Navve wasserman

PhD student @ Weizmann | Brain–Vision Models, Multimodal AI, Brain Interpretability

[DailyPapers @HuggingPapers]

From Activation to Causality

BrainCause discovers true visual representations in the human brain by combining generative models with controlled counterfactual stimuli, moving beyond activation maps to validate neural causality.

[Chris Ge @ChrisGe05]

FLUX.2's @bfl_ml text tokens aren't just holding your prompt.

During image editing, they absorb reference image content, and some of that absorbed content, like color and style, causally drives the output appearance.

New paper 🧵👇

[Martin Hebart @martin_hebart]

I'm proud we are releasing LAION-fMRI, a densely sampled 7T fMRI dataset of natural images, with very broad stimulus sampling for testing countless hypotheses & deeply exploring brain representations. It is now available at
https://t.co/hOnILHonf9
What does LAION-fMRI offer? 🧵

[Yossi Gandelsman @CVPR @YGandelsman]

Diffusion models are great, but we can squeeze out so much more from them. The only problem is that it usually requires extra training or manual representation editing. In our new paper, we show that with the current capabilities of LLMs, it is much simpler than we thought!

[Navve wasserman @NavveW]

Thanks @rohanpaul_ai for sharing our new work!

Automatic Interpretability Pipeline + Human Brain Data = 🧠🔍🔥

See how we use a large-scale automatic interpretability pipeline to discover what concepts are represented in the human brain.

Page & Demo: https://t.co/3cStXhXs0b

[Rohan Paul @rohanpaul_ai]

This paper uses AI-style interpretability tools to map which images trigger which visual concepts in the human brain.

It scales by adding about 120K extra images using predicted functional magnetic resonance imaging (fMRI) signals.

The problem is that fMRI data has about 40K voxels per person, each voxel is a tiny 3D pixel, and manual labeling does not scale.

The pipeline first breaks each brain region’s activity into patterns that can be mixed to rebuild any response, and a sparse autoencoder pushes each response to use only a few patterns.

For every pattern, it finds the top images that trigger it, captions those images, and has a model that writes text suggest shared meanings like “kitchen” or “hands in action”.

To avoid random labels, it builds a big concept list, marks each image as true or false for each concept, then keeps the concept that shows up most consistently in that pattern’s top images.

The payoff is a searchable map from image concepts to brain areas, plus a fair way to compare breakdown methods using held-out real scans.

----

Paper Link – arxiv. org/abs/2512.08560

Paper Title: "BrainExplore: Large-Scale Discovery of Interpretable Visual Representations in the Human Brain"

[DailyPapers @HuggingPapers]

Unlocking the brain's visual secrets with a new AI framework from MIT

Researchers introduce BrainExplore, an automated system that maps thousands of interpretable visual concepts directly from fMRI activity. It's a huge leap towards understanding how our minds process the world.

[DailyPapers @HuggingPapers]

Brain-IT: Reconstructs images from fMRI with unprecedented faithfulness & data efficiency

This brain-inspired approach uses a Brain-Interaction Transformer to faithfully recover visual content from fMRI. It outperforms current SoTA and achieves strong results with just 1 hour of data from new subjects, matching models trained on 40 hours.