Phantom transitions in language model fine-tuning

Fine-tuning a language model on contexts with near-synonym competitors can fail silently, with the correct token never overtaking the competitor in rank even as the cross-entropy loss declines monotonically, according to a study posted to arXiv [1]. The study examined this regime across five transformer architectures spanning two families and a fivefold parameter range, using ten hand-selected near-synonym contexts [1]. The researchers instrumented these failures with an order parameter that combines the predicted distribution and pairwise embedding overlaps. This parameter decomposes additively into a signal, which tracks the model's commitment to the correct token over its nearest competitor, and a background drag, set by how the embedding bulk leaks probability into the score [1]. The decomposition isolates two failure modes: kinematic failure, where the signal stays small, and structural failure, where the drag actively worsens as fine-tuning proceeds [1]. The authors observed sharp, catapult-like jumps in the order parameter that resemble a phase transition. However, a central negative result organizes the paper: the transitions are phantoms. The spontaneous-symmetry-breaking interpretation is ruled out by direct measurement [1]. Catapult-like jumps still appear under LoRA fine-tuning with the token embedding matrix exactly unchanged during training, a condition where no geometric phase transition is possible. The discontinuity lives entirely in the softmax readout [1]. The work, posted on the preprint repository arXiv, arrives as large language models—defined as models with many parameters trained with self-supervised learning on vast text corpora—continue to underpin a wide range of AI systems [11]. The transformer architecture at the core of these models was introduced in the 2017 paper "Attention Is All You Need" and has since become the dominant design for large language models [9]. arXiv itself, launched in 1991, now receives about 24,000 submissions per month and serves as the primary distribution channel for machine-learning preprints [10]. A small number of dimensionless quantities organize the trajectory across architectures in the new study. One is consistent across all five under full fine-tuning. A second sorts architectures into two classes by bulk embedding distribution and predicts LoRA sufficiency [1]. As a blind test, the framework predicts the critical learning rate of a held-out architecture, not used to fit any parameter, to within 2.1% of a subsequent learning-rate sweep [1]. The authors caution that the findings concern the near-synonym mechanism only and should not be extrapolated without recalibration [1].