HKUST paper: policy secretly encodes harness preferences; return = harness ceiling minus mismatch
HKUSTGZ · hf · 2026-10-28
A paper by Jian Mu, Ruibo Shen, and Yao Shu (HKUST Guangzhou/HKUST) asks what an LLM agent policy actually learns from its harness—the prompts, tools, and control logic—and what survives when the harness changes.
Key findings:
- Under KL-regularized control, the log-odds of an optimal policy recover how strongly the harness favors each action, so a trained policy carries its harness's preferences
- Deployed return equals the harness's ceiling (the best return it offers) minus the policy's mismatch (return lost to its own suboptimal decisions)
- Training only changes the mismatch term; changing the harness moves both ceiling and mismatch
Implication: harness comparisons must separate these two terms to predict training gains—deployed performance reflects harness potential, not just policy skill.
More from coding & agent
- VibeVideoBench: multimodal agents coding videos often fail on time management — Muxi Diao · 2026-10-28
- Dev builds open-source Excalidraw MCP after AI-generated diagrams kept overlapping text and arrows — pikahikmag · 2026-10-28
- Amazon tells engineers to get customer-facing or lose their future as AI reshapes coding — DynamicWebPaige · 2026-10-28
- DeepSeek vs Haiku pricing debunk: accounting for peak rates and token use, 1017 is 2-2.5x cheaper — teortaxesTex · 2026-10-28
- Viral screenshot shows an agent spawning 32 parallel workspaces in ~30 seconds — cephaloform · 2026-10-28
- Grok Bot 0.78.1 hides unreleased per-bot logins with manager-only team sharing — faraz0x · 2026-10-28