Lightnews — Scholar-powered news

Reposted by Martin Hebart

Patxi Elosegi @patxelos.bsky.social · 9d

1/Preprint Alert🔔: Across two experiments plus a computational model, we show the visual system compresses complex scenes into summary statistics that can guide behavior without conscious access to the task-defining features. We term this the Ensemble Blindsight effect.

1 7 29

Martin Hebart @martinhebart.bsky.social · 9d

Very sorry to hear! 😢

1

Reposted by Martin Hebart

Malcolm Campbell @malcolmgcampbell.bsky.social · 19d

🚨Our preprint is online!🚨

www.biorxiv.org/content/10.1...

How do #dopamine neurons perform the key calculations in reinforcement #learning?

Read on to find out more! 🧵

10 67 190

Martin Hebart @martinhebart.bsky.social · 21d

I’m also curious to see the manuscript. I’d like to understand better how this is smarter than hierarchical partitioning (i.e. non-Bayesian model averaging ;) since this is just a different way to assign the shared variance to unique parts, but it doesn’t solve the indeterminacy. Can BMA address it?

Martin Hebart @martinhebart.bsky.social · 21d

Can’t we compute an upper bound for the “shared variance”, i.e. the area of indistinguishability, by simply taking 1 minus the sum of the unique parts of all other variables, since the shared cannot possibly explain more than that?

And a lower by assuming all other variables are unrelated to Y?

1

Reposted by Martin Hebart

Joachim Baumann @joachimbaumann.bsky.social · 26d

🚨 New paper alert 🚨 Using LLMs as data annotators, you can produce any scientific result you want. We call this **LLM Hacking**.

Paper: arxiv.org/pdf/2509.08825

$We present our new preprint titled "Large Language Model Hacking: Quantifying the Hidden Risks of Using LLMs for Text Annotation". We quantify LLM hacking risk through systematic replication of 37 diverse computational social science annotation tasks. For these tasks, we use a combined set of 2,361 realistic hypotheses that researchers might test using these annotations. Then, we collect 13 million LLM annotations across plausible LLM configurations. These annotations feed into 1.4 million regressions testing the hypotheses. For a hypothesis with no true effect (ground truth $p > 0.05$), different LLM configurations yield conflicting conclusions. Checkmarks indicate correct statistical conclusions matching ground truth; crosses indicate LLM hacking -- incorrect conclusions due to annotation errors. Across all experiments, LLM hacking occurs in 31-50\% of cases even with highly capable models. Since minor configuration changes can flip scientific conclusions, from correct to incorrect, LLM hacking can be exploited to present anything as statistically significant.$

5 94 260

Martin Hebart @martinhebart.bsky.social · 26d

P.P.S.: re-reading this, I notice the inaccuracies in my language (e.g. not expected by chance - expected given no actually shared variance). Apologies! I just quickly typed my thoughts down, hopefully in an understandable fashion!

1 1

Martin Hebart @martinhebart.bsky.social · 26d

P.S.: you might wonder where bounds for shared variance could be helpful. If your model doesn’t explain more shared variance than expected by the baseline (which is given by the lower bound), you can confidently say there is no evidence for shared variance.

1 1

Martin Hebart @martinhebart.bsky.social · 26d

6. We now use hierarchical partitioning to address interpretability issues for “unique parts” in the presence of highly collinear variables. I saw @diedrichsenjorn.bsky.social has a Bayesian method up his sleeve that is a variant of hierarchical partitioning, so it’s worth keeping an eye out for it.

1 1

Martin Hebart @martinhebart.bsky.social · 26d

5. So while the interpretation of variance partitioning can be tricky w/ collinear variables, bounds can at least help limit the range of possibilities. Curious if my intuition is correct or if anyone has looked into this.

1 2

Martin Hebart @martinhebart.bsky.social · 26d

4. Unique variance becomes very hard to interpret when we have more than 2 independent variables if two of them are highly collinear, because it may seem as if the third variable is better. Again, the upper bounds are really relevant here.

1

Martin Hebart @martinhebart.bsky.social · 26d

3. However, we can compute bounds for the shared variance. I wonder if this has ever been done. Assuming correlations of x1 with x2, x1 with y and setting r(x2,y)=0, we can figure out what shared variance we expect by chance. This doesn’t solve the interpretation problem but it constrains it.

1

Martin Hebart @martinhebart.bsky.social · 26d

2. The shared variance just means the part that cannot be assigned uniquely, and there are different ways of fairly distributing variance but we cannot figure out where the variance is coming from.

