WHALE: A Simple Recipe for Joint Harness-Weight Optimization

  • 类型:arxiv
  • 标识:2609.00196
  • 链接:https://arxiv.org/abs/2609.00196
  • 主分类:evaluation
  • 形态:method
  • 被引:8
  • 被引来源:Semantic Scholar
  • S2被引:8
  • OpenAlex被引:0
  • 影响力被引:3
  • TLDR:Weight-Harness Alternating LEarning (WHALE), a simple recipe that alternates two phases: updating the model under the current harness, then searching for a better harness under the updated model with online rejection-sampling fine-tuning and Meta-Harness, is proposed.
  • OpenAlex ID:W7206174292
  • OpenAlex DOI:10.48550/arxiv.2609.00196
  • DOI:10.48550/arxiv.2609.00196
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://arxiv.org/pdf/2609.00196
  • OpenAlex更新:2026-09-06
  • 待LLM分类:否
  • 标题中文:WHALE:联合 Harness-权重优化的简洁方案
  • TLDR中文:本文提出 Weight-Harness Alternating LEarning (WHALE),一种简单的方法,交替进行两个阶段:先在当前 harness 下更新模型,再在更新后的模型下通过在线拒绝采样微调与 Meta-Harness 搜索更优的 harness。
  • 来源文件:
  • /inbox/tom/_candidates/2026-09-04-agent-rag-longcontext-candidates.json
  • [OpenAlex backfill]
  • [S2 enrich]