JEPA-Anything (Gen-Verse, arXiv 2026)
github.com/gen-verse/jepa-anythingDomain-agnostic predictive world-model framework built on joint-embedding predictive architectures (JEPAs), organizing latent world states into complementary factors via Orthogonal Predictive Factorization (OPF) with dedicated prediction pathways recombined for downstream readout, intervention prediction, planning, and multi-step rollout; shares one predictive core and latent state interface across cells, molecules, physical fields, control environments, and clinical trajectories, with a reusable core library, task-design tools, and HuggingFace model collection (377+ stars, Apache 2.0, 2026)
Sourced from
- Awesome AI for Science — github.com/gen-verse/jepa-anything
- GitHub — github.com/gen-verse/jepa-anything
Related resources
Scientific equation discovery and symbolic regression using LLMs, combining code generation with evolutionary search (ICLR 2025 Oral)
Fast, differentiable, JIT-free finite element library for PyTorch enabling GPU-native PDE solving with native autograd, tensorized assembly, and sparse linear algebra; part of the TensorGalerkin framework (218+ stars, Apache 2.0)
Parallel symbolic regression network evaluating millions of expressions on GPU with automated subtree reuse, Nature Computational Science cover article (MIT, 2026)
Advanced OCR with PP-StructureV3 document parsing, 13% accuracy improvement, supports 80+ languages
SOTA multimodal document parsing with 1.2B parameters outperforming GPT-4o, converts PDFs to LLM-ready Markdown/JSON
Kolmogorov-Arnold Networks with learnable activation functions on edges instead of fixed node activations, achieving strong performance in function fitting, PDE solving, and scientific discovery with enhanced interpretability as an alternative to MLPs (MIT, 16.3K+ stars, 2024)