1 3

Martin Hebart @martinhebart.bsky.social · 26d

I wanted to add some thoughts to this excellent blog post, not detailed, maybe wrong, maybe useful:
1. Unique variance is easy to interpret as a lower bound of what a variable explains (the upper bound being either what the variable explains alone or what the other variables cannot explain uniquely)

Jörn Diedrichsen @diedrichsenjorn.bsky.social · 27d

Variance partitioning is used to quantify the overlap of two models. Over the years, I have found that this can be a very confusing and misleading concept. So we finally we decided to write a short blog to explain why.
@martinhebart.bsky.social @gallantlab.org
diedrichsenlab.org/BrainDataSci...

1 3 13

Reposted by Martin Hebart

Steven Scholte @neurosteven.bsky.social · 29d

🧠 New preprint: Why do deep neural networks predict brain responses so well?
We find a striking dissociation: it’s not shared object recognition. Alignment is driven by sensitivity to texture-like local statistics.
📊 Study: n=57, 624k trials, 5 models doi.org/10.1101/2025...

4 36 110

Reposted by Martin Hebart

assafbreska.bsky.social @assafbreska.bsky.social · Aug 20

We’re hiring! 🥳fully funded PhD and postdoc positions in human cognitive neuroscience @mpicybernetics.bsky.social

M/EEG and brain stimulation methods to study timing/ prediction/ attention/ oscillations
tinyurl.com/5dku4du9

@timingresforum.bsky.social @gtc-tuebingen.bsky.social

please repost 🙏

PhD and Postdoc Positions (m/f/d) in Human Cognitive Neuroscience of Dynamic Cognition – M/EEG, brain stimulation

Job Offer from August 11, 2025

tinyurl.com

2 48 59

Martin Hebart @martinhebart.bsky.social · Sep 7

Pretty amazing how you can learn something in books and in school and it explains everything around you - how the planets move, etc. But then it’s just next level when you can see it is real, with your own eyes.

11

Reposted by Martin Hebart

Max Planck Institute for Human Cognitive and Brain Sciences @mpicbs.bsky.social · Sep 5

Huge congratulations 🥳 to
@sofievalk.bsky.social
&
@reznikdan.bsky.social
for receiving ERC Starting Grants 2025! We're proud of you!
Learn more about their projects: www.cbs.mpg.de/2397348/2025...

Sofie Valk and Daniel Reznik receive ERC Starting Grants

www.cbs.mpg.de

2 5 49

Martin Hebart @martinhebart.bsky.social · Sep 5

Cool stuff! So you are looking at the unique multivariate corr structure shared between two sets of variables, which provides a lower bound of the true unique correl.? Is there a difference between keeping one set of variables uncorrelated with the confounds vs. both? (akin to semi vs partial corr)

1 2

Reposted by Martin Hebart

Sebastian Michelmann @s-michelmann.bsky.social · Sep 5

🚀Excited to share our project: Canonical Representational Mapping for Cognitive Neuroscience. @schottdorflab.bsky.social and I propose a novel multivariate method to isolate neural representations aligned with specific cognitive hypotheses 🧵https://www.biorxiv.org/content/10.1101/2025.09.01.673485v1

1 30 78

Reposted by Martin Hebart

Imaging Neuroscience @imagingneurosci.bsky.social · Sep 5

Launched in 2023, Imaging Neuroscience is now firmly established, with full indexing (PubMed, etc.) and 700 papers to date.

We're very happy to announce that we are able to reduce the APC to $1400.

Huge thanks to all authors, reviewers, editorial team+board, and MIT Press.

2 79 230

Martin Hebart @martinhebart.bsky.social · Sep 4

It’s also really hard for me to see it as something different than a bat. And it’s also nonsense. It makes it even more interesting that kiddo saw it for longer periods of time in either state. Even though our brain can have competing interpretations, here the other interpretation doesn’t win over.

1 1

Martin Hebart @martinhebart.bsky.social · Sep 4

bsky.app/profile/mart...

Martin Hebart @martinhebart.bsky.social · Sep 4

It’s easier to see if you focus on the bottom or top! It’s just a pattern of blobs. Do you see it now? On the left like bush or like half butterflies. On the right like eye glasses.

2 3

Martin Hebart @martinhebart.bsky.social · Sep 4

It’s easier to see if you focus on the bottom or top! It’s just a pattern of blobs. Do you see it now? On the left like bush or like half butterflies. On the right like eye glasses.

2

Martin Hebart @martinhebart.bsky.social · Sep 4

The photo was blurry, so that reduces the effect! It’s easier to see here.

I would also like to know more about what is known in that age range!

2 2