Shubhashis Roy Dipta

PhD Researcher Ā· UMBC

sroydip1@umbc.edu


Amazon Science (Alexa)
Seattle, WA
Applied Scientist Intern
Summer 2026
Manager: Dr. Lichao Wang
Mentors: Dr. Xiaohu Xie, Dr. Daniel Bis
Amazon Science (Alexa)
Seattle, WA
Applied Scientist Intern
Summer 2025
Manager: Dr. Lichao Wang
Mentors: Dr. Daniel Bis, Dr. Kun Zhou
Paper: PA3: Policy-Aware Agent Alignment
Scale AI
San Francisco, CA
Machine Learning Research Intern
Summer 2024
Manager: Dr. Adrian Lam
Mentor: Vijay Kalmath
Blog: RLHF for Text-to-SQL
See more
University of Maryland, Baltimore County
Ph.D. in Computer Science
Fall 2023 - Present
Advisor: Dr. Frank Ferraro
Grade: 4.00/4.00
Publications: See Here (From 2022)
University of Maryland, Baltimore County
M.Sc. in Computer Science
Spring 2021 - Spring 2023
Awards: Phi Kappa Phi
Grade: 4.00/4.00
Morgan State University
Research Assistant
2017 - 2019
Advisor: Dr. Iman Dehzangi
Publications: 4 Journal
UniShopr.com
Bangladesh
Founder
2017 - 2021

Upcoming Travel

  • ACL 2026 in San Diego, CA (Jul 3-7)
Previous
  • āœ… NeurIPS 2025 in San Diego, CA (Dec 2-7)
  • āŒ AACL 2025 in Mumbai, India (Dec 20-24) (canceled)
šŸ‘‹ I'm open to meet! Email me to schedule a chat!

Peer Review

Reviewed 28+ papers across top venues (2023–2025).

Conferences
ACLNeurIPSNAACLCOLING*SEM
Workshops
SemEvalTrustNLPSRWW-NUTELVM
Journals
Scientific ReportsBMC BioinformaticsPlant MethodsComputational and Structural Biotechnology

I am a final-year Ph.D. researcher in Computer Science at the University of Maryland, Baltimore County (UMBC), advised by Dr. Frank Ferraro. I’ve also interned at Amazon Science (Alexa AI; Summer 2025 + Summer 2026) and Scale AI (Summer 2024). My research focuses on three areas where modern LLMs fail predictably - (1) complex reasoning, (2) tool-use, and (3) modality conflict - to make them more reliable, efficient, and aligned.

  • Reasoning & Decomposition
    • Semi-supervised RL for traceable decomposition-based claim verification [DecomposeRL]
    • Atomic, presupposition-free decomposition for robust claim verification [De-Presuppose]
    • Token-efficient math reasoning via distractor-aware computational graphs [DAGGER]
    • Curriculum-driven GRPO for math reasoning in under-resourced languages [GanitLLM]
    • Hierarchical event abstraction for compositional sequence modeling [SHEM]
  • Agentic LLMs & Reinforcement Learning
    • Tool-calling alignment via policy-grounded deliberation [PA3]
    • Multi-agent benchmarks for diagnosing collaboration failures [AgentCollabBench]
    • Mechanistic analysis of token saliency in on-policy distillation [Rock Tokens]
    • Metacognitive control in LLMs under resource constraints [TRIAGE]
  • Multimodal Learning & Evaluation
    • Reference-free factuality metric for video captions [VC-Inspector]
    • Calibrated abstention under modality conflict in omni-modal models [OMD]
    • Zero-shot multilingual text-to-video retrieval via temporal event decomposition [Q2E]

Graduating Spring 2027 Ā· No visa sponsorship needed Ā· actively seeking Research Scientist roles in NLP / Multimodal AI. Please reach out if you have an opening.

Recent News (See All)

Jun 1, 2026 šŸŽ‰ Joined Amazon Science (Alexa AI) for my second summer - researching self-distillation with RL to push LLM reasoning on agentic tasks.
May 27, 2026 šŸš€ New preprint - DecomposeRL: a 7B claim-verifier that matches GPT-4.1-mini across 11 benchmarks - with fully inspectable reasoning traces.
May 26, 2026 🄳 AgentCollabBench accepted at the FAGEN workshop @ ICML 2026 - 900 tasks that catch when a multi-agent LLM team’s final answer is right but the reasoning quietly broke.
May 20, 2026 🄳 5 of my papers got accepted at the MeLLM workshop @ ACL 2026 (Multilinguality in the Era of LLMs) - spanning text-to-gloss translation, math reasoning, sentiment auditing, VLM dialect benchmarks, and pretraining corpora.
May 14, 2026 🄳 OMD-Bench got accepted at 3 CVPR 2026 workshops (Any2Any MLLM, CVinWild, KnowledgeMR).
May 7, 2026 🄳 1 of my ACL papers (VC Inspector) also got accepted to MAGMaR 2026.

