Saumya Malik

Saumya Malik is an AI researcher whose work spans training data, language-model post-training, instruction following and evaluation. She earned a computer-science degree at Princeton in 2024, worked in Danqi Chen's lab, and coauthored QuRating and a bachelor's thesis on LSAT logic games. During her 2025–2026 PhD application cycle, she was a predoctoral young investigator on Ai2's AllenNLP and Olmo teams. Her work includes Tulu 3 in 2024–2025, IFBench and Olmo 3 in 2025, and RewardBench 2 at ICLR 2026. As of October 2026, she is a PhD student at the University of Washington.

Works: 8
Citations: 29
h-index: 4

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2023–2024

Princeton and Danqi Chen's lab

At Princeton, she works in Danqi Chen's lab as an undergraduate in 2023–2024. She earns an undergraduate degree in 2024, majoring in computer science and minoring in linguistics and cognitive science.
2024

QuRating data selection

She coauthors QuRating in 2024, a method for scoring pre-training data on writing style, expertise, facts and educational value. Models trained on selected data improve perplexity and in-context learning.
2024

LSAT logic-game reasoning

In 2024, she publishes Lost in the Logic, her bachelor's thesis on how language models handle the LSAT's Logic Games. It builds a dataset and reports 70 percent accuracy for GPT-4 on a smaller subset using a prompt approach adapted from Reflexion.
2024–2025

Tulu 3 post-training

She coauthors Tulu 3 in 2024, a family of fully open post-trained language models built on Llama 3.1. The release includes data, code and recipes for supervised fine-tuning, Direct Preference Optimization and reinforcement learning with verifiable rewards.
2025

Generalizing instruction following

In 2025, she coauthors Generalizing Verifiable Instruction Following, which introduces IFBench to test whether language models generalize to 58 new output constraints. The paper reports that reinforcement learning with verifiable rewards improves precise instruction following.
2025–2026

Olmo 3 at Ai2

During the 2025–2026 PhD application cycle, she works as a predoctoral young investigator on Ai2's AllenNLP and Olmo teams, advised by Nathan Lambert. In 2025, she coauthors Olmo 3, a fully open family of language models at 7B and 32B parameter scales.
2025–2026

RewardBench 2

At ICLR 2026, she coauthors RewardBench 2, a multi-skill benchmark for reward-model evaluation. Models score about 20 points lower on it than on the first RewardBench, and its scores correlate with downstream performance.
As of October 2026

A UW doctoral student

As of October 2026, she is a PhD student at the University of Washington.