iiinova — where your digital self keeps evolving

Andrea de Varda @andreadevarda.bsky.social · 01/09/2026

👀 Behavior: effort dominates. SFL (three numbers) performs better than 1600-dim embeddings, esp. in eye-tracking. Adding EMB to SFL gives only small gains. 🧠 Brain: the opposite. Embeddings predict fMRI/N400 responses far better than effort. (8/10)

100

Andrea de Varda @andreadevarda.bsky.social · 01/09/2026

We test this on 13 datasets: 👀 4 eye-tracking, 3 self-paced reading, 1 Maze, 🧠 1 ERP (N400), and 4 fMRI. All responses are averaged across participants and brain responses are averaged across voxels/electrodes so brain and behavior are comparable (1-dimensional). (7/10)

100

Andrea de Varda @andreadevarda.bsky.social · 01/09/2026

Our hypothesis: behavior shows a bottleneck. Rich representations get compressed into effort dimensions (SFL) before they can influence processing times. Brain responses have more direct access to the high-dimensional representations. (6/10)

100

Andrea de Varda @andreadevarda.bsky.social · 01/09/2026

We use LMs to operationalize effort and meaning in one framework. From the same GPT/GPT-2 models we get (i) surprisal (+freq and len; SFL) for effort, and (ii) contextual embeddings (EMB), high-dimensional vectors that encode form and meaning. (5/10)

100

Andrea de Varda @andreadevarda.bsky.social · 01/09/2026

New preprint! 🧠👀🤖 Behavioral and brain responses to language reflect different levels of linguistic representation w/ @whylikethis.bsky.social , @evfedorenko.bsky.social , and @rplevy.bsky.social (1/10)

1356

Andrea de Varda @andreadevarda.bsky.social · 19/11/2025

Token count also captures differences across tasks. Avg. token count predicts avg. RT across domains (r = 0.97, left), and even item-level RTs across all tasks (r = 0.92 (!!), right). (5/6)

100

Andrea de Varda @andreadevarda.bsky.social · 19/11/2025

We found that the number of reasoning tokens generated by the model reliably correlates with human RTs within each task (mean r = 0.57, all ps < .001). (4/6)

110

Andrea de Varda @andreadevarda.bsky.social · 19/11/2025

Large reasoning models can solve many reasoning problems, but do their computations reflect how humans think? We compared human RTs to DeepSeek-R1’s CoT length across seven tasks: arithmetic (numeric & verbal), logic (syllogisms & ALE), relational reasoning, intuitive reasoning, and ARC (3/6)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

Encoding models trained on existing fMRI datasets successfully predicted responses in new languages, generalizing across stimuli types and modalities (11/)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

In the “across” condition, performance improves for models with stronger cross-lingual semantic alignment (where translations cluster together in the embedding space) (9/)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

But what kind of model properties influence LM-to-brain alignment across languages? In the “within” condition, encoding performance is highest for models with good next-word prediction abilities (8/)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

We also replicated in a cross-lingual setting the finding that the best fit to brain responses is obtained in intermediate-to-deep layers (for each subplot pair, the left one is “within”, the right one “across”) (7/)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

We evaluated 20 multilingual LMs with different architectures and training objectives, and all of them were able to predict brain responses in the various languages (“within”) and critically, generalized zero-shot to unseen languages (“across”) (6/)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

Critically, we fit two kinds of encoding models: 1️⃣ “within” encoding models, training and testing on data from a single language with cross-validation 2️⃣ “across” encoding models, training in N-1 languages and testing in the left-out language (5/)

100

Andrea de Varda @andreadevarda.bsky.social · 04/02/2025

In Study I, we: 1️⃣ Present participants with auditory passages and record their brain responses in the language network 2️⃣ Extract contextualized word embeddings from multilingual LMs 3️⃣ Fit encoding models predicting brain activity from the embeddings (4/)

100