Partha Sarathi Purkayastha

M.Sc. Computer Science at ETH Zürich (D-INFK)

I am a master’s student in Computer Science at ETH Zürich and a Research Assistant at SCAI Lab, working with Diego Paez-Granados. My research interests span AI in education and healthcare, human–AI interaction, and reliable agentic systems. I am particularly interested in how AI systems reason about context, make decisions, and support people.

At ETH’s LRE Lab, I worked with Mrinmaya Sachan and Heejin Do on dynamic student simulation and the evaluation of adaptive LLM tutors. My current work at SCAI uses multimodal wearable and clinical data for cardiac-recovery monitoring.

Previously, I was a Research Fellow at Microsoft Research India, working with Akshay Nambi on contextual privacy and LLM summarization. I received my B.E. in Computer Science from BITS Pilani in 2024, graduating with the Director’s Gold Medal and a 9.83/10 CGPA.

news

Sep 2026 DySAS accepted at REALM @ EMNLP 2026.
Jul 2026 TraceBound accepted at FAGEN @ ICML 2026.
1 Jul 2026 Presented Reliability as a Teammate at AIED 2026 in Seoul, South Korea.
Jun 2026 Reliability as a Teammate published at AIED 2026; Grounded Teacher published in IJCV.
May 2026 Joined ETH SCAI Lab as a Research Assistant.
Apr 2026 Our team SCAI-ropractors won Track 2 of the ASIA 2026 Data Challenge. I received a travel grant and presented our winning work at the ASIA Annual Scientific Meeting in San Antonio, Texas, on 24 April 2026.
2025 Started my M.Sc. in Computer Science at ETH Zürich.

selected publications

  1. AIED

    Reliability as a Teammate: Budgeted Verifier-in-the-Loop (BVIL) Policies for Reliable LLM Tutoring Actions

    Partha Sarathi Purkayastha

    Artificial Intelligence in Education (AIED 2026), Springer LNAI 16583

    Paper
    Summary

    Policies for deciding when to invoke verification under a budget, with the goal of making LLM tutoring actions more reliable.

  2. IJCV

    Context Aware Grounded Teacher for Source Free Object Detection

    Tajamul Ashraf, Rajes Manna*, Partha Sarathi Purkayastha*, Tavaheed Tariq, and Janibul Bashir

    International Journal of Computer Vision, 134, Article 306 (2026)

    * Equal Contribution

    Paper
    Summary

    A context-aware teacher framework for source-free object detection that addresses class imbalance and unstable pseudo-labels.

  3. REALM

    DySAS: Dynamic State-Aware Student Simulation Beyond Static Personas

    Partha Sarathi Purkayastha and Heejin Do

    Research on Agent Language Models (REALM) Workshop at EMNLP 2026 · Accepted

    Workshop
    Summary

    Psychologically grounded student simulation with evolving states, designed to evaluate adaptive LLM tutors across multi-turn interactions.

  4. FAGEN

    When Thinking Before Retrieval Hurts: TraceBound Diagnostics for Adaptive Knowledge-Graph Retrieval

    Partha Sarathi Purkayastha

    Failure Modes in Agentic AI (FAGEN) Workshop at ICML 2026 · Accepted

    arXivPDF
    Summary

    A diagnostic study of adaptive knowledge-graph retrieval, examining how extra reasoning and trace conditioning can worsen retrieval through ineffective action selection.

All publications →

Feel free to reach out via email.