Phantom transitions in language model fine-tuning

• URL: http://arxiv.org/abs/2606.07559v1
• Date: Mon, 25 May 2026 10:44:42 GMT
• Title: Phantom transitions in language model fine-tuning
• Abstract summary: Fine-tuning a language model on contexts with a near-synonym competitor often fails silently.<n>We study this regime across five transformer architectures spanning two families and a fivefold parameter range.<n>We observe sharp catapult-like jumps in the order parameter that resemble a phase transition.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Fine-tuning a language model on contexts whose correct completion has a near-synonym competitor often fails silently. The cross-entropy loss decreases monotonically while the correct token never overtakes the competitor in rank. We study this regime across five transformer architectures spanning two families and a fivefold parameter range, on ten hand-selected near-synonym contexts. We instrument these failures with an order parameter combining the predicted distribution and pairwise embedding overlaps. It decomposes additively into a signal, tracking 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. This isolates two failure modes. In kinematic failure the signal stays small. In structural failure the drag actively worsens as fine-tuning proceeds. We observe sharp catapult-like jumps in the order parameter that resemble a phase transition. A central negative result organises the paper. The transitions are phantoms. The spontaneous-symmetry-breaking interpretation is ruled out by direct measurement. Catapult-like jumps still appear under LoRA fine-tuning with the token embedding matrix exactly unchanged during training, where no geometric phase transition is possible. The discontinuity lives entirely in the softmax readout. A small number of dimensionless quantities organise the trajectory across architectures. 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. 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. Findings concern the near-synonym mechanism only and should not be extrapolated without recalibration.

