WHALE: A Simple Recipe for Joint Harness-Weight Optimization
WHALE alternates model fine-tuning with harness search instead of treating either side as fixed.
The paper argues that agent gains can stall when weights and executable harness code are optimized separately. WHALE updates the model under the current harness, then searches for a better harness for the updated model. In tests with Qwen3.5-2B/4B agents on search QA, math, and chess puzzles, it beats weight-only, harness-only, and Fast-Slow Training baselines by 4.15 to 24.38 percentage points in best mean@8 accuracy. Source: HF Daily Papers' note.
The paper argues that agent gains can stall when weights and executable harness code are optimized separately. WHALE updates the model under the current harness, then searches for a better harness for the updated model. In tests with Qwen3.5-2B/4B agents on search QA, math, and chess puzzles, it beats weight-only, harness-only, and Fast-Slow Training baselines by 4.15 to 24.38 percentage points in best mean@8 accuracy. Source: HF Daily Papers' note.
score 5