Decoded
← Feed
arXiv:2608.20186LGq-bio.NC20 aug 2026

Decoding silent reading from non-invasive EEG

1 min readHN ×2PDF ↗
CONTEXTUAL, 100% DATA15NON-CONTEXTUAL, 100% DATA10.1CONTEXTUAL, 10% DATA6.2NON-CONTEXTUAL, 10% DATA5020 WITHIN-RUN RETRIEVAL GAIN
fig. within-run retrieval gainMore training data, steadily better word decoding

49 hours of one person's EEG while reading lets a model pick out individual words.

7.5pp

Word-level retrieval gain above chance

Decoding works even where topic tracking is impossible

How it works

  1. No inner-speech labels

    Nobody can record what a person silently thinks with exact timing, so inner-speech decoders lack training data.

  2. Silent reading as proxy

    One participant read ~240,000 words in random fonts while dry EEG was recorded; a contrastive model matched EEG to Llama embeddings.

  3. Words decodable, data-limited

    Every fit beat chance, rare words decoded too, and accuracy kept rising log-linearly with more data.

recordencodetargetrankRSVP wordreading19-channelEEG windowConv encoder+ transformerLlama-3.1word embeddingCLIP contrastivematchingTop-10 wordretrievalrecordencodetargetrankRSVP wordreading19-channelEEG windowConv encoder+ transformerLlama-3.1word embeddingCLIP contrastivematchingTop-10 wordretrieval

What they found

Which comes out on top on Within-run retrieval gain?

More EEG hours, better decoding

try it

Slide the training data size up and watch the error keep falling.

decoding errorfloor10K100K1M~192,100 training trials
training trials
200K log10 trials
decoding error
0.747
10× more cuts the gap by
29%
No plateau at 49 hours: decoding is limited by data, not by EEG signal.Illustrative model, not the paper's data

Terms

6 terms

EEG Scalp electrodes recording brain voltage; here 19 dry electrodes at 600 Hz.

Like picking a song from a crowd's muffled humming: noisy alone, but with hours of practice you learn which hum matches which tune.

Where it breaks
Humming is deliberate output; EEG here records passive reading, possibly mostly visual.

Limits

  • Only one participant; cross-subject results still pending.
  • No held-out test set, so absolute numbers are upper bounds.
  • Signal may partly reflect visual word shape rather than meaning.
From the paper