Himanshu Gupta

I am an Applied Scientist at Amazon (Stores Foundation AI team), where I work on mid-training, post-training optimization, synthetic data creation and benchmarking of large foundation models. More recently, I have focused on RL-based post-training (GRPO/GSPO) with LLM-as-a-Judge reward design for product question answering, large-scale data curation and quality filtering for mid-training (including distilled code-quality classifiers), behavioral mid-training for customer personalization, and LLM-based product recommendation. I have hands on experience with Megatron-LM, verl (SFT, DPO, GRPO), vLLM , SGLang and TRL .

Email  /  Google Scholar  /  X  /  Linkedin

My research spans Instruction Tuning ( Sample Efficiency, Long-Sequence Medical Tasks, Event Detection ), Efficient Pretraining, Efficient LLMs-as-a-Judge, Mathematical Benchmarking ( Adversarial Math Word Problems, Multimodal Benchmarking, Humanity's Last Exam and Formal Math Discovery ), Agentic Benchmarking ( Terminal-Bench and Terminal-Bench-Science ) and Synthetic Data Generation. My recent work centers on foundation models with strong instruction following, guideline adherence, and high-precision reasoning — built through better mid-training corpora and RL-based post-training alignment.