A new arXiv preprint investigates how sampling temperature affects long-range correlations in autoregressive language generation. The authors note that temperature can substantially reshape output, but it remains unclear whether observed changes in correlations indicate a genuine phase transition or simply a smooth change in the underlying statistics.
The paper, titled "Power-Law Correlations Without Apparent Temperature-Tuned Criticality in Fixed-Tree Hierarchical Languages," appears to study this question in the context of fixed-tree hierarchical languages. The abstract raises the core distinction between critical phenomena and ordinary statistical variation, but the provided text cuts off mid-sentence, so the actual results and conclusions are not yet available from this source.
Because only the abstract's opening is available, this digest can report the motivating question but not the paper's findings. Readers should consult the full arXiv entry for the authors' analysis and evidence.