I am a Computer Science Ph.D. candidate at the University of Maryland, College Park, advised by Prof. Sanghamitra Dutta. My research focuses on LLM/VLM post-training and reasoning, with an emphasis on improving their efficiency and effectiveness from both data and algorithmic perspectives, as well as on interpretability. Previously, I worked on code LLMs, AI security, and trustworthy ML.
Geometry-Based RLVR Credit Assignment from Hidden-State Trajectories
Yanjun Fu, Shunchi Zhang, Zhongtao Jiang, Kun Xu, Sanghamitra Dutta
Preprint, 2026
Preprint coming soon.
TL;DR
Uses hidden-state geometry for step-level credit assignment in RLVR. On AIME 2024–2026, our method outperforms DAPO by an average of 2.38 points and surpasses DAPO's peak performance with 1.31× fewer GPU hours, placing it on the performance–compute Pareto frontier among the evaluated RLVR methods.
Proposes an agentic data engine that synthesizes SFT data for VLM tool use. Using SFT alone, LogT-SFT surpasses the SOTA model trained with both SFT and RL by 1.9 points, and LogT-RL surpasses the SOTA by 9.6 points.
Keywords: Agentic data synthesis for VLM agents
2025
T-Shirt: Token-Selective Hierarchical Data Selection for Instruction Tuning
Yanjun Fu, Faisal Hamman, Sanghamitra Dutta
Conference on Neural Information Processing Systems (NeurIPS), 2025
Selects SFT examples by focusing on informative tokens and checking quality consistency among similar samples. Models trained on just 5% of the data can outperform those trained on the full dataset. The method is also highly scalable: it processes 52k samples in just 40 minutes on a single GPU.
Keywords: Data selection for SFT
Few-Shot Knowledge Distillation of LLMs With Counterfactual Explanations
Uses minimally changed examples that flip a teacher model's prediction to help a smaller model learn its decision boundary. This improves knowledge distillation with very little training data.
Introduces PrimeVul, a vulnerability-detection dataset with much higher quality and realistic evaluation settings, and shows that existing benchmarks substantially overestimate code models' ability to detect vulnerabilities.
Introduces CodeGuard+ to evaluate whether generated code is both secure and correct, and uses constrained decoding to improve security without a specialized training dataset.
