Probabilistic Token Alignment for Large Language Model Fusion

• URL: http://arxiv.org/abs/2509.17276v1
• Date: Sun, 21 Sep 2025 23:18:24 GMT
• Title: Probabilistic Token Alignment for Large Language Model Fusion
• Authors: Runjia Zeng, James Chenhao Liang, Cheng Han, Zhiwen Cao, Jiahao Liu, Xiaojun Quan, Yingjie Victor Chen, Lifu Huang, Tong Geng, Qifan Wang, Dongfang Liu,
• Abstract summary: Training large language models (LLMs) from scratch can yield models with unique functionalities and strengths, but it is costly and often leads to redundant capabilities.<n>A key challenge in existing model fusion is their dependence on manually predefined vocabulary alignment.<n>We propose the probabilistic token alignment method as a general and soft mapping for alignment, named as PTA-LLM.
• Score: 100.30692772017238
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Training large language models (LLMs) from scratch can yield models with unique functionalities and strengths, but it is costly and often leads to redundant capabilities. A more cost-effective alternative is to fuse existing pre-trained LLMs with different architectures into a more powerful model. However, a key challenge in existing model fusion is their dependence on manually predefined vocabulary alignment, which may not generalize well across diverse contexts, leading to performance degradation in several evaluation. To solve this, we draw inspiration from distribution learning and propose the probabilistic token alignment method as a general and soft mapping for alignment, named as PTA-LLM. Our approach innovatively reformulates token alignment into a classic mathematical problem: optimal transport, seamlessly leveraging distribution-aware learning to facilitate more coherent model fusion. Apart from its inherent generality, PTA-LLM exhibits interpretability from a distributional perspective, offering insights into the essence of the token alignment. Empirical results demonstrate that probabilistic token alignment enhances the target model's performance across multiple capabilities. Our code is avaliable at https://runjia.tech/neurips_pta-llm/.

Related papers

• Every Step Counts: Decoding Trajectories as Authorship Fingerprints of dLLMs [63.82840470917859]
  We show that the decoding mechanism of dLLMs can be used as a powerful tool for model attribution.<n>We propose a novel information extraction scheme called the Directed Decoding Map (DDM), which captures structural relationships between decoding steps and better reveals model-specific behaviors.
  arXiv  Detail & Related papers  (2025-10-02T06:25:10Z)
• Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo [90.78001821963008]
  A wide range of LM applications require generating text that conforms to syntactic or semantic constraints.<n>We develop an architecture for controlled LM generation based on sequential Monte Carlo (SMC)<n>Our system builds on the framework of Lew et al. (2023) and integrates with its language model probabilistic programming language.
  arXiv  Detail & Related papers  (2025-04-17T17:49:40Z)
• Generative Modeling of Class Probability for Multi-Modal Representation Learning [7.5696616045063845]
  Multi-modal understanding plays a crucial role in artificial intelligence by enabling models to jointly interpret inputs from different modalities.<n>We propose a novel class anchor alignment approach that leverages class probability distributions for multi-modal representation learning.<n>Our method, Class-anchor-ALigned generative Modeling (CALM), encodes class anchors as prompts to generate and align class probability distributions for each modality.
  arXiv  Detail & Related papers  (2025-03-21T01:17:44Z)
• Promises and Pitfalls of Generative Masked Language Modeling: Theoretical Framework and Practical Guidelines [74.42485647685272]
  We focus on Generative Masked Language Models (GMLMs) We train a model to fit conditional probabilities of the data distribution via masking, which are subsequently used as inputs to a Markov Chain to draw samples from the model. We adapt the T5 model for iteratively-refined parallel decoding, achieving 2-3x speedup in machine translation with minimal sacrifice in quality.
  arXiv  Detail & Related papers  (2024-07-22T18:00:00Z)
• CharED: Character-wise Ensemble Decoding for Large Language Models [24.993790740335243]
  We present an inference-time ensembling algorithm aimed at "averaging" outputs from multiple large language models. Our proposed model is able to combine complimentary strengths of multiple LLMs, regardless of vocabulary, tokenization, or model size.
  arXiv  Detail & Related papers  (2024-06-25T22:35:07Z)
• Ensemble Learning for Heterogeneous Large Language Models with Deep Parallel Collaboration [39.35476224845088]
  Large language models (LLMs) exhibit complementary strengths in various tasks, motivating the research of LLM ensembling. We propose a training-free ensemble framework DeePEn, fusing the informative probability distributions yielded by different LLMs at each decoding step.
  arXiv  Detail & Related papers  (2024-04-19T08:52:22Z)
• Meaning Representations from Trajectories in Autoregressive Models [106.63181745054571]
  We propose to extract meaning representations from autoregressive language models by considering the distribution of all possible trajectories extending an input text. This strategy is prompt-free, does not require fine-tuning, and is applicable to any pre-trained autoregressive model. We empirically show that the representations obtained from large models align well with human annotations, outperform other zero-shot and prompt-free methods on semantic similarity tasks, and can be used to solve more complex entailment and containment tasks that standard embeddings cannot handle.
  arXiv  Detail & Related papers  (2023-10-23T04:35:58Z)
• Fusing Sentence Embeddings Into LSTM-based Autoregressive Language Models [20.24851041248274]
  We present an LSTM-based autoregressive language model which uses prefix embeddings (from a pretrained masked language model) via fusion. We find that fusion helps reliably in lowering the perplexity (16.74 $rightarrow$ 15.80), which is even preserved after a transfer to a dataset from a different domain. We also evaluate the best-performing fusion model by correlating its next word surprisal estimates with human reading times.
  arXiv  Detail & Related papers  (2022-08-04T02:13:03Z)
• AvgOut: A Simple Output-Probability Measure to Eliminate Dull Responses [97.50616524350123]
  We build dialogue models that are dynamically aware of what utterances or tokens are dull without any feature-engineering. The first model, MinAvgOut, directly maximizes the diversity score through the output distributions of each batch. The second model, Label Fine-Tuning (LFT), prepends to the source sequence a label continuously scaled by the diversity score to control the diversity level. The third model, RL, adopts Reinforcement Learning and treats the diversity score as a reward signal.
  arXiv  Detail & Related papers  (2020-01-15T18:32:06Z)

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.