Gemini 4 Argon achieved state-of-the-art benchmark scores on the VALS Index Automation Bench, VALS Finance Agent, and Harvey Legal Agent benchmark

Overview Debates Related claims Topics

Handshake AI developed the VIALS benchmark to evaluate the capability of AI models to interpret life science laboratory assays

The VIALS benchmark evaluates model interpretation on tasks including plasmid DNA sequence percentages, molecular hydrogen bond counts, and medical histology images

Handshake AI developed the ATLAS finance benchmark, comprising 100 tasks evaluated across 13 simulated financial environments

Frontier models including MiniMax, GLM, and Grok score below a 20 percent pass rate on the VIALS benchmark

Autonomous AI agents provided with scientific Python libraries like SciPy and OpenCV achieve under 60 percent accuracy on VIALS benchmark tasks

Frontier AI models achieve a pass rate of under 20 percent on the ATLAS finance benchmark

Out of the box, frontier AI models achieve pass rates of 50 percent or worse across the VIALS life science benchmark

Tasks in the ATLAS finance benchmark require constructing financial spreadsheets containing over 4,000 active cells and 2,600 live formulas

Large language models generate outputs based on predictive statistical weights between words learned from human training data

Undergraduate researchers at Princeton University developed the SWE-bench software engineering benchmark

Google reported Agentic Video Understanding achieves up to 88% higher token efficiency and approximately 7% higher accuracy on large files

In Ballot Bench's House and Senate races, frontier artificial intelligence agents outperformed human election modelers while performing slightly worse than prediction markets

Thoughtful developed Ballot Bench, a benchmark evaluating nine frontier artificial intelligence agents forecasting the 2026 United States midterm elections

In Ballot Bench, each artificial intelligence agent received an isolated Modal sandbox, a $500 budget for paid news subscriptions and data, and seven days to complete forecasts

In Ballot Bench's Senate forecasting, Claude Fable 5.1 achieved the highest Brier score among artificial intelligence models while trailing human forecaster Nate Silver

In Ballot Bench, OpenAI's GPT-6.1 SOL aggregated election forecasts by analyzing the historical error shifts of human modelers over the decade prior to 2026

In Ballot Bench, Anthropic's Claude Opus 5.5 deployed approximately 170 to 180 sub-agents to parallelize state and ballot initiative polling research

METR found that the pace of capability improvement for Anthropic AI models roughly doubled over the course of a year

CellFlow applies flow-matching generative modeling to latent representations to predict cell state trajectories resulting from biological perturbations

The ARC Institute hosted the Virtual Cell Challenge to benchmark machine learning methods on single-cell perturbation prediction tasks