WHALE Paper: Alternating Weight and Agent Harness Optimization Gains Up to 24 Points
Kangwook_Lee · x · 2026-09-03
A new arXiv paper from Kangwook Lee's group, WHALE (Weight-Harness Alternating LEarning), proposes a simple recipe for jointly optimizing an agent's model weights and its executable harness code:
- Motivation: Agent performance depends on both model parameters and harness code; optimizing either in isolation leaves the system bottlenecked by the frozen counterpart — weight updates change which harness is effective, and harness updates change which capabilities are exposed.
- Method: Alternate two phases — fine-tune the model under the current harness (online rejection-sampling fine-tuning), then search for a better harness under the updated model (Meta-Harness). Switching uses fixed phase durations or an adaptive patience rule to separate real gains from noise.
- Results: With Qwen3.5-2B/4B across search QA, math reasoning, and chess puzzles, WHALE beats weight-only, harness-only, and Fast-Slow Training by 4.15–24.38 points in best mean@8 accuracy. Either component can be the bottleneck: harness search matches peak weight-only accuracy with far fewer rollouts on SearchQA, but math improves only after a weight update.
More from coding & agent
- NVIDIA's NOOA represents AI agents as plain Python objects, ditching prompt/config sprawl — burkov · 2026-09-03
- NVIDIA's NOOA turns agent scaffolding into ordinary Python objects — SimplyAnnisa · 2026-09-03
- Codex computer-use pro-tip: a shared task-title convention to track agent browser tabs — danshipper · 2026-09-03
- Pro tip: use a shared emoji task-title convention to track Codex computer-use sessions across tabs — danshipper · 2026-09-03
- Google ships a Genkit Go skill for agents building AI apps in Go — adonis_singh · 2026-09-03
- Codex Deleted the 6 Test Cases That Uncovered a Bug So Its Tests Would Pass — PuzzleheadedAnt9503 · 2026-09-03