zip2zip: Inference-Time Adaptive Vocabularies for Language Models via Token Compression

• URL: http://arxiv.org/abs/2506.01084v1
• Date: Sun, 01 Jun 2025 17:03:02 GMT
• Title: zip2zip: Inference-Time Adaptive Vocabularies for Language Models via Token Compression
• Abstract summary: zip2zip is a framework that enables large language models to dynamically adjust token vocabulary at inference time.<n>We show that an existing LLM can be zip2zip-fied in 10 GPU-hours via parameter-efficient finetuning.<n>The resulting zip2zip LLMs effectively learn to use hypertokens at inference time, reducing input and output sequence length by 20-60%.
• Score: 32.01058227175771
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Tokenization efficiency plays a critical role in the performance and cost of large language models (LLMs), yet most models rely on static tokenizers optimized for general-purpose corpora. These tokenizers' fixed vocabularies often fail to adapt to domain- or language-specific inputs, leading to longer token sequences and higher computational costs. We introduce zip2zip, a framework that enables LLMs to dynamically adjust token vocabulary at inference time, allowing for fewer generated tokens and thus faster inference. zip2zip consists of three key components: (1) a tokenizer based on Lempel-Ziv-Welch (LZW) compression that incrementally compresses tokens into reusable "hypertokens" on the fly; (2) an embedding layer that computes embeddings for newly formed hypertokens at runtime; and (3) a causal language modeling variant that trains the model to operate on hypertokenized, compressed sequences. We show that an existing LLM can be zip2zip-fied in 10 GPU-hours via parameter-efficient finetuning. The resulting zip2zip LLMs effectively learn to use hypertokens at inference time, reducing input and output sequence length by 20-60\%, with significant improvements in inference latency.

