AI Models Expand into Scientific Research with New Benchmarks Emerging

연구/벤치마크 | Fri Jun 19 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 6 sources

AI applications in research progressed across rare disease diagnosis, black hole simulation, and chemistry analysis, alongside advances in agent and fine-tuning benchmarks.

Analysis

[OpenAI · Boston Children's Hospital] published research using o3 Deep Research for diagnosing rare pediatric genetic diseases ^[1]

Re-analyzed 376 unresolved cases
Achieved 4.8% additional diagnosis rate (18 confirmed cases)
Published in NEJM AI
Evidence-linked hypothesis generation workflow

[Chi-kwan Chan · OpenAI Codex] accelerated black hole plasma simulations using Codex ^[2]

Event Horizon Telescope collaboration research
Targeting imaging of supermassive black hole in M87 galaxy
Magnetic field simulation with trillions of particles
Validation of general relativity

[ServiceNow · Hugging Face] released MosaicLeaks benchmark and PA-DR training method ^[3]

Evaluates privacy leakage of deep research agents
Measures three types of leakage: intent/answer/full-information
Strict chain success rate 48.7% → 58.7%
Information leakage reduced from 34.0% to 9.9%

[Hugging Face PEFT] published comparative analysis of fine-tuning techniques beyond LoRA ^[4]

95–98% of PEFT checkpoints are LoRA
Introduced alternatives including LoHa
AdaLoRA
DoRA
Unified API and quantization support
Memory efficiency and resistance to catastrophic forgetting

[Hugging Face] released agent-friendliness benchmark for the Transformers library ^[5]

Harness based on pi coding agent
Sweep evaluation across models × revisions × tasks
Measures process and token usage in addition to final answer
1.3–6x token savings in hf CLI case

[Anthropic] announced project to enhance Claude's chemistry capabilities ^[6]

Evaluation of NMR spectrum analysis capability
Collaboration with synthetic
computational
and analytical chemists
Translation capability between molecular structure representations
Applications in retrosynthesis and reaction prediction

AI Models Expand into Scientific Research with New Benchmarks Emerging

Analysis

Sources