Reproduce “Toward Compiler World Models: Learning Latent Dynamics for Efficient Tensor Program Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haolin Pan, Lianghong Huang, Xvlin Zhou, Mingjie Xing, Yanjun Wu
Researchers from the Chinese Academy of Sciences developed a framework that models tensor program optimization as action-conditioned latent dynamics, moving beyond static code evaluation. The approach simulates step-by-step state transitions in a continuous latent space, leading to a 1.29x geometric-mean speedup in end-to-end latency with a 64-trial budget and matching Ansor-10240 performance with 10x fewer measurements.
Key insights
- Treating tensor program schedule evaluation as action-conditioned latent dynamics allows for capturing contextual dependencies and the sequential nature of compiler transformations, addressing limitations of static evaluation.
- Modeling program state evolution in a continuous latent space enables efficient prediction of terminal states without the overhead of explicit AST mutations or text encoding of intermediate states.
- Combining a contrastively trained state encoder with a dynamic transition model and a ranking-based objective provides a more robust and semantically aware evaluation for tensor programs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.