本文提出 Superposition in Transformers 架构，利用自动编码器、B 样条混合系数和自适应隐藏状态重建，在共享参数空间中叠加基础与微调模型，以缓解灾难性遗忘并支持推理时动态切换。
摘要：
灾难性遗忘仍是使大型语言模型（LLM）适应新任务或领域时的一项重大挑战。传统微调往往会覆盖已有知识，导致原始任务性能下降。我们提出 Superposition in Transformers，这是一种新颖架构，利用自动编码器在共享参数空间内叠加基础模型与微调模型的隐藏表示。通过使用基于 B 样条的混合系数，以及根据输入数据分布自适应重建隐藏状态的自动编码器，我们的方法有效缓解了灾难性遗忘，并实现了“模型内”叠加的新范式。该方法在保留原始模型能力的同时，允许添加紧凑的领域专门知识，并支持推理期间在不同模型状态之间动态切换。
英文摘要：
Catastrophic forgetting remains a major challenge when adapting large language models (LLMs) to new tasks or domains. Conventional fine-tuning often overwrites existing knowledge, causing performance degradation on original tasks. We introduce Superposition in Transformers, a novel architecture that leverages autoencoders to superimpose the hidden representations of a base model and a fine-tuned model within a shared parameter space. By using B-spline-based blending coefficients and autoencoders that adaptively reconstruct hidden states based on the input data distribution, our method effectively mitigates catastrophic forgetting and enables a new paradigm of "in-model" superposition. This approach preserves original model capabilities while allowing compact domain-specific expertise to be added, and it supports dynamic switching between model states during inference.
University of Amsterdam（阿姆斯特丹大学）；Xerox Research Centre Europe（施乐欧洲研究中心）
We introduce an LSTM-based method for dynamically integrating several word-prediction experts to obtain a conditional language model which can be good simultaneously at several subtasks. We illustrate this general approach with an application to dialogue where we integrate a neural chat model, good at conversational aspects, with a neural question-answering model, good at retrieving precise information from a knowledge-base, and show how the integration combines the strengths of the independent components. We hope that this focused contribution will attract attention on the benefits of using s…
Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；University of Chinese Academy of Sciences（中国科学院大学）
Simultaneous machine translation (SiMT) generates translation before reading the entire source sentence and hence it has to trade off between translation quality and latency. To fulfill the requirements of different translation quality and latency in practical applications, the previous methods usually need to train multiple SiMT models for different latency levels, resulting in large computational costs. In this paper, we propose a universal SiMT model with Mixture-of-Experts Wait-k Policy to achieve the best translation quality under arbitrary latency with only one trained model. Specificall…
Tsinghua University（清华大学）；THiFLY Research（THiFLY研究院）；State Key Laboratory of Cognitive Intelligence（认知智能国家重点实验室）
The table-based fact verification task has recently gained widespread attention and yet remains to be a very challenging problem. It inherently requires informative reasoning over natural language together with different numerical and logical reasoning on tables (e.g., count, superlative, comparative). Considering that, we exploit mixture-of-experts and present in this paper a new method: Self-adaptive Mixture-of-Experts Network (SaMoE). Specifically, we have developed a mixture-of-experts neural network to recognize and execute different types of reasoning -- the network is composed of multip…
DAG-MoE: From Simple Mixture to Structural Aggregation in Mixture-of-Experts
University of Science and Technology of China（中国科学技术大学）；Tsinghua University（清华大学）；Nanyang Technological University（南洋理工大学）；University of California, Berkeley（加州大学伯克利分校）；University of Washington（华盛顿大学）
Understanding Multilingualism in Mixture-of-Experts LLMs: Routing Mechanism, Expert Specialization, and Layerwise Steering
National University of Singapore（新加坡国立大学）；Beijing University of Posts and Telecommunication（北京邮电大学）；Harbin Institute of Technology（哈尔滨工业大学）；University of Science and Technology of China（中国科学技术大学）
Sparsely-activated Mixture-of-experts (MoE) models allow the number of parameters to greatly increase while keeping the amount of computation for a given token or a given sample unchanged. However, a poor expert routing strategy (e.g. one resulting in load imbalance) can cause certain experts to be under-trained, leading to an expert being under or over-specialized. Prior work allocates a fixed number of experts to each token using a top-k function regardless of the relative importance of different tokens. To address this, we propose a heterogeneous mixture-of-experts employing an expert choic…
MC#: Mixture Compressor for Mixture-of-Experts Large Models
Department of Electrical and Electronic Engineering, The University of Hong Kong（香港大学电子与电气工程系）；School of Computing, National University of Singapore（新加坡国立大学计算机学院）；NVIDIA Research（NVIDIA研究）
Harvard University（哈佛大学）；Microsoft Research（微软研究院）；Massachusetts General Hospital（麻省总医院）；Harvard Medical School（哈佛医学院）；University of Sienna（锡耶纳大学）
We propose Preferential MoE, a novel human-ML mixture-of-experts model that augments human expertise in decision making with a data-based classifier only when necessary for predictive performance. Our model exhibits an interpretable gating function that provides information on when human rules should be followed or avoided. The gating function is maximized for using human-based rules, and classification errors are minimized. We propose solving a coupled multi-objective problem with convex subproblems. We develop approximate algorithms and study their performance and convergence. Finally, we de…
Beihang University（北京航空航天大学）；The Institute of Acoustics of the Chinese Academy of Sciences（中国科学院声学研究所）；Microsoft（微软）
The sparse Mixture-of-Experts (MoE) model is powerful for large-scale pre-training and has achieved promising results due to its model capacity. However, with trillions of parameters, MoE is hard to be deployed on cloud or mobile environment. The inference of MoE requires expert parallelism, which is not hardware-friendly and communication expensive. Especially for resource-limited downstream tasks, such sparse structure has to sacrifice a lot of computing efficiency for limited performance gains. In this work, we observe most experts contribute scarcely little to the MoE fine-tuning and infer…
HyperMoE: Towards Better Mixture of Experts via Transferring Among Experts
Beijing University of Posts and Telecommunications（北京邮电大学）；Tsinghua University（清华大学）；INF Technology（INF科技）；Hong Kong University of Science and Technology（香港科技大学）
We develop a mixtures-of-experts (ME) approach to the multiclass classification where the predictors are univariate functions. It consists of a ME model in which both the gating network and the experts network are constructed upon multinomial logistic activation functions with functional inputs. We perform a regularized maximum likelihood estimation in which the coefficient functions enjoy interpretable sparsity constraints on targeted derivatives. We develop an EM-Lasso like algorithm to compute the regularized MLE and evaluate the proposed approach on simulated and real data.
Mixture of experts (MoE), introduced over 20 years ago, is the simplest gated modular neural network architecture. There is renewed interest in MoE because the conditional computation allows only parts of the network to be used during each inference, as was recently demonstrated in large scale natural language processing models. MoE is also of potential interest for continual learning, as experts may be reused for new tasks, and new experts introduced. The gate in the MoE architecture learns task decompositions and individual experts learn simpler functions appropriate to the gate's decomposit…
The Illusion of Specialization: Unveiling the Domain-Invariant "Standing Committee" in Mixture-of-Experts Models
The Fin AI（Fin AI）；Georgia Institute of Technology（佐治亚理工学院）；Cornell University（康奈尔大学）；Stevens Institute of Technology（史蒂文斯理工学院）；The University of Manchester（曼彻斯特大学）
