New WHALE method jointly optimizes AI agent weights and harness code

Researchers have introduced WHALE (Weight-Harness Alternating LEarning), a novel method for optimizing AI agents by jointly adapting model weights and harness code. This approach alternates between updating model parameters and searching for improved harness code, addressing the issue where optimizing one component in isolation can bottleneck the system. Experiments with Qwen3.5-2B/4B agents across search, math, and chess tasks demonstrated that WHALE significantly outperforms existing methods, including weight-only and harness-only optimization, by achieving higher accuracy with fewer rollouts. AI

IMPACT This new optimization technique could lead to more efficient and higher-performing AI agents across various complex tasks.

RANK_REASON The cluster contains a research paper detailing a new method for optimizing AI agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Fast-Slow Training
• Hugging Face
• Meta-Harness
• Qwen3.5-2B
• Qwen3.5-4B
• SearchQA
• WHALE

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →