danilobzdok

Our paper at #NeurIPS 2025:

1) Pre-trained transformer models impose semantic structure on inputs.

2) More skewing of transformer processing as input uncertainty increases.

3) Spurious concept activations can predict hallucinations.

arxiv.org/pdf/2509.069...

#large #language #models + #safety

📢 job opening: software developer at Mila Quebec AI Institute & McGill

As of now, my team is looking to hire a new technical staff member with abackground in a STEM area for running and curating existing AI analysis pipelines, as well as designing GUIs around them.

Please DM or email me, with CV

1/11 Excited to share our @Naturestudy led by @leonooi.bsky.social @csabaorban.bsky.social @shaoshiz.bsky.social

AI performance is known to scale with logarithm of sample size (Kaplan 2020), but in many domains, sample size can be # participants or # measurements...

doi.org/10.1038/s415...

Our Nature paper on the hashtag#scaling hashtag#behavior and economics of hashtag#machine hashtag#learning predictions in high-dimensional brain scans is out !

Congrats to the whole team.

www.nature.com/articles/s41...

Looking forward to speak at Yale University next friday.

impact-mh.org/event/impact...

I am impressed by this book I am currently reading: Kissinger's posthumous work on:
-strong anthropological views on the current AI evolution
-modes in which #AI could take space as a physical and intellectual actor
-how modes of decision making of humans and machines can coalesce and completment

Great opportunity to get generous postdoc funding to work at a Canadian university

Big congrats to our former postdoc Andrea @andrealuppi.bsky.social for earning the Brain Star Award for his work !

can-acn.org/brain-star-a...

you mean gracious?

'delineate' is a great verb

Multimodal population study reveals the neurobiological underpinnings of chronotype - Nature Human Behaviour Capitalizing on multimodel data from the UK Biobank, Bzdok and colleagues delineate the deep phenotyping and brain basis of early risers and night owls at a population scale, showing links to habit fo...

Using multi-modal data from the UK Biobank, Zhou et al. delineate the phenotyping and brain basis of early risers and night owls at a population scale, showing links to habit formation and emotional regulation.
www.nature.com/articles/s41...

Multimodal population study reveals the neurobiological underpinnings of chronotype - Nature Human Behaviour Capitalizing on multimodel data from the UK Biobank, Bzdok and colleagues delineate the deep phenotyping and brain basis of early risers and night owls at a population scale, showing links to habit fo...

Everything you always wanted to know about
how #night #owls and #early #risers are different.

Out today !

www.nature.com/articles/s41...
rdcu.be/eh8ix

@pessoabrain.bsky.social @ohbmofficial.bsky.social @sfn.org @nature.com @mcgilluniversity.bsky.social

Just finished my slides for World AI Summit Headliner talk this week.

Data Sharing - ABCD Study

FYI, if you want to be prepared to apply for the next ABCD Study data release, prepare for these new NIST security requirements 🧠: @ohbmofficial.bsky.social abcdstudy.org/scientists/d...

Using LLMs to understand how autism gets diagnosed In a study that used large language models to analyze healthcare records of Quebec francophone children, scientists argue that current criteria for diagnosing autism are in need of revision.

#AUTISM “A data-driven revision of criteria ... grounded on clinical certainty, would complement what has historically been done by expert panels and ... humans, who can be wrong." @danilobzdok.bsky.social @mila-quebec.bsky.social @cellpress.bsky.social #AI
nouvelles.umontreal.ca/en/article/2...

Jack Stanley shows how to tailor #LLMs for single-sentence prediction in #autism:

www.cell.com/cell/fulltex...

Great collabo with Eugene Belilovsky + Siva Reddy from @mila-quebec.bsky.social, and Laurent Mottron + Emmett Rabot

Clinical assessment is the gold standard for autism diagnosis, and by using amply pre-trained language models paired with a suitable interpretability strategy, thenative analysis of clinical intuition itself is now possible, guided by detailed observational text reports from clinicians. This interpretable language model framework identified repetitive and stereotyped behaviors as being empirically more relevant for autism diagnosis than social signs in the reports, motivating the possible revision of existing diagnostic instruments.

Trained language model finds repetitive behaviors and special interests are more indicative of an autism diagnosis than a lack of social skills. https://www.cell.com/cell/fulltext/S0092-8674(25)00213-2

@mila-quebec.bsky.social @Danilo Bzdo

Mila researchers Jack Stanley and Danilo Bzdok published a paper in the journal @cellpress.bsky.social. Their work uses LLMs to analyse data from written reports by medical professionals in order to better understand and improve how autism diagnosis is performed. mila.quebec/en/article/u...

Understanding #Parkinsons subtypes is key to more targeted RCTs & better therapeutics! 🧠 Our new study finds that LRRK2 PD patients show milder cortical atrophy, with preserved temporal & occipital regions compared to sporadic PD. #LRRK2 #dataviz

www.medrxiv.org/content/10.1...

🔥🔥🔥

@oprah-winfrey.bsky.social and @michaelpollan.bsky.social make unmistakably clear: there is widespread interest in #psychedelics

>1M views in less than a week, bravo !

My talk on “Neuroscience and AI: a symbiosis” @SimonsInstitute
I start with a whirlwind history of metaphors for intelligence, spanning Hippocratic hydraulics, Hobbesian clocks, Cartesian dualism to Turing, telegraphs, Hopfield, and deep learning. simons.berkeley.edu/talks/surya-...

I read that book at the time…

Out soon !

pubmed.ncbi.nlm.nih.gov/39149460/

@evfedorenko.bsky.social at Berkley LLM workshop gave insightful talk today:

-Broca / Wernicke s are speech areas, not language areas per se
-language networks similar across 50 languages
-most language voxel are INsensitive to semantics
-lesion studies suggest thought intact but speech impaired

@suryaganguli.bsky.social gives great talk at LLM workshop at Berkley

-LLM<->brain is still a new topic, less progress so far than LLM<->vision
-lets train LLM foundation models of specific brain systems and then reverse engineer them
-emerging paradigm: read-write experiments in brains+machines

Now accepted at Nature Computational Science:

Nicole shows how #Bayesian #hierarchical #models can explicitly model major #source of #population #stratification in 10,000 ABCD subjects.

pmc.ncbi.nlm.nih.gov/articles/PMC...

Special Year on Large Language Models and Transformers, Part 2 This is the second part in a special year-long program on large language models and transformers that spans the 2024–2025 academic year. This program's overarching goal is to try to understand the ong...

Will be in San Francisco beginning february for the @IVADO_Qc workshop on large language models and #transformers at Berkley’s
Simons Institute for the Theory of Computing.

simons.berkeley.edu/programs/spe...

Please DM if you want to chat.