Teaching AI
to do real physics

We build and rigorously test AI systems that can carry theoretical physics from a well-posed question to a calculation

Read the log →See the benchmark →

Research log

NLL-log/2607.001 · Lattice gauge theory · July 2026

Calculating the glueball mass with Codex

An AI agent ran a first-principles lattice calculation of the SU(3) scalar glueball mass — writing its own GPU-accelerated code, blinding itself to the target, and gating its own results — landing on a qualified continuum estimate of M = 2.13(43) GeV.

Benchmarks

NLL-bench/2607.001 · July 2026

NLL Physics Research 1

Twelve research-grade problems reconstructed from June 2026 preprints and run with the source paper absent from the environment. The strongest model solved 7 of 22; 14 problems are still standing against the whole panel.

  1. GPT-5.6-solCodex · xhigh7/2232% solved
  2. Qwen 3.8 Max PreviewOpenCode3/2214% solved
  3. Claude Opus 4.8Claude Code · xhigh3/2214% solved
  4. Grok 4.5Grok Build · high3/2214% solved
  5. Gemini 3.6 FlashOpenCode · high2/229% solved
  6. GLM 5.2OpenCode · xhigh1/205% solved
  7. Nemotron 3 UltraOpenCode · high1/225% solved