A language encoding model is a black-box model of how the brain processes language.

The model takes a language snippet as input and outputs a prediction of how a person's brain will light up when it hears that snippet.

Language encoding models are useful tools for studying biological language processing, as they allow us to rapidly perform experiments that would be impractical in humans.

Drag a snippet into the box to see how the model predicts this person’s brain will respond.

In our newest paper, we use this method, Generative Causal Testing, to drive activity in targeted areas across cortex.

Generative causal testing to bridge data-driven models and scientific theories in language neuroscience

Richard J. Antonello^†, Chandan Singh^†, Shailee Jain, Aliyah Hsu, Sihang Guo, Jianfeng Gao, Bin Yu^‡, and Alexander G. Huth^‡

† Equal contribution‡ Joint supervision

Nature Neuroscience (accepted, 2026)

Click the box to read the paper

beside the lake she put the kettle on turn left at the bridge the dog was asleep back to the station he almost missed the joke overlooking the town square a little more lemon the path through the woods she recognized his voice near the old harbor the music grew louder at the end of the hall we split the last piece outside the library he kept his promise beyond the hills the coffee smelled good the familiar way home she counted the change across from the bakery the rain had stopped along the riverbank we laughed about it later through the courtyard he rolled up his sleeves the room facing the garden her hands were cold the phone rang twice a bowl of warm soup he had forgotten her name the children were singing she folded the letter we waited until morning the bread was still warm he changed his mind a familiar melody she opened the envelope the glass felt heavy we ordered the same thing he smiled at the news the baby finally slept she stirred in the sugar it cost twenty dollars we saved some for later the cat stretched lazily he took a deep breath she was glad he called

Generate paragraphs with this input as the theme.

LLM

We took the train from Lisbon to Porto and walked from São Bento station down to the Douro River. The next morning, we headed north to Braga, then crossed into Spain and spent the evening in Santiago de Compostela.

A language encoding model is a black-box model of how the brain processes language.

Language encoding models are useful tools for studying biological language processing, as they allow us to rapidly perform experiments that would be impractical in humans.

These models are powerful, but as black boxes, they are difficult to understand.

Let's say we want to better understand how the model predicts responses in retrosplenial cortex, an area in the brain's limbic lobe.

We can learn about the model by examining its responses to a broad collection of snippets.

Some snippets will drive activations in RSC more than others. Can you notice any patterns?

All of the top snippets mention locations.

An LLM can recognize these kinds of patterns too.

This gives us an automated method for generating candidate explanations for areas like RSC.

We can use these candidate explanations to generate driving stimuli to test in vivo.

This gives us a causal feedback mechanism to evaluate the semantic hypotheses generated by encoding models.

In our newest paper, we use this method, Generative Causal Testing, to drive activity in targeted areas across cortex.

Generative causal testing to bridge data-driven models and scientific theories in language neuroscience