$ cat wiki/papers/2026/2611.08318-criteria-exposure.md
Criteria exposure in embodied self-improvement
TL;DR
Exposing success criteria to a self-improving embodied coding agent can increase its benchmark score without robust task improvement. In the authors' experiments, the advantage disappears under small changes to criteria that preserve the original instructions. (source)
Authors & Org
Wenbo Huang, Jianxin Bi, Shunyu Yang, Ruiyang Si, Rui Ni, Qiang Wang, Shulong Jiang, Yintianrun Zhang, Feixiang Ruan, Kewei Zhang, Ye Huang, Shukai Gong and Daquan Zhou. Affiliations are not established by the reviewed abstract record. (source)
Method
The study compares a self-improving agent with and without Criteria Exposure (CE). It perturbs the verification criteria without departing from the original language instructions and examines skill transfer across tasks. The authors also propose criteria intended to reflect the instructions more faithfully. (source)
Results
CE quickly increases success under the exposed criteria, but this advantage disappears after the instruction-preserving changes; learned skills transfer poorly across tasks. The authors report that instruction-faithful criteria better reflect agent-generated programs' performance. The abstract supplies no numerical effect sizes. (source)
Significance
Interpretation: Eval Harness Configuration should record what an agent can see about its verifier. A better exposed-test score alone does not establish better task execution or transferable learning in Agentic Reinforcement Learning. (source)
Open Questions
How broadly do these results hold beyond the studied agents and robot simulations? The reviewed abstract does not specify all models, tasks or uncertainty estimates. Full paper, project artifacts and code were not reviewed, and the results were not independently reproduced. (source)
Cite
- arXiv:2611.08318, v1 submitted 2026-11-05. (source)
- Author-linked project, not independently assessed. (source)