Demystifying Singular Defects in Large Language Models

• URL: http://arxiv.org/abs/2502.07004v1
• Date: Mon, 10 Feb 2025 20:09:16 GMT
• Title: Demystifying Singular Defects in Large Language Models
• Authors: Haoqi Wang, Tong Zhang, Mathieu Salzmann,
• Abstract summary: In large language models (LLMs), the underlying causes of high-norm tokens remain largely unexplored. We provide both theoretical insights and empirical validation across a range of recent models. We showcase two practical applications of these findings: the improvement of quantization schemes and the design of LLM signatures.
• Score: 61.98878352956125
• License:
• Abstract: Large transformer models are known to produce high-norm tokens. In vision transformers (ViTs), such tokens have been mathematically modeled through the singular vectors of the linear approximations of layers. However, in large language models (LLMs), the underlying causes of high-norm tokens remain largely unexplored, and their different properties from those of ViTs require a new analysis framework. In this paper, we provide both theoretical insights and empirical validation across a range of recent models, leading to the following observations: i) The layer-wise singular direction predicts the abrupt explosion of token norms in LLMs. ii) The negative eigenvalues of a layer explain its sudden decay. iii) The computational pathways leading to high-norm tokens differ between initial and noninitial tokens. iv) High-norm tokens are triggered by the right leading singular vector of the matrix approximating the corresponding modules. We showcase two practical applications of these findings: the improvement of quantization schemes and the design of LLM signatures. Our findings not only advance the understanding of singular defects in LLMs but also open new avenues for their application. We expect that this work will stimulate further research into the internal mechanisms of LLMs and will therefore publicly release our code.

Related papers

• Scalable Language Models with Posterior Inference of Latent Thought Vectors [52.63299874322121]
  Latent-Thought Language Models (LTMs) incorporate explicit latent thought vectors that follow an explicit prior model in latent space. LTMs possess additional scaling dimensions beyond traditional LLMs, yielding a structured design space. LTMs significantly outperform conventional autoregressive models and discrete diffusion models in validation perplexity and zero-shot language modeling.
  arXiv  Detail & Related papers  (2025-02-03T17:50:34Z)
• FIRP: Faster LLM inference via future intermediate representation prediction [54.897493351694195]
  FIRP generates multiple tokens instead of one at each decoding step. We conduct extensive experiments, showing a speedup ratio of 1.9x-3x in several models and datasets.
  arXiv  Detail & Related papers  (2024-10-27T15:53:49Z)
• Counting Ability of Large Language Models and Impact of Tokenization [17.53620419920189]
  We study the impact of tokenization on the counting abilities of large language models (LLMs) Our work investigates the impact of tokenization on the counting abilities of LLMs, uncovering substantial performance variations based on input tokenization differences.
  arXiv  Detail & Related papers  (2024-10-25T17:56:24Z)
• Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers [54.20763128054692]
  We study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model.
  arXiv  Detail & Related papers  (2024-09-09T18:10:26Z)
• LLMs are Not Just Next Token Predictors [0.0]
  LLMs are statistical models of language learning through gradient descent with a next token prediction objective. While LLMs are engineered using next token prediction, and trained based on their success at this task, our view is that a reduction to just next token predictor sells LLMs short. In order to draw this out, we will make an analogy with a once prominent research program in biology explaining evolution and development from the gene's eye view.
  arXiv  Detail & Related papers  (2024-08-06T16:36:28Z)
• Transformers need glasses! Information over-squashing in language tasks [18.81066657470662]
  We study how information propagates in decoder-only Transformers. We show that certain sequences of inputs to the Transformer can yield arbitrarily close representations in the final token. We also show that decoder-only Transformer language models can lose sensitivity to specific tokens in the input.
  arXiv  Detail & Related papers  (2024-06-06T17:14:44Z)
• Language Model Cascades: Token-level uncertainty and beyond [65.38515344964647]
  Recent advances in language models (LMs) have led to significant improvements in quality on complex NLP tasks. Cascading offers a simple strategy to achieve more favorable cost-quality tradeoffs. We show that incorporating token-level uncertainty through learned post-hoc deferral rules can significantly outperform simple aggregation strategies.
  arXiv  Detail & Related papers  (2024-04-15T21:02:48Z)
• A Theoretical Understanding of Shallow Vision Transformers: Learning, Generalization, and Sample Complexity [71.11795737362459]
  ViTs with self-attention modules have recently achieved great empirical success in many tasks. However, theoretical learning generalization analysis is mostly noisy and elusive. This paper provides the first theoretical analysis of a shallow ViT for a classification task.
  arXiv  Detail & Related papers  (2023-02-12T22:12:35Z)
• The Predictive Normalized Maximum Likelihood for Over-parameterized Linear Regression with Norm Constraint: Regret and Double Descent [12.929639356256928]
  We show that modern machine learning models do not obey a trade-off between the complexity of a prediction rule and its ability to generalize. We use the recently proposed predictive normalized maximum likelihood (pNML) which is the min-max regret solution for individual data. We demonstrate the use of the pNML regret as a point-wise learnability measure on synthetic data and that it can successfully predict the double-decent phenomenon.
  arXiv  Detail & Related papers  (2021-02-14T15:49:04Z)

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.