Come check out my poster at #SfN, Wed (8am-12pm, NN1)!
"Auditory and multisensory representations in the human hippocampus"

Using high-resolution fMRI, we ask how does the human hippocampus represent sensory modalities beyond vision, and how does it integrate sense information across modalities?

strikingly, our results reveal that similarity to prototypical narrative patterns predicts factual errors (distortions of actual content), while contextual surprise drives confabulations (entirely fabricated details).

back to our main question.

to identify narrative features that drive false memories, we computed three measures using autoregressive models and sentence embeddings at the level of individual narrative events (see below).

to validate the model, we ran two behavioral experiments in which humans provided ratings on a representative subset of recalls.

we found that human-AI agreement matched (or exceeded) inter-rater reliability in detecting errors.

collecting human ratings of false memories is time-consuming, expensive, and not feasible at this scale.

To address this, we developed an approach using LLMs to detect and classify errors in transcribed recollections. specifically, we use GPT-4o trained with in-context modeling: