What Shapes Emergent Misalignment? Insights from Training Dynamics, Model Priors, and Data

• URL: http://arxiv.org/abs/2606.20814v1
• Date: Thu, 18 Jun 2026 18:04:20 GMT
• Title: What Shapes Emergent Misalignment? Insights from Training Dynamics, Model Priors, and Data
• Abstract summary: Emergent misalignment (EM) is a phenomenon in which models generalize with narrow fine-tuning.<n>We study EM and its variability directly through the components of fine-tuning: training dynamics, model priors, and data.
• Score: 20.335735139409245
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Emergent misalignment (EM) is a phenomenon in which models generalize with narrow fine-tuning, leading to broad (yet uneven) misalignment across evaluation questions. We study EM and its variability directly through the components of fine-tuning: training dynamics, model priors, and data. (1) We first explored how in-domain training loss relates to out-of-domain alignment scores across datasets and model families. Then, we tried to induce potential alternative local minima through different learning schedules for one narrow fine-tuning, but did not find strong runs with better broad alignment scores conditioned on similar or lower training loss. (2) We found that although the mean and standard deviations of the misaligned model scores are usually statistically different from those of the pre-trained model, there are some potential signals on overall positive correlation. The evaluation prompt-only activations from both the pre-trained and the original instruct models (prior to narrow fine-tuning) could predict fine-grained alignment scores after narrow fine-tuning. (3) Finally, we compared activation deltas before and after narrow fine-tuning and found moderate-to-high subspace overlap and similarity between the resulting activation shifts for training and evaluation prompts. Subspace overlaps between training and evaluation prompt activations correlate with their shifts' similarities when measuring with the last prompt-token activations. The train-evaluation data prompt overlap is controlled against overlap computed from random vectors and evaluation prompts activations.

Related papers

• Evaluating Federated Pre-Training: On the Reliability of Downstream Fine-Tuning and Intrinsic Evaluation [3.488204158798457]
  Federated pre-training offers a way to train foundation models on private or distributed data without centralizing the underlying datasets.<n>We study which evaluation protocol more reliably reflects federated pre-training quality.
  arXiv  Detail & Related papers  (2026-07-20T15:14:12Z)
• Online Bayesian Imbalanced Learning with Bregman-Calibrated Deep Networks [0.7106986689736825]
  We present textitOnline Bayesian Imbalanced Learning (OBIL), a principled framework that decouples likelihood-ratio estimation from class-prior assumptions.<n>Our approach builds on the established connection between Bregman divergences and proper scoring rules to show that deep networks trained with such losses produce posterior probability estimates.<n>We prove that these likelihood-ratio estimates remain valid under arbitrary changes in class priors and cost structures, requiring only a threshold adjustment for optimal Bayes decisions.
  arXiv  Detail & Related papers  (2026-02-08T21:23:00Z)
• Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts [1.9458156037869137]
  convergent learning is the degree to which neural systems arrive at similar internal representations.<n>We present a large-scale audit of convergent learning spanning dozens of vision models and thousands of layer-pair comparisons.<n>Findings fill critical gaps in our understanding of representational convergence, with implications for neuroscience and AI.
  arXiv  Detail & Related papers  (2025-02-26T00:04:24Z)
• Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks [91.15120211190519]
  This paper aims to understand the nature of noise in pre-training datasets and to mitigate its impact on downstream tasks. We propose a light-weight black-box tuning method (NMTune) to affine the feature space to mitigate the malignant effect of noise.
  arXiv  Detail & Related papers  (2023-09-29T06:18:15Z)
• What Happens During Finetuning of Vision Transformers: An Invariance Based Investigation [7.432224771219168]
  The pretrain-finetune paradigm usually improves downstream performance over training a model from scratch on the same task. In this work, we examine the relationship between pretrained vision transformers and the corresponding finetuned versions on several benchmark datasets and tasks.
  arXiv  Detail & Related papers  (2023-07-12T08:35:24Z)
• TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization [89.54947228958494]
  This paper focuses on the fine-tuning of an adversarially pre-trained model in various classification tasks. We propose a novel statistics-based approach, Two-WIng NormliSation (TWINS) fine-tuning framework. TWINS is shown to be effective on a wide range of image classification datasets in terms of both generalization and robustness.
  arXiv  Detail & Related papers  (2023-03-20T14:12:55Z)
• Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistent [97.64313409741614]
  We propose to enforce a emphconsistency property which states that predictions of the model on its own generated data are consistent across time. We show that our novel training objective yields state-of-the-art results for conditional and unconditional generation in CIFAR-10 and baseline improvements in AFHQ and FFHQ.
  arXiv  Detail & Related papers  (2023-02-17T18:45:04Z)
• Domain Adaptation with Adversarial Training on Penultimate Activations [82.9977759320565]
  Enhancing model prediction confidence on unlabeled target data is an important objective in Unsupervised Domain Adaptation (UDA) We show that this strategy is more efficient and better correlated with the objective of boosting prediction confidence than adversarial training on input images or intermediate features.
  arXiv  Detail & Related papers  (2022-08-26T19:50:46Z)
• Intersection of Parallels as an Early Stopping Criterion [64.8387564654474]
  We propose a method to spot an early stopping point in the training iterations without the need for a validation set. For a wide range of learning rates, our method, called Cosine-Distance Criterion (CDC), leads to better generalization on average than all the methods that we compare against.
  arXiv  Detail & Related papers  (2022-08-19T19:42:41Z)
• CAFA: Class-Aware Feature Alignment for Test-Time Adaptation [50.26963784271912]
  Test-time adaptation (TTA) aims to address this challenge by adapting a model to unlabeled data at test time. We propose a simple yet effective feature alignment loss, termed as Class-Aware Feature Alignment (CAFA), which simultaneously encourages a model to learn target representations in a class-discriminative manner.
  arXiv  Detail & Related papers  (2022-06-01T03:02:07Z)

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.