Reproduce “Learning State-Tracking from Code Using Linear RNNs” 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
Julien Siems, Riccardo Grazzi, Korbinian Pöppel, Kirill Kalinin, Hitesh Ballani, Babak Rahmani
The authors test state tracking in a next-token format by representing permutations as Python REPL traces with swaps and partial print reveals. They also analyze probabilistic transitions and partial observations through linear representations. DeltaNet with eigenvalues from −1 to 1 reaches 100% versus 25% for the restricted range in one setting, while adversarial reveals can destabilize representations in synthetic tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.