| Sep 24, 2026 | Our paper Understanding Layer Patching in Model Size Interpolation was accepted to NeurIPS 2026! |
| Sep 04, 2026 | I helped train IFM’s suite of fully open-source language models, coming in sizes from 0.9B to 375B! We also developed Uno, a lossless discrete diffusion framework which accelerates decoding by up to 2.5x compared to autoregressive sampling. We provide Uno adapters for Qwen3-8B and our own 0.9B and 7B models. Check out the paper: Unlocking Lossless Speedups in LLMs via Discrete Diffusion. |
| Aug 26, 2026 | Two new follow-up works to boomerang distillation out: Understanding Layer Patching in Model Size Interpolation, accepted to the AdaptFM workshop at ICML 2026, and Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs, accepted to EMNLP 2026! These papers explore extending boomerang distillation to post-trained models, and better understanding its layer patching mechanism. |
| May 11, 2026 | I’ll work as an AI Research Intern at the Institute of Foundation Models in Sunnyvale, CA for the summer! I’ll be advised by Subham Sahoo. |
| Apr 30, 2026 | Our paper Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Training was accepted to ICML 2026! |
| Feb 12, 2026 | New preprint out: Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Training! We propose a simple modification to the forward masking process of Masked Diffusion Models that speeds up training by up to 2.5x. |
| Jan 28, 2026 | Two papers accepted to ICLR 2026! Guided Speculative Inference for Efficient Test-Time Alignment of LLMs and Boomerang Distillation Enables Zero-Shot Model Size Interpolation. |
| Oct 08, 2025 | New preprint out: Boomerang Distillation Enables Zero-Shot Model Size Interpolation. We show that given a teacher and a single distilled student model, you can create models of intermediate sizes without any additional training! |
| May 08, 2025 | I was selected to receive a Kempner Institute Graduate Fellowship! |
| May 01, 2025 | Three papers accepted to ICML 2025! Universal Neural Optimal Transport (main conference), Entropy-Driven Pre-Tokenization for Byte-Pair Encoding (Tokenization Workshop), and Guided Speculative Inference for Efficient Test-Time Alignment of LLMs (Spotlight at ES-FoMo Workshop) |