Disentangling Geometry, Performance, and Training in Language Models

2026-02-24

AI Summary

This study investigates whether universal or causal relationships exist between the geometric properties—particularly the effective rank—of language model output embedding matrices and downstream performance. By systematically training 108 OLMo-style models under rigorously controlled conditions, the work presents the first large-scale experimental disentanglement of geometric characteristics, training hyperparameters, and performance. The findings reveal that while high effective rank often coincides with strong performance, this association is not universal; low effective rank co-occurs with performance saturation but is not its cause. Instead, geometric metrics primarily reflect training configurations—such as batch size and weight decay—and prove unreliable as standalone predictors of model performance. These results challenge the prevailing hypothesis that low effective rank directly causes performance saturation, instead positioning geometric properties as epiphenomena of the training process.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• unembedding matrix
• effective rank
• language model performance
• model geometry
• training hyperparameters

Innovation

• effective rank
• unembedding matrix
• language model geometry
• training hyperparameters
• performance disentanglement