Yavuz Faruk Bakman

PhD Student in Computer Science Capital One Responsible AI Fellow

University of Southern California

Biography

Welcome to my website! I’m a 4th year PhD student in Computer Science, co-advised by Salman Avestimehr and Sai Praneeth Karimireddy. Before starting my PhD, I worked at Hyperbee.ai as Research Engineer, focusing on making neural networks more efficient through compression and quantization. I got my Bachelor’s degree in Computer Science from Bilkent University, where I also researched machine learning security, specifically Trojan Attacks, with Professor Tudor Dumitras.

Currently, I’m interested in Reliable and Safe LLMs. I’m exploring uncertainty quantification of LLMs for reliable decision making and adversarial robustness of LLMs in continual learning scenarios. Furthermore, I have been doing research on LLM Efficiency, Reasoning, Continual Learning, Self-supervised Contrastive Learning and Federated Learning.

Outside of my research, I love playing video games, especially the Soulsborne series:

“Facing a challenge? Keep Calm and Git Gud.”

Interests

Reliable LLMs
LLM Interpretability
LLM Agents
LLM Alignment
Continual Learning

Education

PhD in Computer Science, Present
University of Southern California, CA, US
MSc in Computer Science, 2025
University of Southern California, CA, US
BSc in Computer Science, 2022
Bilkent University, Turkey

Recent News

02-19-2026: I am joining Google Cloud as a Summer Intern in Seattle to work on agents!
02-19-2026: Our new paper “Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignment” is on arXiv!
02-19-2026: “Uncertainty as Feature Gaps: Epistemic Uncertainty Quantification of LLMs in Contextual Question-Answering” got accepted to ICLR 2026! See you in Brazil!
09-20-2025: Our work “Reject Only Critical Tokens: Pivot-Aware Speculative Decoding” got accepted to Neurips Efficient Reasoning Workshop!
09-15-2025: Our open-source library “TruthTorchLM” got accepted to EMNLP 2025!
06-15-2025: Our Paper “Reconsidering Reconsidering LLM Uncertainty Estimation Methods in the Wild” got accepted to ACL 2025!
06-15-2025: Our Paper “Un-considering Contextual Information: Assessing LLMs’ Understanding of Indexical Elements” got accepted to ACL 2025, Findings!
05-02-2025: Our Library for assessing Truthfulness of LLMs, TruthTorchLM, officially released!
02-02-2025: LARS got accepted to NAACL 2025, Findings!
22-10-2024: I gave a talk at Amazon-USC Center on Secure and Trusted Machine Learning about Trustworthy and Efficient LLMs!
20-10-2024: I have been selected as Capital One Responsible AI Fellow 2024 !
01-07-2024: Our paper “CroMo-Mixup: Augmenting Cross-Model Representations for Continual Self-Supervised Learning” got accepted to ECCV 2024 ! See you in Milano!
28-06-2024: “Do LLMs Recognize me, When I is not me: Assessment of LLMs Understanding of Turkish Indexical Pronouns in Indexical Shift Contexts” got accepted to ACL Turkic Languages 2024 Workshop .
16-06-2024: New paper “Do Not Design, Learn: A Trainable Scoring Function for Uncertainty Estimation in Generative LLMs” in collaboration with Amazon AI posted to Arxiv!
16-05-2024: Our paper, “MARS: Meaning-Aware Response Scoring for Uncertainty Estimation in Generative LLMs” co-authored with Amazon AI, has been accepted to the ACL 2024.
01-06-2024: Thrilled to announce that my first-author paper, “Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting in Continual Federated Learning,” has been accepted at ICLR 2024 . Looking forward to presenting our findings in Vienna!
10-12-2023: I gave a talk at Amazon-USC Center on Secure and Trusted Machine Learning about Advancing Continual & Federated Learning with Self & Mixed Supervision.
08-24-2022: Began my journey towards a PhD in Computer Science at University of Southern California(USC).
06-15-2022: Graduated with the highest honors from Bilkent University, majoring in Computer Science.
05-15-2022: Made the decision to join USC for my PhD studies under the guidance of Salman Avestimehr.
04-14-2022: Honored to have received PhD offers from several prestigious institutions: Princeton , Cornell, USC, UCSB, UCSD, Wisconsin-Madison, and Northeastern.

Experience

Applied Research Intern PhD

Capital One

June 2025 – August 2025

Worked on uncertainty quantification of language models in the AI Foundations team, developed a novel epistemic uncertainty quantification method for contextual question answering, and submitted the resulting paper to ICLR 2026.

Research Engineer

Hyperbee.AI

June 2019 – September 2022

Worked on accelerating and compressing neural networks for various computer vision tasks including. Mostly focused on the quantization aspect.

Research Intern

University of Maryland, College Park

June 2021 – December 2021

Attended TrojAI competetion in UMD - UC Berkeley Team supervised by Tudor Dumitras. Developed novel methods using loss surface, layer similarity and trojan transferability to detect backdoored models.

Recent Publications

Quickly discover relevant content by filtering publications.

Yavuz Faruk Bakman, Duygu Nur Yaldiz, Salman Avestimehr, Sai Praneeth Karimireddy (2026). Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignment. arXiv.

PDF Source Document

Amir Ziashahabi, Yavuz Faruk Bakman, Duygu Nur Yaldiz, Mostafa El-Khamy, Sai Praneeth Karimireddy, Salman Avestimehr (2025). Reject Only Critical Tokens: Pivot-Aware Speculative Decoding. NeurIPS 2025 Workshop.

PDF Code Source Document

Duygu Nur Yaldiz, Yavuz Faruk Bakman, Sungmin Kang, Alperen Ozis, Hayrettin Eren Yildiz, Mitash Ashish Shah, Zhiqi Huang, Anoop Kumar, Alfy Samuel, Daben Liu, Sai Praneeth Karimireddy, Salman Avestimehr (2025). TruthTorchLM: A Comprehensive Library for Predicting Truthfulness in LLM Outputs. EMNLP 2025.

PDF Code Source Document DOI

Yavuz Faruk Bakman, Sungmin Kang, Zhiqi Huang, Duygu Nur Yaldiz, Catarina Belem, Chenyang Zhu, Anoop Kumar, Alfy Samuel, Daben Liu, Sai Praneeth Karimireddy, Salman Avestimehr (2025). Uncertainty as Feature Gaps: Epistemic Uncertainty Quantification of LLMs in Contextual Question-Answering. ICLR 2026.