Reproduce “A Fine-Grained Hierarchy and Benchmark for State Tracking in Language Models” 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
Jackson Petty, William Merrill, Ashish Sabharwal
The authors introduce StateBench, a synthetic benchmark that separates state tracking into four update tiers, from constant writes to arbitrary reference operations. They compare recurrent and Transformer architectures on sequences of up to 64,000 assignments, finding recurrent models generalize well on the first three tiers but less reliably on general-finite tasks; pretrained models struggle earlier. The benchmark is synthetic.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.