BioXArena-Pro turns biomedical modeling into 64 research-grade tasks: the best agent scores 55.1 and no task saturates

Evaluation of biomedical modeling agents now rests on 64 research-level tasks, eight each in chemical biology, imaging, network biology, perturbation biology, phenotype–disease modeling, sequence and genomics, single-cell and spatial omics, and structural biology. The earlier BioXArena was approaching saturation: on its 76 tasks the two strongest configurations already averaged 66 of 100, and on 12 tasks the best of four configurations came within 10 points of the maximum. For the four agent configurations run on both benchmarks, mean scores on the same 0–100 scale are 11 points lower on BioXArena-Pro than on BioXArena (51.6 versus 62.8), and near-saturated tasks with a best score of at least 90 fall from 12 of 76 to 0 of 64.

Each task is a self-contained capsule: the agent receives public training data, unlabelled test inputs, native-format assets and a description, and must return test-set predictions within one GPU and a two-hour budget. Relative to BioXArena, the median task provides 19,663 training units instead of 9,042 and the public data total 179.6 GiB instead of 71.6 GiB, and only 16% of tasks can be trained from a single train and test table (BioXArena: 50%). 83% of tasks use a split that holds out an entire group rather than random rows (BioXArena: 66%), mostly biological or clinical units such as species, cohorts, donors, patients, cell lines and protein families; test labels stay private and every score maps to a common 0–100 scale.

The 22 configurations follow two designs: three vendor coding command-line interfaces each paired with two of its vendor's models, and four model-agnostic scaffolds each crossed with four backbones. Codex CLI with GPT-6 Astra reaches 55.1 (95% confidence interval 50.0–59.8), followed by Claude Code with Opus 5.5 and OpenCode with GPT-6 Astra or Opus 5.5 (all 53.2), while the minimal scaffold with DeepSeek V4.1 Flash is last at 33.6. Across the 16 cells of the four-by-four design the backbone accounts for 53% of the variance in mean score and the scaffold for 37%, and a full coding scaffold adds more as the backbone weakens, from 6.7 points with GPT-6 Astra and Opus 5.5 to 14.2 with DeepSeek V4.1 Flash.

Every task carries a reproduced published method as a reference, a method published for that dataset on 33 tasks and the strongest published method for the same problem class on the other 31. The references average 48.6, below Codex CLI with GPT-6 Astra at 55.1 and Claude Code with Opus 5.5 at 53.2, and the agent is higher on 53 of the 64 tasks. On the nine tasks whose test set stays close to the original study and whose published recipe applies unchanged, the methods score 54.7 and the strongest agent 54.9, a non-significant difference; on the 48 tasks rebuilt with new splits, removed inputs or new metrics, the references trail by 8.2 points (confidence interval 6.2 to 10.1).

The demonstrated scope is 22 agent configurations on 64 biomedical modeling tasks, each run on one A100 with a two-hour budget and scored on a common 0–100 scale; the 0–100 scale is an index rather than a fraction solved, so means across tasks and domains should be read comparatively.

Lookup of public source data cannot be fully prevented when agents have network access; the audit found no download of a source dataset and no database query with test inputs, but 40 runs used public checkpoints whose training data may overlap the test samples, and that overlap was not measured. De-identification removes prior knowledge that a domain method or foundation model could legitimately use. Most configurations were run once per task, and differences among the top configurations are smaller than their bootstrap intervals. Human experts were not measured directly, so the references are best read as a lower bound on expert performance.