I am Piyush Jha, a second-year Ph.D. student in Computer Science at Georgia Tech, specializing in Artificial Intelligence and Logic under the guidance of Prof. Vijay Ganesh. Drawing on my experience in Natural Language Processing and Computer Vision, my research focuses on efficiently combining Machine Learning with symbolic reasoning tools.
Recently, I developed Reinforcement Learning via Symbolic Feedback (RLSF), which integrates small language models with symbolic reasoning, outperforming GPT-4 on complex tasks using 1000x smaller models. I also created AlphaMapleSAT, a novel SAT solver utilizing symbolic feedback within MCTS, reducing computation time by 27x vs SOTA for hard combinatorial problems. Additionally, I designed an LLM jailbreak framework with a +50% success rate (vs SOTA at 2%) against highly safety-trained LLMs. My work spans diverse fields, including energy, e-commerce, law, cybersecurity, chemistry, and formal methods, delivering real-world AI innovations.
Areas I have worked on prior to pursuing my Ph.D.:
- Natural Language Processing: Hindi-Sanskrit translation with limited linguistic resources
- Representation Learning: Cross-modal Deep Learning algorithm (CorrMCNN)
- Trustworthy ML: Testing and verification tools for Neural Networks
- AI ∩ Formal Methods: SMT tactic selection
- AI ∩ Software Engineering: Reducing runtime errors in Java programs
- AI ∩ Cybersecurity: LLM in an RL loop to fuzz web applications
- AI ∩ Environmental Science: Predicting seasonal groundwater levels on a nationwide scale
- AI ∩ Law: Legal issues around Deepfakes