Ariel Goldstein

This paper challenges the traditional view that neural plasticity is essential for adaptive learning by showing that sufficiently trained large language models can achieve adaptability through computation alone.

This study shows that the multiple-demand network encodes content-sensitive linguistic representations during naturalistic comprehension, independent of cognitive demand.

The paper shows that word prediction and semantic processing continue during mind-wandering, even when subjective experience is decoupled from speech content.

The paper investigates neural dynamics of lexical alternatives during speech production using electrocorticography and language models, finding three principles of how these alternatives are represented in the brain.

The paper introduces a dataset of ECoG recordings from nine participants listening to a podcast, along with linguistic features and tutorials for analysis.

This study compares how LLMs and the brain process narrative context, introducing an incremental LLM that better predicts DMN neural activity.

This study shows that speakers' brains generate information-rich words through extra neural computation, evidenced by stronger neural activity before articulation, and that LLMs similarly require deeper computation for improbable words.

The paper shows that linguistic embedding spaces can capture neural alignment during conversations, revealing shared meaning between speakers and listeners.

This paper analyzes the functional specialization of transformer-based language models and finds similarities with the human brain's cortical language network.

The paper shows that larger language models better capture linguistic structure and predict neural activity, with performance peaking in earlier layers as model size increases.

The paper shows that aligning brains into a shared space improves LLM-based encoding model performance and denoising of individual brain responses.

The paper uses a deep speech-to-text model to predict neural responses during spontaneous conversations, revealing a distributed cortical hierarchy for speech and language processing.

This study finds a correspondence between the layered structure of deep language models and the temporal structure of neural activity during language processing in the human brain.

The paper shows that brain embeddings from the IFG share geometric structure with DLM contextual embeddings, enabling zero-shot encoding and decoding of unseen words.

This paper uses fine-tuned deep language models to model differences in narrative interpretation based on listeners' perspectives, showing alignment with neural responses and cognitive networks.

The paper introduces a benchmark dataset of fMRI data from human subjects listening to naturalistic stories for evaluating language and narrative comprehension models.

In nine epilepsy patients, cortical activity carries next-word information up to 800 ms before onset, tracks GPT-2 surprise after onset, and fits GPT-2 contextual embeddings better than static ones.

The paper compares evolutionary processes and over-parameterized neural networks, arguing that both use blind fitting to solve complex tasks without explicit representation.