Ved Chadderwala
Graduate Student · Thomas Lord Department of Computer Science · University of Southern California
I am a first-year MSCS student at USC working on LLM reliability, focusing on how models decide when to trust internal knowledge versus retrieve or self-correct. My work studies inference-time reasoning and post-training alignment using reinforcement learning.
Research Focus
LLM Reliability · Inference-time Reasoning · RL-based Alignment
I study reliable language models, with an emphasis on inference-time decision making, uncertainty estimation, and alignment. My work explores how prompting, retrieval, and reinforcement learning influence model behavior, aiming to improve robustness and controllability in real-world systems.
Projects
When Should a Language Model Trust Itself? Routing and Reflection for Reliable Generation
ActAgen: A Memory- and Compute-Efficient Runtime Framework for RL Systems
Environment-Aware Hierarchical Reinforcement Learning for Minecraft
Intrusion Detection in IoT Networks Using Machine Learning
Identifying Gender from Social Media Posts and Comments