Reproduce “The Dynamics of Continuous Mixture Collapse 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
Ali Backour
The paper studies how continuous mixtures of token embeddings are preserved through transformer processing and autoregressive generation. Controlled probes compare trained models with matched random networks, while a two-token analysis models softmax feedback and its coupling threshold. On a 1,000-item benchmark, Qwen3.5-4B has final-layer CALIB 0.332 versus 0.732 ± 0.003 for controls, measuring mixture preservation rather than reasoning ability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.