Interested in efficiently decoding these brain signals? Go check out our companion AI paper:
ai.meta.com/research/pub...

7/8

Result 4: This dynamic code is observed for all levels of the language hierarchy. Critically, it is level-dependent: context representations “move” more slowly in brain activity than letter representations, allowing a seamless unfolding of language representations.

6/8

Result 3:
How does the brain avoid the interference induced by such overlapping representations?

Thanks to a dynamic code! The representations of successive letters continuously move across different neural subspaces.

Result 2: Paradoxically, the representations of letters last much longer than their respective corresponding actions, resulting in a representational overlap of successive letters in brain activity.

4/8

Result 1: We find that, before typing each word, the brain activity is marked by a top-down sequence of representations: context-level representations can be decoded before those of words, syllables, and letters.

3/8

Method: We used MEG to record the brain activity of participants while they typed sentences.

Using linear decoding, we then evaluate whether the brain represents a hierarchy of linguistic features before each word is typed.

2/8

Our paper from AI at Meta and @bcbl.bsky.social is out on arxiv 🔥
“From Thought to Action: How a Hierarchy of Neural Dynamics Supports Language Production”
arxiv.org/abs/2502.07429

How does the brain transform a thought into a sequence of motor actions?

Results summarized in 🧵1/8

Result 2: Paradoxically, the representations of letters last much longer than their respective corresponding actions, resulting in a representational overlap of successive letters in brain activity.

4/8

Result 1: We find that, before typing each word, the brain activity is marked by a top-down sequence of representations: context-level representations can be decoded before those of words, syllables, and letters.

3/8

Method: We used MEG to record the brain activity of participants while they typed sentences.

Using linear decoding, we then evaluate whether the brain represents a hierarchy of linguistic features before each word is typed.

2/8