Projects

Selected Industry Projects

– Large-scale synthetic data generation for LLM training and evaluation (E.g., reasoning traces, agentic tool use benchmark)

– LLM post-training: SFT and RL (e.g., PPO, GRPO, DAPO, GSPO, CISPO and their variants) for coding and ads agents.

– Self-evolving agents for agent harness evolution and LLM post-training as automated AI researcher.

– Coding agents for large-scale code decomposition, migration, documentation, and evaluation.

Selected PhD Research Projects

• Large Language Models for Controllable Text Generation via Reinforcement Learning
  • Advior: Prof. Mustaque Ahamad, and Prof. Srijan Kumar
  • Research output: ACM WWW’23 code and data
    • Design novel BERT-based reward models in offline RL framework to capture desirable textual properties in text generation
    • Implemented a transformer decoder-based GPT-kind language model for counter-misinformation generation
    • Curate high-quality misinformation-related conversational datasets and fine-tuned LLaMA2 using LoRA
• Multi-task Learning-based Personalized Text Generation for Attack/Defense
  • Advior: Prof. Mustaque Ahamad, and Prof. Srijan Kumar
  • Research output: ACM KDD’21 code and data
    • Proposed a Generative Adversarial Network (GAN)-based text generation framework for personalization and adversarial attack
    • Design a multi-stage multi-task optimization strategy to improve the generation performance
    • Leverage adversarial machine learning for fine-tuning and enhancing the model robustness
• BERT-based Fine-grained Text and Textual Sequence Classification
  • Advior: Prof. Mustaque Ahamad, and Prof. Srijan Kumar
  • Research output: IEEE BigData’20 code and data and ACM ASONAM’21 code and data
    • Implement the masked language model-enhanced fine-tuning- and feature-based BERT for misinformation and hate speech classification
    • Propose the transformer-based sequence-aware deep sequence classifier powered by data augmentation from OpenAI ChatGPT and GPT4
    • Crawl large-scale tweets and user networks, and analyze the spread of misinformation and hate speech on Twitter