Local Normalization Distortion and the Thermodynamic Formalism of Decoding Strategies for Large Language Models
Advances in hardware and language model architecture have spurred a revolution in natural language generation. However, autoregressive models compute probability distributions over next-token choices, and sampling from these distributions, known as decoding, has received significantly less attention than other design choices. Existing decoding strategies are largely based on heuristics, resulting in methods that are hard to apply or improve in a principled manner. We develop the theory of decoding strategies for language models by expressing popular decoding algorithms as equilibrium states in the language of ergodic theory and stating the functions they optimize. Using this, we analyze the effect of the local normalization step of top-k, nucleus, and temperature sampling, used to make probabilities sum to one. We argue that local normalization distortion is a fundamental defect of decoding strategies and quantify the size of this distortion and its effect on mathematical proxies for the quality and diversity of generated text. Contrary to the prevailing explanation, we argue that the major cause of the under-performance of top-k sampling relative to nucleus sampling is local normalization distortion. This yields conclusions for the future design of decoding algorithms and the detection of machine-generated text.
Code (0)
등록된 구현이 없습니다.
Tasks
Text GenerationMethods 이 논문이 사용한 방법론
Similar Papers 제목 키워드 기반
Sentence Compression as Tree Transduction
This paper presents a tree-to-tree transduction method for sentence compression. Our model is based on synchronous tree substitution grammar, a formalism that allows local distortion of the tree topology and can thus nat…
Local DistortionSentenceSentence CompressionThermodynamic Formalism in Neuronal Dynamics and Spike Train Statistics
The Thermodynamic Formalism provides a rigorous mathematical framework to study quantitative and qualitative aspects of dynamical systems. At its core there is a variational principle corresponding, in its simplest form,…
A comparison of Single- and Double-generator formalisms for Thermodynamics-Informed Neural Networks
The development of inductive biases has been shown to be a very effective way to increase the accuracy and robustness of neural networks, particularly when they are used to predict physical phenomena. These biases signif…
validPort-metriplectic neural networks: thermodynamics-informed machine learning of complex physical systems
We develop inductive biases for the machine learning of complex physical systems based on the port-Hamiltonian formalism. To satisfy by construction the principles of thermodynamics in the learned physics (conservation o…
Local and Global Decoding in Text Generation
Text generation, a key component in applications such as dialogue systems, relies on decoding algorithms that sample strings from a language model distribution. Traditional methods, such as top-$k$ and top-$\pi$, apply l…
Language ModelingLanguage ModellingText Generation