BioAgent Bench.

Can an AI agent run real bioinformatics, reliably, for the right reasons?

Read the paper Explore the code

BioAgent Bench is an evaluation suite designed for measuring the performance and robustness of AI agents in common bioinformatics tasks. The suite consists of manually curated end-to-end tasks (e.g., RNA-seq, variant calling, metagenomics) accompanied by task-specific prompts and concrete output artifacts to support automated assessment. We also introduce controlled perturbation testing as a form of adversarial evaluation, using corrupted inputs, decoy files, and prompt bloat to probe biological reasoning beyond pipeline completion rates.

Pipeline completion

Claude Opus 4.5100.0
Gemini 3 Pro96.6
GPT-5.292.5
Claude Sonnet 4.592.5
GLM-4.782.5
GPT-5.1 Codex Max81.7
Kimi K2 Thinking80.5
MiniMax M2.173.6
Qwen3 Coder68.8
Devstral 251267.9
Completion is the fraction of required pipeline steps completed. Results reflect model-plus-harness systems reported in the current paper.

Completion is only the baseline for reliability. We evaluated GPT-5.2 in the Codex CLI harness under three controlled perturbations: corrupting selected inputs, adding plausible but irrelevant decoy files, and padding prompts with related but unnecessary text. The agent missed corrupted inputs in 3 of 10 tasks, used decoys incorrectly in 2 of 10, and completed 28% fewer steps under prompt bloat. Across up to four clean runs per task, mean Jaccard overlap was 0.43 and mean Pearson correlation was 0.73, showing that even a completed pipeline can produce meaningfully different results from one run to the next.

Adversarial testing

Controlled perturbations and repeated clean trials with GPT-5.2 in the Codex CLI harness.

Corrupted inputs 3/10 missed
Decoy files 2/10 used
Prompt bloat 28% fewer steps

Relative steps completed; clean run = 100.

Repeated clean trials Outputs varied across runs

Continue reading