Wavefunction Bench Tests Coding Agents on Quantum Chemistry: Only 3 of 9 Systems Passed
CatAstro_Piyush · x · 2026-11-06
Primateria released Wavefunction Bench, testing whether AI coding agents can improve scientific simulation software rather than typical app code.
- Each of 8 models got 8 hours on one GPU to modify Orbformer, a neural network approximating solutions to the electronic Schrödinger equation, and train a wavefunction for a specified molecule across 9 molecular systems, including transition-metal compounds outside its pretraining.
- Scoring: independently validated variational energy within 1.594 mHa of targets; LiH and CO₂ targets from high-accuracy theory, others combine calculated atomic energies with corrected experimental atomization energies.
- Results: success on only 3 of 9 systems. LiH: all 8 models passed at least once (31/32 trials). NaCl: only GPT-6 Astra, Claude Opus 5.5 and Muse Spark 1.3. MgCl₂: only Astra and Opus. The other 6 systems: zero passes.
- Context: Primateria builds simulators predicting molecular/material properties at experimental fidelity before synthesis.
More from coding & agent
- AI revives Facebook's abandoned Prepack, speeding up React renderToString 5.4x — Vjeux · 2026-11-06
- If You're Compute-Constrained, Don't Burn Agent-Compute on Social Agents — nptacek · 2026-11-06
- AI Engineer SH talk: models see the world through their action space, not ours — vincent_koc · 2026-11-06
- teortaxesTex: Singularity still needs bigger models; public AGI easy to dunk on — teortaxesTex · 2026-11-06
- Late to a talk, his AI agent tapped the live transcript and took notes for him — vincent_koc · 2026-11-06
- VoidZero October Update: Rust-Based TypeScript 8x Faster, SDK Goes Open Source — cnakazawa · 2026-11-06