Large Language Model Alignment with Brain Representations during Language Production

Poster Session A - Saturday, April 13, 2024, 2:30 – 4:30 pm EDT, Sheraton Hall ABC

Roger Beaty¹, John Patterson¹, Elisa Kwon², Kosa Goucher-Lambert²; ¹Pennsylvania State University, ²University of California, Berkeley

Recent work has demonstrated that representations from Large Language Models (LLMs) accurately align with human neural responses for the same linguistic inputs. This alignment suggests far-reaching uses for LLMs in both neuroengineering and basic research, if sufficiently general. However, brain-LLM alignment has scarcely been tested beyond language reception; the circumstances under which LLMs align with human neural representations during language production, and under different task goals, remains an open question. In this work, we explore how LLMs align with human neural representations on a language production task under two distinct goals, within-subject. To investigate the impact of different goals, human subjects were asked to produce word associations that were either appropriate (bench - sit) or novel (coin - veil) in response to cue nouns (bench/coin) during fMRI. To relate brain representations to model-based ones, we obtained language representations from Llama-2, a contemporary 7B-parameter LLM that is architecturally similar to, but improved over, leading models of human language reception (e.g. GPT-2). We then assessed how task goals shape brain-LLM alignment during language production using representational similarity analysis. Specifically, we compared the representational geometries from the brain and LLM that resulted from each goal. We hypothesized the appropriate goal would produce greater brain-LLM alignment relative to the novel goal—as common associates are more likely to be present in training corpora—but our findings suggest a more nuanced relationship. We conclude by discussing future directions and implications for neuroscience and neuroengineering.

Topic Area: LANGUAGE: Other