WHALE Paper Proposes Alternating Weight and Harness Optimization

KRAFTON released and open-sourced WHALE (Weight-Harness Alternating LEarning), a paper by Kangwook Lee et al. of UW-Madison also posted to arXiv with a blog and code. The recipe alternates model weight updates with harness search—the executable code managing tools, context, and control flow—jointly optimizing LLM agents on reasoning tasks and reporting 4-24 point accuracy gains, breaking the paradigm of tuning only weights or only the harness.

Confirmed

• Core observation: harness optimization is sample-efficient but plateaus quickly; if budget allows, optimizing weights too yields larger gains. An LLM agent is weights plus harness, and optimizing only one end is limited by the frozen other.
• The method alternates between weight updates and harness search; paper, blog, and code are public.
• Kangwook Lee separately argued in "Why we should stop designing harnesses for AI agents" that the best harness may be one agents engineer for themselves, and that continued human hand-design becomes a bottleneck.

Why it matters

• WHALE offers a viable path for co-evolving harness and model, treating scaffolding not as fixed externals but as jointly trainable objects.
• Asked whether harness modifications would naturally become part of rollouts, Lee replied that with data, weights and harness can and should be trained together, promising a longer thread.
• He clarified that the meta-harness vs. trained-harness distinction in the paper was a deliberate choice for the studied setting, not a necessity; once it disappears and agents are capable enough, harnesses could evolve on their own during rollout—evidence the author is actively pushing the agents-build-their-own-harness agenda.

2026-09-03 ~ 2026-09-04 · 7 related posts

• [source] WHALE Paper: Alternating Weight and Agent Harness Optimization Gains Up to 24 Points — Kangwook_Lee · 2026-09-03
• Kangwook Lee: with data, you should train both model weights and the harness together — sheggle_ · 2026-09-03
• WHALE: a simple recipe to jointly optimize an LLM's weights and harness — lateinteraction · 2026-09-04
• [source] KRAFTON's WHALE jointly optimizes model weights and harness for agent performance — KRAFTON · 2026-09-04
• [source] Kangwook Lee: agent harness improvement can happen inside rollouts once agents are capable enough — Kangwook_Lee · 2026-09-04
• Prof. Kangwook Lee: stop hand-designing harnesses for AI agents — let them build their own — Kangwook_Lee · 2026-09-04

1 near-duplicate retellings: Kangwook_Lee