repositories
Selected open-source projects I have contributed to.
RELAY: Propagating continuous context across discrete diffusion steps via learned latent representations. Code for Learned Relay Representations for Forward-Thinking Discrete Diffusion Models (SPIGM & FoGen Workshops @ ICML 2026) — full-parameter adaptation of Fast-dLLM v2 1.5B, plus a Sudoku design study.
A modular, research-friendly framework for non-autoregressive language models. Built on PyTorch Lightning with Hydra configs. Supports MLM, ILM, MDLM, FlexMDM, and more. Co-developed at IESL.
A continuous-time Markov chain framework for insertion language models. Code for A Continuous Time Markov Chain Framework for Insertion Language Models (AISTATS 2026 Spotlight).
Insertion-based sequence generation with learnable order dynamics — variable-length masked diffusion that learns where to insert as it generates. Code for Insertion Based Sequence Generation with Learnable Order Dynamics (ICML 2026).
Interpretable information extraction over sentence- and document-level parse structures, framing inference as subgraph matching and learning as subgraph mining. Code for Claim Extraction via Subgraph Matching (DMR 2023).
A finite-state morphological analyzer for Blackfoot (Siksiká) nouns, implemented with the Helsinki Finite-State Toolkit. Based on Frantz's Blackfoot Grammar and Blackfoot Dictionary.
🤗 On Hugging Face
Model checkpoints and datasets released with the Relay paper, on my Hugging Face profile.
Relay adaptation of Fast-dLLM v2 1.5B with stop-gradient training.
Relay adaptation of Fast-dLLM v2 1.5B trained with full BPTT.
Sudoku planning benchmark used in the Relay design study (4.25M rows).