// forward pass
Hey! I'm Marcel, a recent grad from Brown University with a degree in Computer Science and Economics. I'm a backend engineer and ML/AI researcher with a passion for intersecting software engineering with cutting-edge research in ML/AI. My current research interests lie in LLMs, model interpretability, and self-supervised learning. I'm a member of the GalilAI Group with Prof. Randall Balestriero and an incoming Software Engineer at Pinterest!
// layer_01 · experience_units
One unit per role, wired in chronological order — t-0 is the most recent. Hover to read a unit's activation, click to probe it.
signal flows t-8 (past) ⟶ t-0 (now)
// checkpoints · publications
states of the research track, saved to disk
ckpt_0001 · saved 2025
Marcel Mateos Salles, Praney Goyal, Pradyut Sekhsaria, Hai Huang, Randall Balestriero
Shows that a handful of spurious tokens in finetuning data is enough to steer the behavior of LoRA-finetuned language models.
ckpt_0002 · saved 2026
Marcel Mateos Salles
Investigates when and why parameter-efficient finetuning amplifies reliance on spurious correlations, and what that means for safe deployment.
// feature_dictionary
Select features to filter which project units activate downstream.
22 features idle — select to feed layer_02
// layer_02 · project_units
Connected units share technologies or collaborators — stronger overlap, stronger connection. Click a unit to open its neuron probe.
probe filter: none — all 6 units active
// output_layer
epoch 000 · loss 2.413 · optimizer: AdamW · descending…
Think I'd be a good fit for your team? I'm always interested in hearing about new projects, opportunities, and challenges!
Email Me!