LLMs Exhibit Human-Like Temporal Order Memory via Time Reinstatement
A new study on arXiv investigates how long-context language models can replicate human episodic memory, particularly regarding the order of events. The researchers used a dataset that simulates how people remember a complete novel and found that these models exhibit a distance effect similar to that seen in humans. Their analysis of how the model works revealed that its performance relies on a one-dimensional temporal code, which is activated by a single attention head during retrieval. This suggests that language models might help us understand the computational aspects behind episodic memory.
Key facts
- Study published on arXiv with ID 2607.22575v1
- Investigates LLMs' ability to capture core behavioral signatures of episodic memory
- Uses a new dataset of human behavior based on memory of a full-length novel
- LLMs exhibit the same distance effect observed in humans
- Performance relies on a one-dimensional temporal code reinstated by a single attention head
- Research aims to uncover computational mechanisms of episodic memory
- Long-context LLMs used as models for human memory retrieval
- Mechanistic interpretability analysis applied to understand model behavior
Entities
Institutions
- arXiv