iiinova — where your digital self keeps evolving

ines-schoenmann.bsky.social

@ines-schoenmann.bsky.social

71 followers 28 following 11 posts

Cognitive scientist focusing on language processing and the phenomenology of vision

Posts Replies Media

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 28/04/2026

Ah no, I didn't know that was possible, but looks like a very cool option. Would definitely something I'd add next time!

000

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

More broadly, the paper proposes a constructive criterion for future work: pre-onset neural effects should only count as evidence for prediction if they go beyond what is explainable from passive stimulus dependencies. Paper: elifesciences.org/articles/106... (8/8)

elifesciences.org

Stimulus dependencies—rather than next-word prediction—can explain pre-onset brain encoding in naturalistic listening designs

The ability to predict brain activity from words before they occur can be explained by information shared between neighbouring words, without requiring next-word prediction by the brain.

1113

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

Crucially, the implication of our work is not that the brain does not perform prediction. Rather, pre-onset embedding-based encoding alone cannot adjudicate that question unless it exceeds a passive stimulus-dependent baseline. (7/8)

1153

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

We also tested proposed fixes from the literature, including repeated-bigram removal and residualising neighbouring word information. Neither was sufficient: the effects persisted in the passive control systems. (6/8)

141

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

Even more strikingly, the same hallmarks also appeared in the speech acoustics. Hence, these signatures need not reflect predictive neural computations, but can emerge from temporal dependencies in the stimulus alone. (5/8)

1101

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

They do. Both pre-onset encoding and its modulation by word predictability emerged in passive control systems, including static word embeddings themselves. (4/8)

161

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

We, therefore, asked a simple question: if the same analysis is applied to systems that encode the stimulus but cannot predict future words, do the canonical “prediction” hallmarks still appear? (3/8)

161

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

A growing body of literature has taken successful preonset encoding of future word embeddings as evidence that the brain predicts upcoming words during naturalistic language comprehension. However neighbouring words in naturalistic listening situations already share rich statistical structure. (2/8)

142

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 27/04/2026

New peer-reviewed paper w/ @mheilbron.bsky.social, @predictivebrain.bsky.social & Jakub Szewczyk! Pre-onset brain encoding has been taken as evidence that brains–like LLMs–predict upcoming words. We show that the same signatures arise in systems that cannot predict. (elifesciences.org) (1/8)

312050

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 16/01/2026

So, as a speech act, when we say "What's the update on the shelf?", we are not interested in your specific furniture, but you should read it as "We are listening to the podcast and we are part of your in-group of ideas."

000

ines-schoenmann.bsky.social @ines-schoenmann.bsky.social · 16/01/2026

@michaelhobbes.bsky.social & @notalawyer.bsky.social On the longevity of The Shelf: Similar to kids saying 6-7, it's a meme for us listeners to identify ourselves to you and to one another. Linguistically speaking it has become a shibboleth which is why it persisted after you stopped mentioning it.

120