# AI Agents’ Rankings Matched Each Other More Closely Than Lab Results By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/57702 Published: 2026-11-15T06:53:57.000Z Updated: 2026-11-15T06:53:57.000Z Section: Technology > Averaging rankings improved scores on 73 public tasks, but strong agreement did not prevent a poor match with lab outcomes on one oncology task. AI agents’ rankings tracked one another more closely than laboratory results across 100 public science tasks, while averaging their rankings improved scores but preserved a shared mistake on an oncology task. UniPat AI’s Science Prediction Universe benchmark tested 13 agent configurations spanning 12 models in chemistry, materials science, biology and medicine. The benchmark also included 16 private tasks based on unpublished experiments. For each task, an agent received an experimental setup and candidate conditions, then ranked the candidates by an outcome measured in a lab. The measured values and grading code were kept outside the agents’ environment. Across the public tasks, the mean pairwise Spearman rank correlation between configurations was 0.52. The correlation between their rankings and measured outcomes was 0.33, and agreement was higher on 81 of the 100 tasks. Spearman correlation measures how similarly rankings order items; it does not show the percentage of predictions that were correct. A consensus ranking formed by averaging each candidate’s position across configurations scored 0.081 points above the average individual submission. Its 95% confidence interval was 0.054 to 0.110, and it improved scores on 73 public tasks. The paper’s authors wrote, “Averaging cancels errors that differ between configurations and keeps those they share.” One oncology task showed the limit of that approach: 11 configurations recorded 0.90 agreement but averaged 0.27 against measured outcomes. Across tasks, agreement and scores were associated, with a Spearman correlation of 0.56. On the private tasks, the benchmark had 159 scored runs out of 208, and 12 of the 13 configurations ranked results significantly better than chance. Coverage varied across the full benchmark. All 13 configurations were scored on the same 28 tasks: 27 public and one private. The study had no human prediction baseline, so its results do not establish how the agents performed compared with scientists. The distinction between agreement and correctness also appeared in [an earlier TokenPost legal AI test](), where repeated model runs agreed with one another more often than with gold annotations. ## Links in this article - [an earlier TokenPost legal AI test](https://www.tokenpost.com/news/technology/48407)