Reproduce “Slot Machines: How LLMs Keep Track of Multiple Entities” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Paul C. Bogdan, Jack Lindsey

The paper investigates whether one token position can carry information about two neighboring entities and whether the model can use both traces. Probes on Qwen3-32B activations find a current-entity slot and a distinct prior-entity slot, with four-slot accuracy 56.3% versus 29.7% for one slot. The prior trace supports relational retrieval, but not tested factual retrieval; shared-object questions remain difficult for many models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.