Advaith Malladi

Ph.D. in Computer Science @ UNC Chapel Hill  |  ML Research @ UNC Chapel Hill, MPI-SWS, IIIT Hyderabad
advaithmalladi0620@gmail.com

I am a Ph.D. student in Computer Science at UNC Chapel Hill, advised by Prof. Shashank Srivastava. My research spans Machine Learning and Natural Language Processing, with a focus on the interpretability and explainability of ML models, post-training of LLMs, and understanding and improving the reasoning capabilities of LLMs. My work has been published at ICML 2026 and EMNLP 2025. Prior to this, I was a Research Fellow at the Max Planck Institute for Software Systems, where I worked on understanding and training LLMs to balance conflicting behaviors, such as creativity and factuality, under the supervision of Prof. Abhilasha Ravichander. I graduated from IIIT Hyderabad in 2026 with a B.Tech. in Computer Science and an M.S. by Research in Computational Linguistics.

Education

University of North Carolina at Chapel Hill

Ph.D. in Computer Science

Advisor: Prof. Shashank Srivastava

August 2026 -

International Institute of Information Technology - Hyderabad

B.Tech in Computer Science (Hons) + M.S. by Research in Computational Linguistics
Advised by Dr. Radhika Mamidi in the integrated bachelor's and master's by research program
2021 - 2026

Publications

Directly Optimizing Natural Language Explanations for Behavioral Faithfulness: Simulatability and Recoverability

Advaith Malladi, Shashank Srivastava

International Conference on Machine Learning 2026 (ICML 2026)

[Paper]

Explaining Differences Between Model Pairs in Natural Language through Sample Learning

Advaith Malladi, Rakesh R. Menon, Yuvraj Jain, Shashank Srivastava

Empirical Methods in Natural Language Processing 2025 (EMNLP 2025)

[Paper]

Research Interests

My research spans Machine Learning and Natural Language Processing. I am especially interested in:

  • Interpretability and Explainability of ML Models. Understanding the internal mechanisms of ML models and developing methods to make their decision-making processes transparent and interpretable to humans.
  • Generating Faithful Natural Language Explanations for Model Behavior. Training language models to generate explanations that are directly optimized for faithfulness and simulatability, enabling humans to better understand and predict model behavior.
  • Faithfulness of Chain-of-Thought in Reasoning Models. Studying whether the chain-of-thought produced by thinking-based models actually reflects the computation behind their final answers, characterizing cases where sound reasoning yields wrong answers or flawed reasoning yields right ones, and developing methods to measure and improve CoT faithfulness.
  • Post-Training of LLMs using RLHF Techniques. Exploring reinforcement learning from human feedback techniques such as GRPO to align LLMs with desired behaviors after pretraining.
  • Conflicting Behaviors in LLMs. Understanding how contradicting properties such as creativity and factuality interact within LLMs, and training models to dynamically alternate between such behaviors at a token level.
  • Mixture-of-Experts (MoE) Language Models. Exploring the architecture and training of MoE models to enable specialized expert routing, particularly for handling conflicting behavioral objectives.

Research Experience

Research Assistant

Learning from Language Lab, UNC Chapel Hill

Under the guidance of Prof. Shashank Srivastava, my research focuses on the interpretability and explainability of machine learning models, post-training of large language models, and understanding and improving the reasoning capabilities of LLMs. My work includes generating faithful natural language explanations of differences in model behavior (EMNLP 2025) and optimizing LLM-generated explanations using reinforcement learning techniques such as GRPO (ICML 2026). I also study the reasoning processes of language models and explore methods to improve their reasoning capabilities.

August 2026 - Present

Research Fellow

Max Planck Institute for Software Systems

Under the guidance of Prof. Abhilasha Ravichander, I investigated how competing behaviors, such as creativity and factuality, interact in large language models. I also explored training Mixture-of-Experts language models to dynamically balance these behaviors at the token level.

March 2026 - August 2026