# Phantom Transitions in Language Model Fine-Tuning: A Density-Matrix Analysis

arXiv: https://arxiv.org/abs/2606.07559 · HTML: https://mlsanity.com/paper/2606.07559/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 28 of its 40 references that are in the map.
- First posted on arXiv 25 May 2026.

## Builds on (most cited first, 28 of 28)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://mlsanity.com/paper/2104.09864/) (2021)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [Parameter-Efficient Transfer Learning for NLP](https://mlsanity.com/paper/1902.00751/) (2019)
- [Prefix-Tuning: Optimizing Continuous Prompts for Generation](https://mlsanity.com/paper/2101.00190/) (2021)
- [The Power of Scale for Parameter-Efficient Prompt Tuning](https://mlsanity.com/paper/2104.08691/) (2021)
- [QLoRA: Efficient Finetuning of Quantized LLMs](https://mlsanity.com/paper/2305.14314/) (2023)
- [Emergent Abilities of Large Language Models](https://mlsanity.com/paper/2206.07682/) (2022)
- [Survey of Hallucination in Natural Language Generation](https://mlsanity.com/paper/2202.03629/) (2022)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [On Faithfulness and Factuality in Abstractive Summarization](https://mlsanity.com/paper/2005.00661/) (2020)
- [Progress measures for grokking via mechanistic interpretability](https://mlsanity.com/paper/2301.05217/) (2023)
- [How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings](https://mlsanity.com/paper/1909.00512/) (2019)
- [Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution](https://mlsanity.com/paper/2202.10054/) (2022)
- [Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning](https://mlsanity.com/paper/2012.13255/) (2020)
- [Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets](https://mlsanity.com/paper/2201.02177/) (2022)
- [Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping](https://mlsanity.com/paper/2002.06305/) (2020)
- [Using the Output Embedding to Improve Language Models](https://mlsanity.com/paper/1608.05859/) (2016)
- [Representation Degeneration Problem in Training Natural Language Generation Models](https://mlsanity.com/paper/1907.12009/) (2019)
- [On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines](https://mlsanity.com/paper/2006.04884/) (2020)
- [Breaking the Softmax Bottleneck: A High-Rank RNN Language Model](https://mlsanity.com/paper/1711.03953/) (2017)
- [A Statistical Physics of Language Model Reasoning](https://mlsanity.com/paper/2506.04374/) (2025)

## Abstract (arXiv)

Language models fine-tuned where the correct completion must outrank a near-synonym competitor often fail silently. The cross-entropy loss falls monotonically while the correct token never overtakes the competitor in the model's ranking. We study this across five transformer architectures from two families spanning a sixfold parameter range, on ten contexts whose correct and competing completions share substantial embedding overlap. We build an order parameter combining the predicted distribution with embedding overlap, as a density matrix because that distribution lives over a non-orthogonal basis. It decomposes additively into a signal term tracking commitment to the correct token and a drag term set by how the embedding bulk leaks probability into the score. This isolates two failure modes. In kinematic failure the signal stays too small and the model never commits. In structural failure the drag worsens during fine-tuning, so the model degrades geometrically as its loss falls. The order parameter also shows sharp jumps resembling phase transitions. We test the spontaneous-symmetry-breaking reading by tracking it after every gradient step, and rule it out. The jumps persist under LoRA even though the token embedding matrix never changes. No geometric phase transition is possible when that geometry cannot move, so the discontinuity lies entirely in the softmax readout. A few dimensionless quantities organize the trajectory across architectures. One is consistent across all five models under full fine-tuning. A second sorts architectures into two classes by their bulk embedding distribution and predicts whether LoRA alone can make a sentence commit. As a blind test, the framework predicts a held-out architecture's critical learning rate to within 2.1% of a later sweep. These results characterize this near-synonym mechanism and need recalibration before extrapolation.