Anton Xue

Postdoctoral Fellow, UT Austin

I work on diffusion language models, AI agents for mathematical research, and provable AI safety and interpretability. I'm currently a postdoc at the Institute for Foundations of Machine Learning (IFML) at UT Austin, working with Sanjay Shakkottai, Adam Klivans, and Swarat Chaudhuri.

I completed my Ph.D. in Computer and Information Science at the University of Pennsylvania with Rajeev Alur and Eric Wong.

CV Email Google Scholar GitHub Twitter

Selected works

Efficient diffusion language models

Reducing the cost of AR-to-DLM pretraining; faster parallel decoding and better code generation.

Try dQwen3.5 →

start

step 1ingapsstep

step 2fillingapseachstep

step 3fillinmanygapsateachstep

PIG: Parallel Inconsistency Guidance for Accelerating Diffusion Language ModelsarXiv 2026
Multi-Target Few-Shot Prompting with Diffusion Language ModelsarXiv 2026
dQwen3.5: Hybrid-Attention Diffusion Language ModelsarXiv 2026
AnCoder: Anchored Code Generation via Discrete Diffusion ModelsarXiv 2026

AI agents for mathematical research

Long-horizon AI agents for mathematical research; new bounds on the Grothendieck constant and AI-guided theorem proving.

Talk slides →

Provable AI safety and interpretability

Provable guarantees for LLM reasoning and interpretability: rule following, reasoning soundness, and explanation stability.

A fraction's denominator is 3× its numerator minus 7. If it equals 2/5, find the numerator.

step 1Let the numerator be x.✓

step 2The denominator is 3x − 7.✓

step 3So x/(3x − 7) = 3/5.✗

step 4Then 5x = 9x − 20, so x = 5.✗

Earlier work spans formal verification, programming languages, convex optimization, and control, including symbolic execution for Haskell and certified compilation for WebAssembly. All publications →