Updated 21 July 2026 · published continuously

What we are
learning about
making molecules.

Papers, benchmarks, and engineering notes, kept in one place, updated as the work happens rather than when marketing is ready.

ICML 2026 papers: 02
Blogs: 02

ICML 2026

Papers at ICML 2026.

Two papers accepted this year. Each opens on arXiv or OpenReview in a new tab.

ICML 2026arXiv

RETROSPECT: RETROsynthesis via Sequential Prediction, and Chemically Transformed-ranking

RETROSPECT is our proprietary retrosynthesis engine, combining transformer-based route generation with learned candidate selection, delivering 55.0% top-1, 86.2% top-10, and 99.9% valid retrosynthesis predictions, with learned reranking lifting first-choice selection to 59.4%, creating a scalable, proprietary foundation for AI-driven synthesis planning.

By Raja Sekhar Pappala, Shreyas Vinaya Sathyanarayana, Ronit Kumar Choudhary, Arjun Verma, Deepak Warrier

Read the paper

ICML 2026OpenReview

LabProc and Tacit: Quantifying the Visual-Textual Prior Gap in Autonomous Laboratory Perception

LabProc is our benchmark for laboratory procedure understanding, spanning six tasks from static state recognition to motion-only reasoning. On its hardest motion-only subset, our 300M-parameter Tacit model beats Claude Opus by 9.1 points, despite being about 1000x smaller and trained in just 28 minutes on a single H100.

By Jagrati Gogia, Monish Kaul

Read the paper

Blogs

Notes from the work.

Engineering notes from the team building the Chemistry World Model.

ModelsChemistry

Sphere-Native Transformers for Molecular Property Prediction

When a domain carries structural priors as rich as chemistry's, it warrants a domain-native architecture rather than a generic transformer rescued by scale.

By Deepak Warrier and Raja Sekhar Pappala

15 May 2026 · Blog · 11 min read

Read the blog

ModelsOptimization

Bayesian Guided Knowledge Graph Embedding for Multi-Campaign Reaction Optimization

Every new reaction campaign shouldn't start from scratch. Introducing BEACON, our framework that wires tens of thousands of historical reactions directly into the optimization loop.

By Jahnvi Yadav and Raja Sekhar Pappala

20 May 2026 · Blog · 10 min read