CV

Arka Mukherjee — Undergraduate researcher in multimodal LLMs and VLM evaluation.

Contact Information

Name Arka Mukherjee
Professional Title Undergraduate Researcher
Email arka.mukherjee078@gmail.com

Professional Summary

Undergraduate researcher focused on multimodal LLM evaluation, reasoning benchmarks, and AI agents. Incoming Research Intern on the AMD AGI team working on LLM inference, Research Fellow at IIT Bhubaneswar (with Dr. Shreya Ghosh), and IUSSTF-Viterbi Summer Research Intern at USC. CS junior at KIIT University (GPA 9.74/10). Published at ICCV 2025, IJCNLP-AACL 2025, and EMNLP 2026 Findings.

Experience

  • -

    Bengaluru, India

    Incoming Research Intern
    AMD
    AMD AGI Team
    • Working on LLM inference.
  • 2026 - 2026

    Los Angeles, CA

    IUSSTF-Viterbi Summer Research Intern
    University of Southern California
    Advisor: Dr. Maja Mataric
    • Built LazySloth, a fast video understanding method that improved retrieval efficiency by up to 8.3x over existing pipelines such as VideoLucy and WorldMM, while retaining task accuracy. (Submitted to AAAI 2027)
    • Contributed to lab infrastructure on LLM inference.
  • 2026 - 2026

    Remote

    Research Intern
    Carnegie Mellon University
    Advisor: Dr. Min Xu
    • Applied GRPO, DPO, and PPO to study RL generalization to codon translation, a low-sampling decoding task (64-token action space).
    • Improved VRAM efficiency of the code to iterate GRPO on 800k+ datapoints.
  • 2024 -

    Bhubaneswar, India

    Undergraduate Research Fellow (Funded)
    IIT Bhubaneswar
    Advisor: Dr. Shreya Ghosh
    • Developed mmJEE-Eval, a 1,460-problem multimodal STEM reasoning benchmark evaluating 17 VLMs. Discovered metacognitive barriers: VLMs detect 53% of errors but correct only 3.5%. (IJCNLP-AACL 2025 Findings)
    • Created the first evaluation framework for VLM cultural competence via multimodal story generation on 5 VLMs. (Oral @ ICCV 2025 ASI Workshop)
    • Proposed ICE, a new evaluation paradigm unlike benchmarks and leaderboards that models real-world tasks.
    • Engineered CoTDL-mini-SWE-agent, a novel coding harness to improve frontier agentic terminal coding performance.
  • 2025 - 2025

    Ropar, India

    Summer Research Fellow (IASc-INSA-NASI)
    IIT Ropar
    Advisor: Dr. Sudarshan Iyengar
    • Developed EduVLM-Bench for STEM prerequisite detection and evaluated 5 open-source LLMs. Top model (Gemma3 27B) achieved 38.5% accuracy.

Education

  • 2023 - 2027

    Bengaluru, India

    B.Tech CSE
    Kalinga Institute of Industrial Technology (KIIT)
    Computer Science and Systems Engineering
    • Agentic AI, Probability & Statistics, Machine Learning, Algorithms, Data Mining, Human-Computer Interaction

Awards

  • 2026
    Singapore AI Safety Hub (SASH) FAST Fellow
    Singapore AI Safety Hub

    Selected for the fully funded Frontier AI Security Training fellowship with a 2.8% acceptance rate.

  • 2025
    KIIT Merit Scholar (Dean's List)
    KIIT University

    Dean’s List recognition for 6 semesters.

  • 2026
    Amgen Scholars Program
    Amgen Foundation

    Selected for the Amgen Scholars Summer Research Program (declined).

  • 2026
    Aalto Science Institute (AScI) Summer Research Fellow
    Aalto University, Finland

    Selected for a summer research position in Finland (declined).

  • 2025
    D&I Subsidy Award ($750)
    IJCNLP-AACL 2025

    Travel grant for presenting at IJCNLP-AACL 2025.

  • 2025
    NeurIPS 2025 DCVLR Challenge - #6/59
    NeurIPS 2025

    Team Blackwell ranked 6th out of 59 teams (top 10th percentile).

Skills

Programming Languages: Python (advanced), Java, C, SQL
ML / NLP Frameworks: PyTorch, HuggingFace Transformers, Datasets & TRL, Unsloth, LM Studio, Docker
Agentic AI: LangChain, Playwright, mini-SWE-agent

Interests

Research: Multimodal LLMs, VLM Evaluation, AI Reasoning, AI Agents, HCI
Outreach: Tech journalism (10M+ reads), hardware reviews with Nvidia / AMD / ASUS (Nvidia RTX 5090/5080/5070 sponsor), YouTube