Farnoush Rezaei Jafari

Farnoush
Rezaei Jafari

PhD candidate
TU Berlin · BIFOLD

Interpretability & efficient AI

I am currently a Student Researcher at . Prior to that, I was a Research Intern at Microsoft Research, where I worked on the interpretability of biological foundation models.

I am also a PhD candidate in the Machine Learning / Intelligent Data analysis group at Technische Universität Berlin and BIFOLD (Berlin Institute for the Foundations of Learning and Data), working with Prof. Dr. Klaus-Robert Müller. My primary focus lies in enhancing the interpretability and efficiency of multi-modal LLMs, as well as vision and language models.

Before embarking on my Ph.D. journey, I accomplished my M.Sc. in Computer Science at TU Berlin in the year 2021. I did my Master’s thesis, titled “Analyzing the Importance of Temporal Information in 3D ConvNets for Human Action Recognition Through Explanations” under the supervision of Prof. Dr. Klaus-Robert Müller and Prof. Dr. Jürgen Gall.

news

Sep 8, 2026 I joined as a Student Researcher.
May 1, 2026 Microsoft I joined Microsoft Research Cambridge as a Research Intern.
Oct 22, 2025 🥳 Our paper, RelP accepted as a spotlight at Mechanistic Interpretability Workshop at NeurIPS 2025.
Mar 17, 2025 🎉 I have been accepted into the 8.0 Training Program, led by Neel Nanda.
Jan 16, 2025 🎙️ Invited talk at Microsoft ASG about our NeurIPS 2024 paper, MambaLRP.
Jan 6, 2025 🥳 Our paper Towards Symbolic XAI has got accepted at Information Fusion.
Sep 27, 2024 🎉 Our paper, MambaLRP, has got accepted at NeurIPS 2024.
Aug 30, 2024 📢 Check out our new preprint Towards Symbolic XAI.
Jun 11, 2024 📢 Check out our new preprint on the interpretability of selective state space sequence models, called MambaLRP.
Jul 3, 2022 🥳 Our paper, ATS is accepted at ECCV 2022 as an oral presentation.
Jun 14, 2022 🎙️ Presented our paper, ATS at Franco-German Workshop at Inria.
Nov 17, 2021 🚀 I started my PhD journey.

Selected Publications

* denotes equal contribution.

  1. Figure for RelP: Faithful and Efficient Circuit Discovery in Language Models via Relevance Patching

    MechInterp Workshop @ NeurIPS 2025, Spotlight

    RelP: Faithful and Efficient Circuit Discovery in Language Models via Relevance Patching

    F. Rezaei Jafari , O. Eberle , A. Khakzar , N. Nanda
    Paper Code

    Circuit/Subgraph-Level Model Analysis; LLMs; Mechanistic Interpretability; Sparse Feature Circuits

  2. Figure for Towards Symbolic XAI – Explanation Through Human Understandable Logical Relationships Between Features

    Information Fusion · 2025

    Towards Symbolic XAI – Explanation Through Human Understandable Logical Relationships Between Features

    T.* Schnake , F.* Rezaei Jafari , J Lederer , P. Xiong , S. Nakajima , S. Gugler , G. Montavon , K-R. Müller
    Paper Code

    Compositional Reasoning in LLMs & Vision Transformers; Subgraph-Level Model Analysis; Bridging Mechanistic Interpretability with Symbolic Reasoning; Logical Explanations for Transformers & GNNs

  3. Figure for MambaLRP: Explaining Selective State Space Sequence Models

    NeurIPS · 2024

    MambaLRP: Explaining Selective State Space Sequence Models

    F. Rezaei Jafari , G. Montavon , K-R. Müller , O. Eberle
    Paper Code

    State-Space Models; Mamba LLMs; Vision Mamba; Interpretability; Identifying Model Biases; Analyzing Long-Range Dependencies; Introducing a Novel Evaluation Metric for Needle-in-a-Haystack

  4. Figure for Adaptive Token Sampling For Efficient Vision Transformers

    ECCV, Oral · 2022

    Adaptive Token Sampling For Efficient Vision Transformers

    M.* Fayyaz , S.* Abbasi Kouhpayegani , F.* Rezaei Jafari , S. Sengupta , E. Sommerlade , H. Vaezi Joze , H. Pirsiavash , J. Gall
    Paper Website

    Test-Time Computation Scaling; Efficient Image/Video Transformers; Parameter-Free Adaptive Token Sampling; Emergent Capabilities in Vision Transformers