Akmal Ashirmatov

Akmal Ashirmatov

M.Sc. Student at ETH Zürich
akmalashirmatov@gmail.com


About Me

I’m an M.Sc. student in Computer Science at ETH Zürich, working on LLM post-training methods in Prof. Andreas Krause’s Learning & Adaptive Systems group. I’m supervised by Jonas Hübotter, Frederike Lübeck and Thomas Kleine Buening. My recent work focuses on self-distillation for reasoning and self-play for code generation.

Previously, I interned at Google in London, where I designed and prototyped a new tool and contributed code to the Gemini Live agent harness, and at Huawei’s Zurich Research Center, where I worked on kernel-level optimization for LLM inference.

Always happy to chat-feel free to reach out!

Find my publications on Google Scholar and my code on GitHub.

News

Publications

  1. Visual abstract of contrastive self-distillation: attractive and repulsive teacher signals separate correct and incorrect answers from student behavior.
    Anton Baumann, Akmal Ashirmatov, Leo Schmidt-Traub, Frederike Lübeck, Jonas Hübotter, Thomas Kleine Buening, Andreas Krause
    FLLMPT @ NeurIPS 2026
    Poster
  2. Self-play framework: a proposer and a solver share one base model; the proposer is rewarded for problems the solver solves about half the time, and the solver for solving them.
    Akira Yoshiyama*, Akmal Ashirmatov*, Jonas Hübotter, Ido Hakimi, Andreas Krause
    DL4C @ ICML 2026 · *equal contribution
    Poster
  3. ESCAPE pipeline: submit a prompt, edit it by adding, swapping or reweighting words, edit cross-attention maps, and compare the resulting images.
    Diego Zafferani*, Giovanni De Muri*, Johanna Hedlund Lindmar*, Akmal Ashirmatov*, Sinie van der Ben, Rita Sevastjanova, Mennatallah El-Assady
    IEEE VIS 2025 · VISxAI workshop · *equal contribution

Experience

Course Projects

Awards & Competitions