Pedram Mouseli
Pedram Mouseli

AI Researcher · Postdoctoral Fellow, University of Toronto

I'm Pedram, an AI researcher interested in the representations models learn: what they capture, whether they generalize, and whether they actually shape behaviour. I come at this from two sides. I build deep learning models of health data, such as physiological and neural signals, that transfer to new people and new conditions. I also study reasoning in large language models: how internal state and reasoning shape what models do.

Research themes

Representation learning for health data

Self-supervised, multimodal models of physiological and neural signals.

Reasoning and internal state in LLMs

Interpretability, and testing which internal states and reasoning traces actually change behaviour.

Generalization

Models that hold up across people, tasks and conditions, and that keep an accurate picture of how a situation changes.

Selected work

InterpretabilityUnder review · 2026

Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents

Does an LLM agent's reasoning actually drive its actions? In a task where agents must track a changing hidden state, we separate generating reasoning at each step from keeping earlier reasoning in context. Most of the benefit comes from reasoning written at the moment of action, and a state that probes can decode from activations is a limited guide to what the agent does.

Reasoning helps as step-by-step computation, not as memory

P. Mouseli, M. Moayedi

Preprint → Code → Project Page →
Deep learning · BiosignalsbioRxiv · 2026

LENS: a transferable neuromuscular signature of musculoskeletal pain intensity

LENS (Latent sEMG Neuromuscular Signature) decodes exertion-evoked musculoskeletal pain intensity from a single surface EMG channel. It is pre-trained with a cross-modal self-supervised objective that reconstructs muscle oxygenation from sEMG in 183 adults. Trained only on healthy muscle physiology, it transfers to an unseen chronic-pain cohort (myogenous TMD) through label-free test-time adaptation.

Pain prediction in unseen chronic-pain cohort: ρ = 0.61 · AUROC = 0.86

P. Mouseli, N. De Vera, S. Sagheer, I. Jurisica, W.D. Reid, M.S. Angst, N. Aghaeepour, M. Moayedi, I. Cioffi

Preprint → Code → Data →
Machine learning · PainPNAS · 2026Co-first author

Nociceptive sensitization and pain perception are dissociable in humans

I built the machine learning model that predicts who perceives the thermal grill illusion, pain evoked by interlaced innocuous warm and cool bars, from behavioral, demographic and psychological features. The model predicted pain perception but not the size of nociceptive sensitization, supporting the paper's central finding: sensitization can occur without pain, so illusion responsiveness reflects a perceptual phenotype.

Predicting TGI responders: AUC = 0.81 · N = 251

M. Cormie*, P. Mouseli*, C. Martini, K. De Silva, I. Lauric, O. Khalil, D. Seminowicz, M. Moayedi  ·  *equal contribution

Paper → Code →

Publications

  • 2026
  • 2026
    LENS: a transferable neuromuscular signature of musculoskeletal pain intensity
    Mouseli, P., De Vera, N., Sagheer, S., Jurisica, I., Reid, W.D., Angst, M.S., Aghaeepour, N., Moayedi, M., Cioffi, I.
    bioRxiv
  • 2026
    Nociceptive sensitization and pain perception are dissociable in humans
    Cormie, M.*, Mouseli, P.*, Martini, C., De Silva, K., Lauric, I., Khalil, O., Seminowicz, D., Moayedi, M.
    Proceedings of the National Academy of Sciences, 123(35), e2603667123
  • 2026
    Prefrontal cortex connectivity during right and left hand dexterity tests in younger and older adults
    Kulkarni, M., Rassam, P., Mouseli, P., Date, M., Mori, T., Van Hollebeke, M., Campos, J.L., Rozenberg, D., Reid, W.D.
    PLoS One, 21(2), e0342547
  • 2025
    Functional connectivity differences in adults with ADHD: a MEG study
    Mouseli, P., Dunkley, B.T., Moayedi, M.
    bioRxiv
  • 2025
    Social media use and jaw motor activity: insights from electromyography and self-report data
    Lee, S., Koshkebaghi, D., Mouseli, P., Cioffi, I.
    Journal of Oral Rehabilitation, 52(11), 1945–1954
  • 2024
    Subjective salience ratings are a reliable proxy for physiological measures of arousal
    Hadjis, G.E., Atlas, L.Y., Mouseli, P., Sexton, C.A., McAndrews, M.P., Moayedi, M.
    Pain
  • 2023
    Leveraging the resting brain to predict memory decline after temporal lobectomy
    Audrain, S., Barnett, A., Mouseli, P., McAndrews, M.P.
    Epilepsia, 64(11), 3061–3072
  • 2023
    Insula-cingulate structural and functional connectivity: an ultra-high field MRI study
    Cormie, M.A., Kaya, B., Hadjis, G.E., Mouseli, P., Moayedi, M.
    Cerebral Cortex, 33(17), 9787–9801
  • 2022
    EEG ERP Preregistration Template
    Govaart, G.H., Schettino, A., … Mouseli, P., … Paul, M.
    MetaArXiv

* equal contribution. Full list, including conference abstracts and talks, in my CV.

About

I'm a Postdoctoral Fellow at the University of Toronto and a Research Analyst at the Krembil Research Institute (University Health Network). My PhD in neuroscience at the University of Toronto focused on multimodal machine learning for pain, and before that I trained as an engineer, with an MSc in Mechatronics Engineering where I studied brain networks with MEG.

I've taught deep learning for several years, currently as Head Teaching Assistant for Applied Fundamentals of Deep Learning at U of T Engineering, and earlier as a project TA for Neuromatch Academy's Deep Learning and Computational Neuroscience courses. I have reviewed for Nature Communications, the Journal of Neuroscience and PAIN.

  • 2026 –Postdoctoral Fellow · University of Toronto
  • 2022 –Research Analyst · Krembil Research Institute, UHN
  • 2022 –(Head) Teaching Assistant, Applied Fundamentals of Deep Learning · U of T Engineering
  • 2021 – 2026PhD, Neuroscience · University of Toronto
  • 2017 – 2020MSc, Mechatronics Engineering · K.N. Toosi University of Technology