Victoria Graf

Timeline

1. May 2023

   A Goldwater Scholarship

   As a junior computer science major at Princeton University, she wins a Goldwater Scholarship in May 2023, one of four awarded to Princeton students. She plans a PhD in machine learning focused on natural language processing and names Karthik Narasimhan and Danqi Chen among her mentors.

2. December 2023

   The C-STS similarity task

   In December 2023, she is one of nine authors of C-STS at EMNLP in Singapore, alongside Danqi Chen and Karthik Narasimhan. The paper defines conditional semantic textual similarity, which measures sentence similarity under a natural-language condition, and contains almost 20,000 instances.

3. June 2024

   Nested PoE backdoor defense

   In June 2024, she coauthors “Two Heads are Better than One: Nested PoE for Robust Defense Against Multi-Backdoors” with Qin Liu and Muhao Chen for NAACL in Mexico City. Its Nested Product of Experts framework uses a trigger-only mixture of experts within a Product of Experts defense to handle multiple backdoor trigger types.

4. November 2024

   Tulu 3 post-training

   In November 2024, she joins Nathan Lambert, Hannaneh Hajishirzi and other authors on Tulu 3, a family of fully open post-trained models built on Llama 3.1 base models. The paper includes the model data, code and training recipes, and introduces Reinforcement Learning with Verifiable Rewards.

5. July 2025

   The IFBench benchmark

   In July 2025, she coauthors “Generalizing Verifiable Instruction Following” with Valentina Pyatkin, Saumya Malik and Hamish Ivison. The paper introduces IFBench with 58 new verifiable out-of-domain constraints and appears in the NeurIPS 2025 Datasets and Benchmarks Track.

6. December 2025

   Core contribution to Olmo 3

   She is a core contributor to Olmo 3, a family of fully open language models at the 7B and 32B parameter scales. The Olmo 3 paper is submitted in December 2025 and describes the full model flow, including every stage, checkpoint, data point and dependency used to build it.

7. July 2026

   Pretraining-data poisoning

   In July 2026, she coauthors “Pretraining Data Can Be Poisoned through Computational Propaganda” with Hanna Hajishirzi, Noah A. Smith, David Kohlbrenner and Kyle Lo. The paper studies public discussion interfaces as a route for poisoning web-scale pretraining data and introduces HalfLife to estimate whether injected content enters training data after crawling and curation.

8. As of October 2026

   PhD student at Washington

   As of October 2026, she is a PhD student at the University of Washington, advised by Hanna Hajishirzi, and a Student Researcher at Ai2. Her interests include language-model alignment and improving language-model usability.