Photo of Aayush Mishra

Aayush Mishra

I'm a fifth-year Computer Science PhD student at Johns Hopkins University, advised by Anqi (Angie) Liu and co-mentored by Daniel Khashabi. I'm currently a MATS 11.0 fellow, mentored by Isak Etinger from the OpenAI Safety Team.

My research focuses on efficiency and reliablity in AI systems (specifically LLMs). I have 3+ years of industry AI research experience at Qualcomm, Adobe, Microsoft and Siemens. I received my bachelors in Computer Science from IIT Mandi.

Selected publications

  1. Enabling Dynamic Computation in Looped LMs
    Aayush Mishra, et al.
    Preprint 2026thread
  2. Steered LLM Activations are Non-Surjective
    Aayush Mishra, Daniel Khashabi, Anqi Liu
    Sci4DL & Re-Align @ ICLR 2026codethread
  3. IA2: Alignment with ICL Activations Improves Supervised Fine-Tuning
    Aayush Mishra, Daniel Khashabi, Anqi Liu
    ICLR 2026codethread
  4. Genomic Next-Token Predictors are In-Context Learners
    Nathan Breslow, Aayush Mishra, et al.
    TMLR 2026Oral @ MLGenX (ICLR 2026)codethread
  5. ODD: Overlap-aware Estimation of Model Performance under Distribution Shift
    Aayush Mishra, Anqi Liu
    UAI 2025codethread
  6. EigenLoRA: Recycle Trained Adapters for Resource-Efficient Adaptation and Inference
    Aayush Mishra*, Prakhar Kaushik*, et al.
    Preprint 2024
  7. Do Pretrained Transformers Learn In-Context by Gradient Descent?
    Aayush Mishra*, Lingfeng Shen*, Daniel Khashabi
    ICML 2024Oraltalk

* equal contribution · Full list on Google Scholar →

Experience

  1. present
    MATS x OpenAI · AI Safety Research Fellow (11.0)

    Misalignment monitoring in LLMs, and steerability/interpretability of looped language models.

  2. 2026
    Qualcomm · AI Research Intern

    Made early exit practical in looped LMs, cutting FLOPs and KV-cache memory by 25% at full-depth quality.

  3. 2024
    Adobe · AI Research Intern

    EigenLoRA: recycles trained LoRA adapters so new tasks train up to 100× fewer parameters, with ~18× less memory for multi-adapter inference.

  4. 2021 – 22
    Microsoft · Data & Applied Scientist, Bing Shopping

    Query-to-product-category classification over ~18k classes and a relevance metric for offer ranking.

  5. 2019 – 21
    Siemens · Research Engineer

    CNN explainability and compression for edge deployment; an RL agent that finds safety violations of self-driving agents in simulation. 3 invention disclosures.