at once. There are the model's parameters, and there is the harness: the executable code that manages context and control flow around the model. Optimize either one alone and the system can end up bottlenecked by the other, still frozen. The two interact. Updating the weights can change which harness is effective, and updating the harness can change which of the model's capabilities get exposed. Existing methods for joint adaptation optimize weights together with textual prompts, but they leave the broader harness fixed.

Weight-Harness Alternating LEarning, or WHALE, is a simple recipe that alternates between two phases. First it updates the model under the current harness. Then it searches for a better harness under the model it has just updated. Each side gets to move while the other holds still, and then they trade places.

The two phases are instantiated with existing tools. Weight updates use online rejection-sampling fine-tuning. Harness search uses Meta-Harness. The key design choice is when to switch. Switch too eagerly and noise looks like progress. Linger too long and one side over-optimizes against a counterpart that is about to change. WHALE handles this in one of two ways: either fixed durations for each phase, or an adaptive patience rule applied over the training signals.

The experiments use Qwen3.5 agents at two billion and at four billion parameters, across three domains: search question answering, mathematical reasoning, and chess puzzles. Against weight-only optimization, harness-only optimization, and Fast-Slow Training, WHALE comes out ahead by between four point one five and twenty-four point three eight percentage points in best mean at eight accuracy.

The domains also show that either component can be the bottleneck. In SearchQA, harness search matches the peak accuracy of weight-only training with far fewer rollouts. In math, harness search improves accuracy only after a weight update has happened. And small interleaved updates beat a stagewise schedule that does weights first and harness second, in accuracy and in rollout cost alike. The code is publicly available.

The argument is about where the ceiling sits. Hold the harness fixed and a better model may be held back by old scaffolding. Hold the weights fixed and a better harness may have nothing new to expose. Letting the two take turns, in small steps, means neither stays the limit for long.