Agentic AI tackles games, genomics and Earth system modeling

A roundup of arXiv preprints: the AAArena benchmark shows Opus 5.5 with Claude Code winning 6 gold medals across 12 adversarial games without weight changes. The ARGUS framework cuts hallucinations in genomics, and legoESM is the first differentiable Earth system model built by AI coding agents.

• AAArena: 12 adversarial games, Opus 5.5 with Claude Code took 6 gold medals
• ARGUS wraps 458 DNABERT models and verifies against ADASTRA, JASPAR and ENCODE
• legoESM on JAX scales to kilometer-scale climate simulations on GPUs
• DataSense-Bench: AI agents rank training subsets inconsistently