Put another way, EEG moves top-10 retrieval out of 512 candidate words from 20.3% -> 27.7%. Hardly "open-vocabulary word decoding"
The paper is also missing a critical control of feeding Llama the preceding ground-truth text and asking how well it ranks the current word among the same candidate set.
Comments
Put another way, EEG moves top-10 retrieval out of 512 candidate words from 20.3% -> 27.7%. Hardly "open-vocabulary word decoding"
The paper is also missing a critical control of feeding Llama the preceding ground-truth text and asking how well it ranks the current word among the same candidate set.