TorchLean (lean-dojo, 2026)
github.com/lean-dojo/torchleanFirst unified Lean 4 framework for neural-network specification, execution, and verification; tensor shapes are part of the types, models are executable Lean programs, and the same definitions can be used by training code, graph transformations, certificate checkers, and proofs with CPU/CUDA backends (123+ stars, MIT License)
Sourced from
- Awesome AI for Science — github.com/lean-dojo/torchlean
- GitHub — github.com/lean-dojo/torchlean
Related resources
LLMs as copilots for theorem proving in Lean 4, exposing native tactics (`suggest_tactics`, `search_proof`, `select_premises`) that embed language model inference and premise retrieval directly inside the Lean proof environment, supporting local CTranslate2/CUDA inference as well as remote model APIs for interactive and automated proof search (Caltech & NVIDIA, NeurIPS 2024, 1.2K+ stars)
Open-source toolkit and benchmark for learning-based theorem proving in Lean, providing programmatic Lean interaction, a 98K+ theorem dataset extracted from 217 Lean projects, and ReProver—the first retrieval-augmented LLM-based theorem prover for Lean—with reproducible training pipelines underpinning much subsequent Lean prover research (Caltech & NVIDIA, NeurIPS 2023 Outstanding Paper, Datasets & Benchmarks)
Neural differential equations in PyTorch
Scientific equation discovery and symbolic regression using LLMs, combining code generation with evolutionary search (ICLR 2025 Oral)
ibm-research/materials.smi-ted
by ibm-researchWelcome to IBM's series of large foundation models for sustainable materials. Our models span a variety of representations and modalities, including SMILES, SELFIES, 3D atom positions, 3D density grids, molecular graphs, and other formats.
NathanRoll/h01-cortex-snn
by NathanRollThe complete neuron-to-neuron connectivity matrix extracted from a nanometer-resolution reconstruction of human temporal cortex (H01 dataset, Google/Harvard/Lichtman Lab).