2026
depth-wise activation steering for honest language models
re4-align workshop, iclr 2026
gracjan góral
/ˈɡrat͡sjan ˈɡural/
i started in math, ended up in language models. mostly i look at where they go wrong — what they make up, what they pretend to know, what they can't see in front of them. lately i've been borrowing tricks from psychology and visual perception.
at lute i work on robotics — small things, like helping robots put items where they belong.
five cats share my office. they aren't as helpful as they think.
the demons.
papers
2026
beyond recognition: evaluating visual perspective taking in vision language models
cvpr 2026 (findings)
2025
opengvl: benchmarking visual temporal progress for data curation
workshop on making sense of data in robotics, 2025
2025
wait, that's not an option: llms robustness with in-correct multiple-choice options
acl 2025
2024
seeing through their eyes: evaluating visual perspective taking in vision language models
workshop on responsibly building the next generation of multimodal foundational models, neurips 2024
2024
what matters in hierarchical search for combinatorial reasoning problems?
generative models for decision making workshop, iclr 2024
datasets
—
the isle — i spy with my little eye
tasks for studying visual perspective taking, scene understanding, and spatial reasoning.
—
a scaled-up successor to the isle.
writing
—
building trust with invisible robots
on humanoids and open source.