Beyond Research

I’ve competed internationally in algorithms and robotics - ranking 8th out of 300+ teams at the 2018 ACM ICPC Asia Dhaka Regional with multiple regional and national placements, reaching the top 70 on Kaggle šŸ„‰ in the Birdcall Identification competition, and placing 9th at the University Rover Challenge 2015 (Utah, USA) and 22nd at the European Rover Challenge 2016 (Poland). Full list of awards →

Before the PhD, I also founded UniShopr (2017-2021), a cross-border e-commerce platform serving consumers in Bangladesh.

Featured Publications

Check out Google Scholar for a full list of my publications.

  1. DecomposeRL: Learning to Ask Useful, Informative, and Diverse Questions for Semi-Supervised, Traceable Claim Verification
    Submitted
    DecomposeRL: Learning to Ask Useful, Informative, and Diverse Questions for Semi-Supervised, Traceable Claim Verification
    Shubhashis Roy Dipta,Ā Ankur Padia,Ā andĀ Francis Ferraro
    Preprint 2026
  2. Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation
    Submitted
    Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation
    Yuxuan Jiang*,Ā Runchao Li*,Ā Shubhashis Roy Dipta*, and 2 more authors
    Preprint 2026
    * Equal contribution
  3. AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators
    Submitted
    AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators
    Aritra Mazumder,Ā Shubhashis Roy Dipta,Ā Nusrat Jahan Lia, and 10 more authors
    Preprint 2026
  4. TRIAGE: Evaluating Prospective Metacognitive Control in LLMs under Resource Constraints
    Preprint
    TRIAGE: Evaluating Prospective Metacognitive Control in LLMs under Resource Constraints
    Zabir Al Nazi,Ā andĀ Shubhashis Roy Dipta
    Preprint 2026
  5. PA3: Policy-Aware Agent Alignment through Chain-of-Thought
    Submitted
    PA3: Policy-Aware Agent Alignment through Chain-of-Thought
    Shubhashis Roy Dipta,Ā Daniel Bis,Ā Kun Zhou, and 4 more authors
    Preprint 2026
    Work done during internship at Amazon Alexa AI
  6. †DAGGER: Distractor-Aware Graph Generation for Executable Reasoning in Math Problems
    Submitted
    †DAGGER: Distractor-Aware Graph Generation for Executable Reasoning in Math Problems
    Zabir Al Nazi,Ā Shubhashis Roy Dipta,Ā andĀ Sudipta Kar
    Preprint 2026
  7. Omni-Modal Dissonance Benchmark: Systematically Breaking Modality Consensus to Probe Robustness and Calibrated Abstention
    Submitted
    Omni-Modal Dissonance Benchmark: Systematically Breaking Modality Consensus to Probe Robustness and Calibrated Abstention
    Zabir Al Nazi*,Ā Shubhashis Roy Dipta*,Ā andĀ Md Rizwan Parvez
    Preprint 2026
    * Equal contribution
  8. GanitLLM: Difficulty-Aware Bengali Mathematical Reasoning through Curriculum-GRPO
    ACL
    GanitLLM: Difficulty-Aware Bengali Mathematical Reasoning through Curriculum-GRPO
    Shubhashis Roy Dipta,Ā Khairul Mahbub,Ā andĀ Nadia Najjar
    ACL 2026
  9. VC-Inspector: Advancing Reference-free Evaluation of Video Captions with Factual Analysis
    ACL
    VC-Inspector: Advancing Reference-free Evaluation of Video Captions with Factual Analysis
    Shubhashis Roy Dipta,Ā Tz-Ying Wu,Ā andĀ Subarna Tripathi
    ACL 2026
  10. Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks
    ACL
    Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks
    Nobin Sarwar,Ā Shubhashis Roy Dipta,Ā Zheyuan Liu, and 1 more author
    ACL 2026
  11. Q2E: Query-to-Event Decomposition for Zero-Shot Multilingual Text-to-Video Retrieval
    AACL
    Q2E: Query-to-Event Decomposition for Zero-Shot Multilingual Text-to-Video Retrieval
    Shubhashis Roy Dipta,Ā andĀ Francis Ferraro
    AACL 2025
  12. If We May De-Presuppose: Robustly Verifying Claims through Presupposition-Free Question Decomposition
    *SEM
    If We May De-Presuppose: Robustly Verifying Claims through Presupposition-Free Question Decomposition
    Shubhashis Roy Dipta,Ā andĀ Francis Ferraro
    *SEM 2025
  13. Learning How to Use Tools, Not Just When: Pattern-Aware Tool-Integrated Reasoning
    MathAI @NeurIPS
    Learning How to Use Tools, Not Just When: Pattern-Aware Tool-Integrated Reasoning
    Ningning Xu,Ā Yuxuan Jiang,Ā andĀ Shubhashis Roy Dipta
    MathAI @NeurIPS 2025
  14. Semantically-informed Hierarchical Event Modeling
    *SEM
    Semantically-informed Hierarchical Event Modeling
    Shubhashis Roy Dipta,Ā Mehdi Rezaee,Ā andĀ Francis Ferraro
    *SEM 2023