Vincent Counathe

I work on making LLM training and inference more efficient. To that end, I develop algorithms that improve model quality at low precision and run efficiently on the latest hardware.

More broadly, I study which limits of low-precision learning are fundamental and which can be overcome with better algorithms. Ultimately, the goal is better models for a given hardware and time budget: faster computation lets us train larger models, or the same model on more tokens.

I’m pursuing a PhD at Cornell, advised by Chris De Sa, and I work on quantization at Together AI.

Selected Research

Standard QAT and QUASAR loss surfaces: bringing reconstructed weights closer to latent weights improves the direction of the applied gradient.

QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction

Vincent Counathe, Ben Athiwaratkun, Christopher De Sa, Tianyi Zhang

QUASAR treats quantization-aware training as a weight reconstruction problem, continuously fitting a loss-aware quantized representation as the latent full-precision weights evolve. Its analysis connects weight reconstruction quality to the optimization dynamics and final quantized-model loss.

Paper Code Checkpoints Blog post (coming soon)

Comparison of causal, bidirectional, and leave-one-out token dependencies, showing training signal at every position for LOOT.

Empowering Masked Diffusion Models to Self-Correct with Leave-One-Out Transformers

Sofian Zalouk, Vincent Counathe, Paul Jünger, Daniel Cao, Daniel Zlotnick, Kilian Q Weinberger, Christopher De Sa

NeurIPS 2026

LOOT predicts each token from the rest of the sequence, allowing masked diffusion language models to learn from every token and self-correct during generation.

Paper

Earlier Research in Probability and Statistics

A random recursive tree generated by uniform attachment, with blue vertices and edges.

Statistical Risk Bounds for Genealogical Reconstruction on Random Recursive Trees

Vincent Counathe

Non-asymptotic risk bounds for a Jordan-ordering estimator in genealogical inference.

Paper

A graph with two communities, with denser connections within each community. — From Vershynin (2018), reproduced in the note.

Community Detection in Random Graphs

Vincent Counathe, Matthieu Dagès

Expository note on spectral methods for community detection in random graphs.

Note

Background

Before Cornell, I completed my Master's in Statistics and Machine Learning at Université Paris-Saclay (Institut de Mathématiques d'Orsay), where I worked on high-dimensional statistics and machine learning with Florent Krzakala, Lenka Zdeborová, and Christophe Giraud.