Hi, I am Saikat, a Senior Researcher at the Research in Software Engineering (RiSE) group at Microsoft Research, working on reliability of large language models for code and post-training. I bring 10 years of experience in training and evaluating code models, with a focus on improving the correctness and fidelity of generated programs under real-world constraints.

My work guides code generation models through static and dynamic correctness signals—via tests, program analysis, and verification—and uses these signals through fine-tuning and reinforcement learning. I view reliability as fundamentally a training problem, driven by structured and composable feedback.

Earlier, I graduated with a Ph.D. in Computer Science from Columbia University, advised by Professor Baishakhi Ray. I wrote my Ph.D. thesis on Learning to Edit Code.

🌐 Website: saikatc.info

• Post-training for code: SFT, RLHF/GRPO, reward modeling, reranking, retrieval-augmented fine-tuning
• Correctness supervision: Reward design using test generation, execution feedback, mutation testing, specification inference, and program analysis
• Agent-driven testing: DeepTest — symbolic analysis + LLMs for testing production code at scale
• Formal verification: DeepProof — post-training models for theorem proving (F*, Rocq, Lean)
• Systems: PyTorch, Megatron-LM, Ray, distributed GPU clusters, Kubernetes

• 🏆 ICSE'25 — Neural Synthesis for Proof-Oriented Programming [Distinguished Paper Award]
• 🏆 ISSTA'23 — Contrastive Learning for Code Understanding [Distinguished Paper Award]
• 📄 ACL'25 — Teaching an Old LLM Secure Coding via Localized Preference Optimization
• 📄 EMNLP'23 — Ranking LLM-Generated Loop Invariants
• 📄 ICSE'24 — Causal Learning for Code Understanding
• 📄 FSE'22 — NatGen: Semantic Rewriting for Pretraining of Code Models
• 📄 NAACL'21 — Unified Pretraining for Code Understanding and Generation

• Use of LLMs for program synthesis, editing, and verification
• Reinforcement learning with execution and correctness feedback for code
• Formal methods meets machine learning (proof generation, specification mining)