publications

2026

  1. ICLR 2026
    In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations
    Mohammad Aflah Khan, Mahsa Amani, Soumi Das, Bishwamittra Ghosh, Qinyuan Wu, Krishna P. Gummadi, Manish Gupta, and Abhilasha Ravichander
    The Fourteenth International Conference on Learning Representations, ICLR 2026, April 23-27, 2026, Rio de Janeiro, Brazil, 2026
    Present in The International Association for Safe & Ethical AI second annual conference (IASEAIʼ26), non-archival
    HTML
  2. WWW 2026
    Algorithmic Self-Portrait:Deconstructing Memory in ChatGPT
    Abhisek Dash*, Soumi Das*, Elisabeth Kirsten*, Qinyuan Wu*, Sai Keerthana Karnam, Krishna P. Gummadi, Thorsten Holz, Muhammad Bilal Zafar, and Savvas Zannettou
    ACM Web Conference 2026, Dubai, UAE, 2026
    Present in The International Association for Safe & Ethical AI second annual conference (IASEAIʼ26), non-archival
    arXiv Code
  3. ICLR 2026
    Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs
    Qinyuan Wu, Soumi Das, Mahsa Amani, Bishwamittra Ghosh, Mohammad Aflah Khan, Krishna P. Gummadi, and Muhammad Bilal Zafar
    The Fourteenth International Conference on Learning Representations, ICLR 2026, April 23-27, 2026, Rio de Janeiro, Brazil, 2026
    Present in The International Association for Safe & Ethical AI second annual conference (IASEAIʼ26), non-archival
    arXiv Supp Code
  4. IASEAI 2026
    Revisiting Privacy, Utility, and Efficiency Trade-offs when Fine-Tuning Large Language Models
    Soumi Das, Camila Kolling, Mohammad Aflah Khan, Mahsa Amani, Bishwamittra Ghosh, Qinyuan Wu, Till Speicher, and Krishna P. Gummadi
    The International Association for Safe & Ethical AI second annual conference (IASEAIʼ26), 2026
    arXiv

2025

  1. Workshop@ICML 2025
    Rethinking Memorization Measures in LLMs: Recollection vs. Counterfactual vs. Contextual Memorization
    Bishwamittra Ghosh, Soumi Das, Qinyuan Wu, Mohammad Aflah Khan, Krishna P. Gummadi, Evimaria Terzi, and Deepak Garg
    The Impact of Memorization on Trustworthy Foundation Models: ICML 2025 Workshop, 2025
    arXiv
  2. ArXiv
    Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents
    Mathis Pink, Qinyuan Wu, Vy Ai Vo, Javier Turek, Jianing Mu, Alexander Huth, and Mariya Toneva
    2025
    arXiv
  3. WSDM 2025
    Towards Reliable Latent Knowledge Estimation in LLMs: Zero-Prompt Many-Shot Based Factual Knowledge Extraction
    Qinyuan Wu, Mohammad Aflah Khan, Soumi Das, Vedant Nanda, Bishwamittra Ghosh, Camila Kolling, Till Speicher, Laurent Bindschaedler, Krishna P Gummadi, and Evimaria Terzi
    The 18th ACM International Conference on Web Search and Data Mining (WSDM 2025), Hannover, Germany, 2025
    arXiv HTML Supp Code Poster

2024

  1. ArXiv
    Assessing Episodic Memory in LLMs with Sequence Order Recall Tasks
    Mathis Pink, Vy A Vo, Qinyuan Wu, Jianing Mu, Javier S Turek, Uri Hasson, Kenneth A Norman, Sebastian Michelmann, Alexander Huth, and Mariya Toneva
    arXiv preprint arXiv:2410.08133, 2024
    arXiv Supp Code
  2. ArXiv
    Understanding Memorisation in LLMs: Dynamics, Influencing Factors, and Implications
    Till Speicher, Mohammad Aflah Khan, Qinyuan Wu, Vedant Nanda, Soumi Das, Bishwamittra Ghosh, Krishna P Gummadi, and Evimaria Terzi
    arXiv preprint arXiv:2407.19262, 2024
    arXiv Code

2022

  1. Journal
    Exponential negation of a probability distribution
    Qinyuan Wu, Yong Deng, and Neal Xiong
    Soft Computing, 2022
    HTML