October 10, 2025Open Access

Can an LLM Induce a Graph? Investigating Memory Drift and Context Length

View Full Paper

Ask AI

Bookmark

Authors

Discussion

Findings reveal shorter context lengths lead to memory drift in LLMs, suggesting critical evaluation benchmarks.

LLMs exhibit memory drift at shorter effective context lengths, undermining their performance on relational reasoning tasks.
When tasked with inducing graphs from noisy text, LLMs show significant limitations in handling structured knowledge.
Evaluation benchmarks may not accurately represent LLM capabilities in complex reasoning scenarios, necessitating updates.
Even reasoning-specialized models like OpenAI o1 are vulnerable to memory drift, highlighting architectural adaptation needs.

Yousuf et al. (2025) studied this question.

View Full Paper

Ask AI

Bookmark