New preprint! Statistical structure skews object memory toward predictable successors. Model simulations show how this bias can arise from the backward expansion of hippocampal representations.
w/co-first @codydong.bsky.social , @marlietandoc.bsky.social & @annaschapiro.bsky.social osf.io/yuxb6_v1
In this review, we compare the properties of MA-LLMs to human episodic memory to identify ways to make MA-LLMs more human-like, such that they will be more effective as cognitive models. Aligning MA-LLMs with useful features of human memory may also help to advance AI. (5/n)
We’ve made progress in understanding how memory systems support real-world event comprehension. Yet we still lack computational models that generate precise predictions about how episodic memory (EM) will be used when processing naturalistic, high-dimensional stimuli. (2/n)
Memory-augmented LLMs (MA-LLMs) may help solve this problem. They combine the rich, context-sensitive semantic knowledge in LLM weights with an added memory system that can retrieve unique events, similar to human episodic memory. (4/n)