Related papers

• Right Direction, Wrong Step: Geometric Analysis of Finite-Step Failure in Looped Transformers [40.43356191604634]
  Looped Transformers offer a parameter-efficient route to test-time scaling by reusing shared layers for iterative latent reasoning.<n>Additional iterations can reduce support for a reference answer, leaving unclear whether an update's direction is locally unhelpful or its full displacement moves too far.<n>We study this distinction by analysing reference utility, which measures this support, along the model's own update direction.
  arXiv  Detail & Related papers  (2026-09-15T05:42:02Z)
• Vision: Data-Centric Anchoring for Robust and Interpretable Agentic AI [21.406298170489507]
  We argue that robustness and interpretability should be engineered into the data environment, not extracted from models after training.<n>Our central contribution is the Data-Centric Agentic Loop, a four-stage framework of Curate, Augment, Constrain, and Attribute.<n>We ground the framework in a failure-driven taxonomy that links four core failure modes to the data lifecycle.
  arXiv  Detail & Related papers  (2026-09-08T03:57:54Z)
• A Calibration Audit of Confidence in Feed-Forward 3D Reconstruction [45.496856835431885]
  3D reconstruction models emit per-pixel confidence that downstream systems read as a reliability signal.<n>We score the confidence on four properties, how well it ranks error, whether its level is right on average, whether it holds across the confidence range, and whether its intervals cover the truth.<n>The confidence ranks error well, but the predicted uncertainty is too low when it is read under conditions that are not exactly those of training.
  arXiv  Detail & Related papers  (2026-08-30T10:21:34Z)
• Prompt-Model Interaction Reaches the Fixed Points: A deterministic, task-free structural readout -- and the factorizations of it that failed [0.0]
  We ask a readout with no task in it: the fixed-point structure of the short-window argmax map x_t+1 = argmax_x p(x | x_t-1, x_t), censused from 96 starts.<n>It is deterministic, so nothing can be helped or hurt, and it exists only at short windows -- four of six models lose it entirely by window 16 -- so everything here concerns how a model reads a fragment.<n>First, the interaction reaches this readout at full magnitude: nine tokens of conditioning move the fixed-point fraction across most of its
  arXiv  Detail & Related papers  (2026-08-21T17:25:18Z)
• What Iterated Self-Feeding Probes of Language Models Measure, and a test that separates the construction from the model [0.0]
  A class of methods probes a language model by feeding it its own output.<n>We ask what such a probe measures, in a construction chosen to make the question sharp.<n>The answer is that it measures two different things at once, in readings that look alike.
  arXiv  Detail & Related papers  (2026-08-11T14:40:35Z)
• Otap:Structure-Aware Optimal Transport for Evaluating Planning and Execution in Agent Trajectories [13.06173852232351]
  Current evaluation metrics reduce such a trajectory to a binary success flag or compare it against a reference by exact matching.<n>We instantiate trajectory evaluation as a distance between the agent's execution graph and a set of valid solution graphs.<n>The resulting score, termed otap, is a pseudo-metric that is provably invariant to dependency-preserving reorderings.
  arXiv  Detail & Related papers  (2026-07-19T05:23:17Z)
• DriftScope: Measuring The Hidden Effects of Diffusion Model Adaptation [50.584834483403675]
  Adapting pre-trained text-to-image diffusion models, whether to learn new visual concepts or erase unwanted ones, is routinely evaluated on its intended effects alone.<n>We demonstrate that adaptation systematically damages semantically unrelated concepts in ways that aggregate metrics structurally cannot surface.<n>We introduce DriftScope, a prompt-level diagnostic tool that takes any two model checkpoints and returns a ranked list of tokens whose visual concepts have shifted most between them.
  arXiv  Detail & Related papers  (2026-06-30T20:57:51Z)
• When Do Autoregressive Sequence Models Forecast Physical Wavefields? A Controlled Study on Synthetic Seismograms [43.8784307709823]
  Long-horizon autoregressive forecasting of physical signals is limited by error accumulation.<n>We ask when such rollout stays stable using synthetic three-component seismograms as a physically structured testbed.
  arXiv  Detail & Related papers  (2026-06-09T13:46:14Z)
• Hacking Generative Perplexity: Why Unconditional Text Evaluation Needs Distributional Metrics [49.443264461057645]
  Diffusion and continuous flow-based language models have emerged as the leading non-autoregressive alternatives to language modeling.<n>Progress in both paradigms is overwhelmingly tracked by generative perplexity (gen-PPL): the per-token negative log-likelihood of samples under a frozen autoregressive (AR) scorer.<n>We argue that this metric is unsound. By construction, gen-PPL measures only predictability under the scoring AR, not grammaticality or semantic coherence.
  arXiv  Detail & Related papers  (2026-06-07T02:35:56Z)
• Divergence is Uncertainty: A Closed-Form Posterior Covariance for Flow Matching [8.603039700922809]
  Flow matching has become a leading framework for generative modeling, but quantifying the uncertainty of its samples remains an open problem.<n>We show that none of these trade-offs is necessary.<n>Experiments on MNIST confirm that the resulting per-pixel uncertainty maps are semantically meaningful.
  arXiv  Detail & Related papers  (2026-05-01T04:25:00Z)
• Test-Time Scaling with Diffusion Language Models via Reward-Guided Stitching [66.39914384073145]
  We propose a self-consistency framework that turns cheap diffusion-sampled reasoning into a reusable pool of step-level candidates.<n>We find that step-level recombination is most beneficial on harder problems.<n>Our training-free framework improves average accuracy by up to 2 across six math and coding tasks.
  arXiv  Detail & Related papers  (2026-02-26T11:08:39Z)
• MERGETUNE: Continued fine-tuning of vision-language models [77.8627788911249]
  Fine-tuning vision-language models (VLMs) often leads to catastrophic forgetting of pretrained knowledge.<n>We introduce a novel paradigm, continued fine-tuning (CFT), which seeks to recover pretrained knowledge after a zero-shot model has already been adapted.
  arXiv  Detail & Related papers  (2026-01-15T15:15:53Z)
• Convergence Dynamics of Over-Parameterized Score Matching for a Single Gaussian [48.340460104014]
  We study gradient descent for training models to learn a single Gaussian distribution.<n>We prove a global convergence result for gradient descent under multiple regimes.<n>This is the first work to establish global convergence guarantees for Gaussian mixtures with at least three components under the score matching framework.
  arXiv  Detail & Related papers  (2025-11-27T03:41:48Z)
• The Remarkable Robustness of LLMs: Stages of Inference? [5.346230590800585]
  We investigate the robustness of Large Language Models (LLMs) to structural interventions by deleting and swapping adjacent layers during inference.<n>Surprisingly, models retain 72-95% of their original top-1 prediction accuracy without any fine-tuning.
  arXiv  Detail & Related papers  (2024-06-27T17:57:03Z)
• The distribution of syntactic dependency distances [0.13812010983144798]
  We contribute to the characterization of the actual distribution of syntactic dependency distances.<n>We propose a new model with two exponential regimes in which the probability decay is allowed to change after a break-point.<n>We find that a two-regime model is the most likely one in all 20 languages we considered, independently of sentence length and annotation style.
  arXiv  Detail & Related papers  (2022-11-26T17:31:25Z)
• Generative Text Modeling through Short Run Inference [47.73892773331617]
  The present work proposes a short run dynamics for inference. It is variation from the prior distribution of the latent variable and then runs a small number of Langevin dynamics steps guided by its posterior distribution. We show that the models trained with short run dynamics more accurately model the data, compared to strong language model and VAE baselines, and exhibit no sign of posterior collapse.
  arXiv  Detail & Related papers  (2021-05-27T09:14:35Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.