Shubhashis Roy Dipta

Shubhashis is a final-year Ph.D. researcher in Computer Science at the University of Maryland, Baltimore County (UMBC), advised by Dr. Frank Ferraro. His work spans Natural Language Processing, Computer Vision, and multimodal reasoning, with a focus on decomposition-based reasoning — making complex problems more reliable by structuring them into atomic, verifiable subproblems.

His research thesis is that complex problems become more reliable when broken into structured, atomic, verifiable subproblems — a principle he has carried across reasoning regimes, modalities, and languages. Recent work includes presupposition-free question decomposition for robust claim verification [1], query-to-event decomposition for zero-shot multilingual text-to-video retrieval [2], and curriculum-driven GRPO for low-resource Bengali mathematical reasoning [3]. His earlier work on Hierarchical Variational Autoencoders for Event Representation Learning [4] underpins downstream applications in summarization, question answering, and counterfactual reasoning.

A recurring theme across his projects is robustness and efficiency over raw accuracy. He builds lightweight, fact-aware reference-free evaluators for video captions [5], designs benchmarks that isolate modality reliance and probe calibrated abstention in omni-modal systems [6], reformulates mathematical reasoning as distractor-aware computational graphs that match reasoning-specialized models at 89% lower token cost [7], and trains tool-calling agents to recall relevant business policies on demand — outperforming in-context baselines by 16 points while using 40% fewer tokens [8].

In Summer 2026, he will return to Amazon Science (Alexa Team) as an Applied Scientist Intern.

In Summer 2025, he interned at Amazon Science (Alexa Team) as an Applied Scientist. He showed how the Deliberative Alignment technique can be used to reduce token usage in tool-calling LLMs and lower inference cost (Paper).

In Summer 2024, he interned at Scale AI (Enterprise Team) as a Machine Learning Researcher. He explored how RL can improve text2SQL generation by using a small LLM-as-a-Judge inside the reinforcement loop (Scale AI Blog).

Beyond Research

Shubhashis founded UniShopr (2017–2021), a cross-border e-commerce platform serving consumers in Bangladesh.

He has also competed internationally in robotics and algorithms — placing 9th at the University Rover Challenge 2015 (Utah, USA) and 22nd at the European Rover Challenge 2016 (Poland), ranking 8th out of 300+ teams at the 2018 ACM ICPC Asia Dhaka Regional with multiple regional and national placements, and reaching the top 70 on Kaggle 🥉 in the Cornell Birdcall Identification competition. See more.