Disentangling Geometry, Performance, and Training in Language Models

• URL: http://arxiv.org/abs/2602.20433v1
• Date: Tue, 24 Feb 2026 00:31:04 GMT
• Title: Disentangling Geometry, Performance, and Training in Language Models
• Authors: Atharva Kulkarni, Jacob Mitchell Springer, Arjun Subramonian, Swabha Swayamdipta,
• Abstract summary: We systematically investigate the relationship between model performance and the unembedding matrix geometry.<n>Our experiments involve a suite of 108 OLMo-style language models trained under controlled variation.<n>While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups.
• Score: 28.748060518731446
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model's performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model's geometry, as captured by existing metrics, primarily reflects training choices rather than performance.

Related papers

• Weight Decay Improves Language Model Plasticity [9.005013915262658]
  We study pretraining from the perspective of model plasticity, that is, the ability of the base model to adapt to downstream tasks.<n>We show that models trained with larger weight decay values are more plastic, meaning they show larger performance gains when fine-tuned on downstream tasks.
  arXiv  Detail & Related papers  (2026-02-11T18:49:26Z)
• Fisher Information, Training and Bias in Fourier Regression Models [0.05997422707234518]
  We study the interplay between the empheffective dimension and the emphbias of a model towards a given task.<n>We show that for a model that is completely agnostic, or unbiased, towards the function to be learned, a higher effective dimension likely results in a better trainability and performance.
  arXiv  Detail & Related papers  (2025-10-08T12:29:11Z)
• Tending Towards Stability: Convergence Challenges in Small Language Models [3.734405405403176]
  Despite their advantages, smaller models frequently underperform compared to their larger counterparts. This is anecdotally attributed to their reduced representational capacity. By linking the convergence of layers' activations to their parameters' effective rank, our analyses can guide future work to address inefficiencies in the learning dynamics of small models.
  arXiv  Detail & Related papers  (2024-10-15T09:57:19Z)
• TRAWL: Tensor Reduced and Approximated Weights for Large Language Models [11.064868044313855]
  We introduce TRAWL (Tensor Reduced and Approximated Weights for Large Language Models), a technique that applies tensor decomposition across multiple weight matrices to effectively denoise LLMs by capturing global structural patterns.<n>Our experiments show that TRAWL improves model performance by up to 16% over baseline models on benchmark datasets, without requiring additional data, training, or fine-tuning.
  arXiv  Detail & Related papers  (2024-06-25T04:01:32Z)
• Observational Scaling Laws and the Predictability of Language Model Performance [51.2336010244645]
  We propose an observational approach that bypasses model training and instead builds scaling laws from 100 publically available models. We show that several emergent phenomena follow a smooth, sigmoidal behavior and are predictable from small models. We show how to predict the impact of post-training interventions like Chain-of-Thought and Self-Consistency as language model capabilities continue to improve.
  arXiv  Detail & Related papers  (2024-05-17T17:49:44Z)
• Theoretical Characterization of the Generalization Performance of Overfitted Meta-Learning [70.52689048213398]
  This paper studies the performance of overfitted meta-learning under a linear regression model with Gaussian features. We find new and interesting properties that do not exist in single-task linear regression. Our analysis suggests that benign overfitting is more significant and easier to observe when the noise and the diversity/fluctuation of the ground truth of each training task are large.
  arXiv  Detail & Related papers  (2023-04-09T20:36:13Z)
• Training Trajectories of Language Models Across Scales [99.38721327771208]
  Scaling up language models has led to unprecedented performance gains. How do language models of different sizes learn during pre-training? Why do larger language models demonstrate more desirable behaviors?
  arXiv  Detail & Related papers  (2022-12-19T19:16:29Z)
• Investigating Ensemble Methods for Model Robustness Improvement of Text Classifiers [66.36045164286854]
  We analyze a set of existing bias features and demonstrate there is no single model that works best for all the cases. By choosing an appropriate bias model, we can obtain a better robustness result than baselines with a more sophisticated model design.
  arXiv  Detail & Related papers  (2022-10-28T17:52:10Z)
• Benign Overfitting in Classification: Provably Counter Label Noise with Larger Models [8.696962915720174]
  We show that a ResNet model overfits benignly on Cifar10 but not benignly on ImageNet. Our work highlights the importance of understanding implicit bias in underfitting regimes as a future direction.
  arXiv  Detail & Related papers  (2022-06-01T14:00:37Z)
• No One Representation to Rule Them All: Overlapping Features of Training Methods [12.58238785151714]
  High-performing models tend to make similar predictions regardless of training methodology. Recent work has made very different training techniques, such as large-scale contrastive learning, yield competitively-high accuracy. We show these models specialize in generalization of the data, leading to higher ensemble performance.
  arXiv  Detail & Related papers  (2021-10-20T21:29:49Z)
• Powerpropagation: A sparsity inducing weight reparameterisation [65.85142037667065]
  We introduce Powerpropagation, a new weight- parameterisation for neural networks that leads to inherently sparse models. Models trained in this manner exhibit similar performance, but have a distribution with markedly higher density at zero, allowing more parameters to be pruned safely. Here, we combine Powerpropagation with a traditional weight-pruning technique as well as recent state-of-the-art sparse-to-sparse algorithms, showing superior performance on the ImageNet benchmark.
  arXiv  Detail & Related papers  (2021-10-01T10:03:57Z)
• Models, Pixels, and Rewards: Evaluating Design Trade-offs in Visual Model-Based Reinforcement Learning [109.74041512359476]
  We study a number of design decisions for the predictive model in visual MBRL algorithms. We find that a range of design decisions that are often considered crucial, such as the use of latent spaces, have little effect on task performance. We show how this phenomenon is related to exploration and how some of the lower-scoring models on standard benchmarks will perform the same as the best-performing models when trained on the same training data.
  arXiv  Detail & Related papers  (2020-12-08T18:03:21Z)

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.