Related papers

• SimSD: Simple Speculative Decoding in Diffusion Language Models [61.33773959352141]
  Diffusion large language models (dLLMs) offer faster inference through parallel or blockwise decoding.<n>Their masked language modeling formulation remains incompatible with standard token-level speculative decoding.<n>We propose a speculative decoding algorithm for dLLMs, named SimSD, that equips dLLMs with temporally valid token-level contexts.<n>Our method achieves up to 7.46x higher decoding throughput while maintaining and even improving average generation quality.
  arXiv  Detail & Related papers  (2026-06-01T17:46:46Z)
• Meta-Soft: Leveraging Composable Meta-Tokens for Context-Preserving KV Cache Compression [12.589754555104768]
  KV cache used in large language models has linearly growing time complexity.<n>Current KV Cache eviction methods rely on a static parameter set as the query to evaluate the importance of KV pairs.<n>We propose Meta-Soft, a dynamic compression framework based on probe-driven context integration.
  arXiv  Detail & Related papers  (2026-05-21T11:24:46Z)
• Large Language Model as Token Compressor and Decompressor [13.847845643427119]
  We show that an off-the-shelf LLM can function as an excellent token compressor and decompressor.<n>We use a self-expressive autoencoding learning framework to translate long texts into a compact internal language of discrete, variable-length latent codes, termed Z-tokens.
  arXiv  Detail & Related papers  (2026-03-26T11:30:44Z)
• LiteToken: Removing Intermediate Merge Residues From BPE Tokenizers [76.59130257385826]
  Intermediate merge residues in BPE vocabularies are frequent during merge learning so that retained in the final vocabulary, but are mostly further merged and rarely emitted when tokenizing the corpus during tokenizer usage.<n>We present a systematic empirical characterization of this phenomenon across commonly used tokenizers and introduce LiteToken, a simple method for removing residue tokens.<n>Experiments show that LiteToken reduces token fragmentation, reduces parameters, and improves robustness to noisy or misspelled inputs, while preserving overall performance.
  arXiv  Detail & Related papers  (2026-02-04T16:19:05Z)
• Continuous Autoregressive Language Models [56.49239051750678]
  We introduce Continuous Autoregressive Language Models (CALM)<n>CALM uses a high-fidelity autoencoder to compress a chunk of K tokens into a single continuous vector.<n>We develop a comprehensive likelihood-free framework that enables robust training, evaluation, and controllable sampling.
  arXiv  Detail & Related papers  (2025-10-31T17:58:11Z)
• Learning to Parallel: Accelerating Diffusion Large Language Models via Learnable Parallel Decoding [21.609237262034636]
  Autoregressive decoding in large language models (LLMs) requires $mathcalO(n)$ sequential steps for $n$ tokens.<n>We propose Learning to Parallel Decode (Learn2PD), a framework that trains a lightweight and adaptive filter model to predict, for each token position, whether the current prediction matches the final output.<n>This learned filter approximates an oracle parallel decoding strategy that unmasks tokens only when correctly predicted.
  arXiv  Detail & Related papers  (2025-09-29T17:59:54Z)
• Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning [23.58934174168992]
  Autoregressive (AR) language models generate text one token at a time, which limits their inference speed.<n>We propose Convolutional decoding (Conv), a normalization-based method that narrows the decoding window without hard segmentation.<n>We also introduce Rejecting Rule-based Fine-Tuning (R2FT), a post-hoc training scheme that better aligns tokens at positions far from context.
  arXiv  Detail & Related papers  (2025-09-18T17:48:21Z)
• Accelerating Diffusion LLMs via Adaptive Parallel Decoding [60.407727995313074]
  We introduce adaptive parallel decoding (APD), a novel method that dynamically adjusts the number of tokens sampled in parallel.<n>APD provides markedly higher throughput with minimal quality degradations on downstream benchmarks.
  arXiv  Detail & Related papers  (2025-05-31T06:10:10Z)
• Vision-centric Token Compression in Large Language Model [51.92055188780033]
  Vision Centric Token Compression (Vist) is a slow-fast compression framework that mirrors human reading.<n>On eleven in-context learning benchmarks, Vist achieves the same accuracy with 2.3 times fewer tokens, cutting FLOPs by 16% and memory by 50%.
  arXiv  Detail & Related papers  (2025-02-02T13:10:06Z)
• Retrofitting Large Language Models with Dynamic Tokenization [3.608780819053423]
  We propose retrofitting current language models with dynamic tokenization.<n>We merge frequent subword sequences in a batch, then apply a pre-trained embedding-prediction hypernetwork to compute the token embeddings on-the-fly.<n>We find that dynamic tokenization can mitigate the limitations of static tokenization by substantially improving inference speed and promoting fairness across languages.
  arXiv  Detail & Related papers  (2024-11-27T17:51:58Z)
• Inference Optimal VLMs Need Fewer Visual Tokens and More Parameters [54.01228554126122]
  Vision Language Models (VLMs) have demonstrated strong capabilities across various visual understanding and reasoning tasks.<n>To reduce inference costs, one can either downsize the Large Language Models (LLMs) or reduce the number of input tokens needed to represent the image.<n>We take the first steps toward designing token compression algorithms tailored for high-compression settings.
  arXiv  Detail & Related papers  (2024-11-05T18:54:21Z)
• MultiTok: Variable-Length Tokenization for Efficient LLMs Adapted from LZW Compression [5.5795785998430185]
  MultiTok is a new tokenization method inspired by universal Lempel-Ziv-Welch data compression.<n>We show that MultiTok achieves a comparable performance to the BERT and GPT-2 standards as both a stand-alone tokenizer and an add-on to existing tokenizers.
  arXiv  Detail & Related papers  (2024-10-28T21:24:51Z)
• LazyLLM: Dynamic Token Pruning for Efficient Long Context LLM Inference [30.722379261991563]
  LazyLLM is a method that selectively computes the KV for tokens important for the next token prediction. We show that LazyLLM accelerates the prefilling stage of the LLama 2 7B model by 2.34x while maintaining accuracy.
  arXiv  Detail & Related papers  (2024-07-19T06:34:45Z)
• A Training-free Sub-quadratic Cost Transformer Model Serving Framework With Hierarchically Pruned Attention [43.211427581302715]
  We propose Hierarchically Pruned Attention (HiP) to increase context length in large language models.<n>HiP reduces the time complexity of the attention mechanism to $O(T log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length.<n>We show that HiP significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation.
  arXiv  Detail & Related papers  (2024-06-14T08:32:45Z)
• Training LLMs over Neurally Compressed Text [55.11828645767342]
  This paper explores the idea of training large language models (LLMs) over highly compressed text.<n>We propose Equal-Info Windows, a novel compression technique whereby text is segmented into blocks that each compress to the same bit length.<n>We demonstrate effective learning over neurally compressed text that improves with scale, and outperforms byte-level baselines by a wide margin on perplexity and inference speed benchmarks.
  arXiv  Detail & Related papers  (2024-04-04T17:48:28Z)
• Extreme Compression of Large Language Models via Additive Quantization [59.3122859349777]
  Our algorithm, called AQLM, generalizes the classic Additive Quantization (AQ) approach for information retrieval. We provide fast GPU and CPU implementations of AQLM for token generation, which enable us to match or outperform optimized FP16 implementations for speed.
  arXiv  Detail & Related papers  (2024-01-11T18:54:44Z)
• Context Compression for Auto-regressive Transformers with Sentinel Tokens [37.07722536907739]
  We propose a plug-and-play approach that is able to incrementally compress the intermediate activation of a specified span of tokens into compact ones. Experiments on both in-domain language modeling and zero-shot open-ended document generation demonstrate the advantage of our approach.
  arXiv  Detail & Related papers  (2023-10-12T09:18:19Z)
• Efficient Streaming Language Models with Attention Sinks [72.20260088848987]
  StreamingLLM is an efficient framework that enables Large Language Models to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more.
  arXiv  Detail & Related papers  (2023-09-29T17:59:56Z)
• ZipLM: Inference-Aware Structured Pruning of Language Models [56.52030193434863]
  We propose a novel structured compression approach for large language models (LLMs) called ZipLM. ZipLM achieves state-of-the-art accuracy-vs-speedup, while matching a set of desired target runtime speedups. ZipLM produces state-of-the-art compressed models across all settings.
  arXiv  Detail & Related papers  (2023-02-07T18:55:28Z)

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.