Reproduce “Task Ecologies and the Evolution of World-Tracking Representations in Large 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
Giulio Valentino Dalla Riva
The authors relate next-token cross-entropy to the distinctions preserved by an encoding under training contexts. For a fixed encoding, excess loss is a weighted Jensen–Shannon divergence between merged states, vanishing when their next-token distributions agree in sampled contexts. In a small model, bracket-containing training raises Lisp balance discrimination from −0.002 to 0.78, but does not show SGD finds optimal encodings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.