Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining

• URL: http://arxiv.org/abs/2509.05291v1
• Date: Fri, 05 Sep 2025 17:56:24 GMT
• Title: Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining
• Authors: Deniz Bayazit, Aaron Mueller, Antoine Bosselut,
• Abstract summary: Large language models learn non-trivial abstractions during pretraining.<n>We use sparse crosscoders to discover and align features across model checkpoints.<n>We show that crosscoders can detect feature emergence, maintenance, and discontinuation during pretraining.
• Score: 33.22703101533052
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models (LLMs) learn non-trivial abstractions during pretraining, like detecting irregular plural noun subjects. However, it is not well understood when and how specific linguistic abilities emerge as traditional evaluation methods such as benchmarking fail to reveal how models acquire concepts and capabilities. To bridge this gap and better understand model training at the concept level, we use sparse crosscoders to discover and align features across model checkpoints. Using this approach, we track the evolution of linguistic features during pretraining. We train crosscoders between open-sourced checkpoint triplets with significant performance and representation shifts, and introduce a novel metric, Relative Indirect Effects (RelIE), to trace training stages at which individual features become causally important for task performance. We show that crosscoders can detect feature emergence, maintenance, and discontinuation during pretraining. Our approach is architecture-agnostic and scalable, offering a promising path toward more interpretable and fine-grained analysis of representation learning throughout pretraining.

Related papers

• Towards Understanding Multimodal Fine-Tuning: Spatial Features [25.349396112139214]
  Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model.<n>We present the first mechanistic analysis of VLM adaptation using stage-wise model diffing.
  arXiv  Detail & Related papers  (2026-02-06T18:48:18Z)
• Evolution of Concepts in Language Model Pre-Training [53.994470178155105]
  We track linear interpretable feature evolution across pre-training snapshots using a sparse dictionary learning method called crosscoders.<n>We find that most features begin to form around a specific point, while more complex patterns emerge in later training stages.
  arXiv  Detail & Related papers  (2025-09-21T18:53:12Z)
• Shortcut Learning Susceptibility in Vision Classifiers [11.599035626374409]
  Shortcut learning is where machine learning models exploit spurious correlations in data instead of capturing meaningful features.<n>This study introduces deliberate shortcuts into the dataset that are correlated with class labels both positionally and via intensity.<n>We evaluate susceptibility to shortcut learning across different learning rates.
  arXiv  Detail & Related papers  (2025-02-13T10:25:52Z)
• Pay Attention to Your Neighbours: Training-Free Open-Vocabulary Semantic Segmentation [19.20874993309959]
  vision-language foundation models, such as CLIP, have showcased remarkable effectiveness in numerous zero-shot image-level tasks. We propose a baseline for training-free OVSS, termed Neighbour-Aware CLIP (NACLIP) Our method enforces localization of patches in the self-attention of CLIP's vision transformer which, despite being crucial for dense prediction tasks, has been overlooked in the OVSS literature.
  arXiv  Detail & Related papers  (2024-04-12T01:08:04Z)
• The Fine Line: Navigating Large Language Model Pretraining with Down-streaming Capability Analysis [27.310894780313618]
  This paper undertakes a comprehensive comparison of model capabilities at various pretraining intermediate checkpoints. We confirm that specific downstream metrics exhibit similar training dynamics across models of different sizes. In addition to our core findings, we've reproduced Amber and OpenLLaMA, releasing their intermediate checkpoints.
  arXiv  Detail & Related papers  (2024-04-01T16:00:01Z)
• OmDet: Large-scale vision-language multi-dataset pre-training with multimodal detection network [17.980765138522322]
  This work introduces OmDet, a novel language-aware object detection architecture. Leveraging natural language as a universal knowledge representation, OmDet accumulates a "visual vocabulary" from diverse datasets. We demonstrate superior performance of OmDet over strong baselines in object detection in the wild, open-vocabulary detection, and phrase grounding.
  arXiv  Detail & Related papers  (2022-09-10T14:25:14Z)
• Bridging the Gap between Language Models and Cross-Lingual Sequence Labeling [101.74165219364264]
  Large-scale cross-lingual pre-trained language models (xPLMs) have shown effectiveness in cross-lingual sequence labeling tasks. Despite the great success, we draw an empirical observation that there is a training objective gap between pre-training and fine-tuning stages. In this paper, we first design a pre-training task tailored for xSL named Cross-lingual Language Informative Span Masking (CLISM) to eliminate the objective gap. Second, we present ContrAstive-Consistency Regularization (CACR), which utilizes contrastive learning to encourage the consistency between representations of input parallel
  arXiv  Detail & Related papers  (2022-04-11T15:55:20Z)
• A Simple Long-Tailed Recognition Baseline via Vision-Language Model [92.2866546058082]
  The visual world naturally exhibits a long-tailed distribution of open classes, which poses great challenges to modern visual systems. Recent advances in contrastive visual-language pretraining shed light on a new pathway for visual recognition. We propose BALLAD to leverage contrastive vision-language models for long-tailed recognition.
  arXiv  Detail & Related papers  (2021-11-29T17:49:24Z)
• Dense Contrastive Visual-Linguistic Pretraining [53.61233531733243]
  Several multimodal representation learning approaches have been proposed that jointly represent image and text. These approaches achieve superior performance by capturing high-level semantic information from large-scale multimodal pretraining. We propose unbiased Dense Contrastive Visual-Linguistic Pretraining to replace the region regression and classification with cross-modality region contrastive learning.
  arXiv  Detail & Related papers  (2021-09-24T07:20:13Z)
• Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models [65.19308052012858]
  Recent Transformer-based large-scale pre-trained models have revolutionized vision-and-language (V+L) research. We present VALUE, a set of meticulously designed probing tasks to decipher the inner workings of multimodal pre-training. Key observations: Pre-trained models exhibit a propensity for attending over text rather than images during inference.
  arXiv  Detail & Related papers  (2020-05-15T01:06:54Z)
• Pre-training Text Representations as Meta Learning [113.3361289756749]
  We introduce a learning algorithm which directly optimize model's ability to learn text representations for effective learning of downstream tasks. We show that there is an intrinsic connection between multi-task pre-training and model-agnostic meta-learning with a sequence of meta-train steps.
  arXiv  Detail & Related papers  (2020-04-12T09:05:47Z)

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.