Riju Marwah

Trustworthy & Reliable LLMs · Token-level Dynamics · Interpretability · Alignment

Riju Marwah headshot
New Delhi · -:-
marwah.riju@gmail.com · CV · LinkedIn · Google Scholar · GitHub

I am a Research Intern at the IRT Group at the University of South Carolina, advised by Dr. Amit Sheth and Vishal Pallagani. I will complete my Bachelor's in Computer Science in June 2026.

My research centers on making large language models trustworthy, robust, and reliable through a focus on token-level dynamics. I study how technical signals and linguistic cues interact during long-context generation, and design retrain-free interventions to mitigate degradation. My work sits at the intersection of robustness, interpretability, and alignment.

I am currently working on formalizing cognitive fatigue as a measurable latent state in autoregressive transformers — see the project page →

I am always open to discussing research, collaborations, or anything else — feel free to reach out!

News

Apr 2026
🎉 Paper accepted at ICML 2026 — Cognitive Fatigue in Autoregressive Transformers: Formalization and Measurement.

Jan 2026
💻 Presented Chatsparent at AAAI 2026 in Singapore; served as a Student Scholar Volunteer.

Oct 2025
🎉 Paper accepted at the AAAI 2026 Demonstration Track — Chatsparent: An Interactive System for Detecting and Mitigating Cognitive Fatigue in LLMs (28% acceptance rate).

Publications

  1. Cognitive Fatigue in Autoregressive Transformers: Formalization and Measurement ICML 2026
    Riju Marwah*, Ritvik Garimella, Vishal Pallagani, Atishay Jain, Michael Stewart, Amit Sheth

    Formalizes cognitive fatigue as a runtime state variable grounded in three token-level signals. Introduces the Fatigue Index with five explicit axioms and validates it across nine models (1B–13B) on HotpotQA, TriviaQA, and SQuAD under long-context, positional, and precision stress conditions. [Project Page]

  2. Chatsparent: An Interactive System for Detecting and Mitigating Cognitive Fatigue in LLMs AAAI 2026 Demo
    Riju Marwah*, Vishal Pallagani, Ritvik Garimella, Amit Sheth

    Introduces Chatsparent, an interactive system that detects and mitigates cognitive fatigue in LLMs via token-level signals and retrain-free interventions, improving reliability and transparency during dialogue. [OpenReview]

  3. MicroDetect-Net (MDN): Leveraging Deep Learning to Detect Microplastics in Clam Blood Springer / ICICC 2025
    8th International Conference on Innovative Computing and Communication, Springer Nature

    Proposes MDN, combining fluorescence microscopy (Nile Red staining) and deep learning to scan blood samples for microplastics — a step toward human blood analysis. [Preprint]

Experience

  1. Research Intern, AI Institute, University of South Carolina · Apr 2025 – Present
    • Advisors: Dr. Amit Sheth (NCR Chair & Director, AIISC), Vishal Pallagani
    • Designed a framework to detect and mitigate cognitive fatigue in LLMs using token-level signals and real-time interventions.
    • First-authored an accepted AAAI 2026 Demonstration paper; extended to full paper accepted at ICML 2026.
    • Conducting research on long-context reliability, entropy collapse, and attention decay in LLMs.
  2. Research Collaborator, University of Illinois Urbana-Champaign · Nov 2025 – Mar 2026
    • Advisor: Soorya Ram Shemgekar
    • Studied politeness framing and reward leakage in LLMs across structured tasks and instruction-following settings.
    • Performed mechanistic interpretability analysis including early-token probing and activation patching.
  3. Generative AI Intern, EY (Ernst & Young) · Jan 2025 – Mar 2025
    • Built a low-code platform for agentic AI workflows using modular DAGs, vector DBs, and LLM toolchains.
    • Implemented Celery–Redis task execution with production-grade scalability.
    • Integrated semantic agent routing, memory components, and external API/tool support (OpenAI, FAISS).
  4. Software Developer Intern, National Thermal Power Corporation · Jul 2024 – Sep 2024
    • Developed ASP.NET Core applications using MVC and Entity Framework for enterprise automation.
    • Optimized MySQL and MongoDB CRUD operations via LINQ, ensuring ACID compliance.
    • Implemented secure authentication using JWT, Identity Framework, Google reCAPTCHA, and SMTP.
  5. Strategy & IS Intern, Indian Oil Corporation Limited
    • QA for Digital Tender Management: black-box testing, UI/UX review, and performance testing.
    • Root-cause analysis and test design to improve system stability and user experience.

Projects

  1. Early Stage Lung Cancer Detection Using Deep CNNs

    Built a CNN model on the LIDC-IDRI dataset to detect pulmonary nodules in CT scans, enabling earlier diagnosis. Integrated Grad-CAM and SHAP to assist radiologists with model-driven decision support.

  2. Bhoomi: AI-Powered Agricultural Productivity

    Full-stack platform integrating drone technology and mobile solutions to enhance agricultural productivity. Developed AI models for crop disease detection and farmer assistance to improve yield and sustainability.

  3. Microphone Array-Based Direction-of-Arrival for Gunshot Detection — Smart India Hackathon 2024 Nominee

    Six-mic array with TDoA algorithms, FPGA processing, band-pass filtering, and KNN/STFT classification. Reduces false positives and improves situational awareness in public safety applications.

  4. Dynamically Optimized Cycling Navigation with Real-Time Adaptive Routing

    Real-time route optimization using OSM/OSMnx/NetworkX and Monte Carlo rollouts with dynamic safety and traffic weights, Dijkstra/A* fallback, and SCC connectivity checks.

Honors & Service

Education

B.Tech, Computer Science Engineering · Guru Gobind Singh Indraprastha University
Nov 2022 – Jun 2026 (Expected) · GPA: 8.12 / 10.0

MIT OpenCourseWare — Graduate & Advanced Coursework (Part-time)
6.825 Techniques in AI / Deep Learning for NLP · 6.864 Advanced NLP · 6.867 Machine Learning (Advanced) · 6.006 Introduction to Algorithms

Relevant Undergraduate Coursework · Prerequisite Mapping Syllabi →