Expanding AI Research Ecosystem and New Trends in Benchmark Experiments

연구/벤치마크 | Thu Jul 23 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 7 sources

AI research and evaluation is diversifying through OpenAI's national science support, Anthropic's economic research fund, and pelican benchmark experiments.

Analysis

[OpenAI] announced US national science infrastructure support program ^[1]

Provided $4M in Codex access to approximately 2
000 researchers participating in the Genesis Mission
$3M API support for large-scale scientific campaigns
Up to $10M API usage available with $2.5M spend
Opened GPT-Rosalind bioscience-specialized model to national laboratory researchers

[Anthropic] unveiled research agenda for $200M Economic Futures Research Fund ^[2]

External research support to prepare for AI's economic impact
Selected 5 priority research areas (AI impact in the workplace
transition support
income support modernization
etc.)
Support centered on large-scale RCTs and creative pilot programs
Shift from existing small grants to large-scale projects

[Anthropic Economic Index] launched Economic Index connector for Claude ^[3]

Enables direct exploration of AI usage data within Claude
Supports queries on AI usage patterns by occupation
region
and task
Activated in claude.ai connector directory
no installation required
Compatible with all Claude models

[Pelicanmaxxing Experiment] released experiment verifying whether AI labs overfit to benchmarks ^[5]

Tested 1
008 SVG generations across 7 frontier models
Designed 48 prompts combining 8 animals × 6 vehicles
Used GPT-5.6 Luna as judge and Gemini 3.1 Flash-Lite for feature extraction
Verified whether the pelican-bicycle cell significantly outperforms other combinations

[Cactus Hybrid] released Gemma 4-based on-device confidence routing model ^[6]

Embedded a probe inside the checkpoint that returns a confidence score of 0 to 1
Automatic re-routing to a larger model when confidence < 0.85
Gemma 4 E2B Hybrid matches Gemini 3.1 Flash-Lite performance by routing only 15-35% of queries
Published benchmark results for ChartQA
MMBench
MMLU-Pro
and others

[DharmaOCR] demonstrated Brazilian Portuguese-specialized OCR model's superiority over newer models ^[7]

Superior performance compared to Mistral OCR4 and Unlimited-OCR
Stage 1 with supervised fine-tuning based on Portuguese files
Stage 2 stability reinforcement with Direct Preference Optimization
Achieved the highest extraction quality score and lowest degeneration rate

[NASA Roman Space Telescope] prepared first space-based active coronagraph deployment and exoplanet observation ^[4]

Equipped with approximately 300-megapixel wide-field camera
Expected to detect approximately 100
000 new exoplanets
Implemented active wavefront control using 2 deformable mirrors based on a 48x48 actuator grid
Approximately 100x wider imaging range compared to Hubble

Expanding AI Research Ecosystem and New Trends in Benchmark Experiments

Analysis

Sources