AI 代理與科學運算
OPERA Constrains Experimental Agents with Physics Residuals, Cutting Reward-Seeking Decisions with No Real Improvement from as High as 39.0% to 1.9%
Rather than returning only a single score, OPERA gives language models interpretable physics residuals before they select, compose, or generate optical operations. Three optical tasks show that residuals can curb specification gaming, but the public code does not include experimental instrument control or the complete statistical pipeline.
Language-model agents can improve a numerical score according to an evaluation function without necessarily improving the actual experimental outcome. Researchers from Tsinghua University and other institutions introduced OPERA, which represents beam shaping, structured-light 3D reconstruction, and interferometric measurement as a set of typed optical operators. An agent can select existing operations, chain multiple steps, or generate a new procedure from permitted primitives. Each execution checks optical-path connectivity and the resource budget.
The key change is to divide feedback into three layers: an agent-visible scalar score, L2; a residual vector, L1, describing insufficient coverage, model mismatch, or deviation from the target; and the true physical performance, L3, calculated from a hidden reference state and used only for offline evaluation. This makes it possible to directly identify decisions where “L2 increases while L3 does not improve.” In paired experiments covering 90 independent problems, such events accounted for 23.6% and 39.0% of decisions under neutral and score-seeking instructions, respectively, when only the score was provided. Adding residuals reduced those rates to 0.9% and 1.9%.
Across another 270 test problems, 75.8% of trajectories using operator–residual feedback reached the target and maintained it through completion, compared with 33.9% to 60.1% for four reference strategies. Average budget usage was also 61.7% of the limit, below the reference strategies’ 70.8% to 86.0%. The advantage appeared mainly under distribution shift; under familiar conditions, OPERA did not consistently outperform constrained Bayesian optimization or task-specific algorithms.
The team transferred fixed workflows selected in a digital twin to three physical instrument setups. In the primary structured-light setup, the average projection budget fell from 95.8% for fixed Gray code to 60.0%, although each configuration had only six paired replicates. The public repository can replay digital-twin trajectories and verify interfaces and hashes, but it does not include real-time instrument control, model-provider SDKs, statistical resampling, or the final figure-generation pipeline. The next question for engineering teams is whether this observable residual interface can extend to chemistry, biology, and materials experiments—and whether incorrectly designed residuals could create new optimization loopholes.