# Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding

arXiv: https://arxiv.org/abs/2503.10135 · HTML: https://mlsanity.com/paper/2503.10135/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- Placed in Speculative decoding, one of 12 directions in Efficient LLM inference and serving.
- Builds on 39 of its 43 references that are in the map.
- First posted on arXiv 13 Mar 2025.

## Placed in

- Efficient LLM inference and serving / [Speculative decoding](https://mlsanity.com/inference/speculative-decoding/)

## Builds on (most cited first, 30 of 39)

- [GPT-4 Technical Report](https://mlsanity.com/paper/2303.08774/) (2023)
- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Evaluating Large Language Models Trained on Code](https://mlsanity.com/paper/2107.03374/) (2021)
- [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://mlsanity.com/paper/2312.00752/) (2023)
- [Fast Inference from Transformers via Speculative Decoding](https://mlsanity.com/paper/2211.17192/) (2022)
- [Accelerating Large Language Model Decoding with Speculative Sampling](https://mlsanity.com/paper/2302.01318/) (2023)
- [Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned](https://mlsanity.com/paper/1905.09418/) (2019)
- [Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads](https://mlsanity.com/paper/2401.10774/) (2024)
- [Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyond](https://mlsanity.com/paper/1602.06023/) (2016)
- [EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty](https://mlsanity.com/paper/2401.15077/) (2024)
- [The State of Sparsity in Deep Neural Networks](https://mlsanity.com/paper/1902.09574/) (2019)
- [Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations](https://mlsanity.com/paper/1609.07061/) (2016)
- [Movement Pruning: Adaptive Sparsity by Fine-Tuning](https://mlsanity.com/paper/2005.07683/) (2020)
- [EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees](https://mlsanity.com/paper/2406.16858/) (2024)
- [Better & Faster Large Language Models via Multi-token Prediction](https://mlsanity.com/paper/2404.19737/) (2024)
- [Q-BERT: Hessian Based Ultra Low Precision Quantization of BERT](https://mlsanity.com/paper/1909.05840/) (2019)
- [LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding](https://mlsanity.com/paper/2404.16710/) (2024)
- [Q8BERT: Quantized 8Bit BERT](https://mlsanity.com/paper/1910.06188/) (2019)
- [Break the Sequential Dependency of LLM Inference Using Lookahead Decoding](https://mlsanity.com/paper/2402.02057/) (2024)
- [Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding](https://mlsanity.com/paper/2401.07851/) (2024)
- [I-BERT: Integer-only BERT Quantization](https://mlsanity.com/paper/2101.01321/) (2021)
- [Speculative Decoding: Exploiting Speculative Execution for Accelerating Seq2seq Generation](https://mlsanity.com/paper/2203.16487/) (2022)
- [Lite Transformer with Long-Short Range Attention](https://mlsanity.com/paper/2004.11886/) (2020)
- [Hydra: Sequentially-Dependent Draft Heads for Medusa Decoding](https://mlsanity.com/paper/2402.05109/) (2024)
- [The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models](https://mlsanity.com/paper/2203.07259/) (2022)
- [GOBO: Quantizing Attention-Based NLP Models for Low Latency and Energy Efficient Inference](https://mlsanity.com/paper/2005.03842/) (2020)
- [GliDe with a CaPE: A Low-Hassle Method to Accelerate Speculative Decoding](https://mlsanity.com/paper/2402.02082/) (2024)

## Built on by (most cited first, 9 of 9)

- [LogitSpec: Accelerating Retrieval-based Speculative Decoding via Next Next Token Speculation](https://mlsanity.com/paper/2507.01449/) (2025)
- [ECHO: Elastic Speculative Decoding with Sparse Gating for High-Concurrency Scenarios](https://mlsanity.com/paper/2604.09603/) (2026)
- [Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match](https://mlsanity.com/paper/2511.22972/) (2025)
- [DynaSpec: Context-aware Dynamic Speculative Sampling for Large-Vocabulary Language Models](https://mlsanity.com/paper/2510.13847/) (2025)
- [PocketLLM: Ultimate Compression of Large Language Models via Meta Networks](https://mlsanity.com/paper/2511.17637/) (2025)
- [Please refuse to answer me! Mitigating Over-Refusal in Large Language Models via Adaptive Contrastive Decoding](https://mlsanity.com/paper/2604.17132/) (2026)
- [P-MTP: Efficient Document Parsing via Multi-Token Prediction with Progressive Depth Scaling](https://mlsanity.com/paper/2606.24447/) (2026)
- [NanoSpec: Accelerating Speculative Decoding using Minimalist In-Context Vocabularies](https://mlsanity.com/paper/2605.26444/) (2026)
- [Learnable Permutation for Structured Sparsity on Transformer Models](https://mlsanity.com/paper/2601.22980/) (2026)

## Abstract (arXiv)

Speculative decoding (SPD) aims to accelerate the auto-regressive token generation process of a target Large Language Model (LLM). Some approaches employ a draft model with multiple heads to predict a sequence of future tokens, where each head handles a token in the sequence. The target LLM verifies the predicted sequence and accepts aligned tokens, enabling efficient multi-token generation. However, existing methods assume that all tokens within a sequence are equally important, employing identical head structures and relying on a single-generation paradigm, either serial or parallel. To this end, we theoretically demonstrate that initial tokens in the draft sequence are more important than later ones. Building on this insight, we propose Gumiho, a hybrid model combining serial and parallel heads. Specifically, given the critical importance of early tokens, we employ a sophisticated Transformer architecture for the early draft heads in a serial configuration to improve accuracy. For later tokens, we utilize multiple lightweight MLP heads operating in parallel to enhance efficiency. By allocating more advanced model structures and longer running times to the early heads, Gumiho achieves improved overall performance. The experimental results demonstrate that our method outperforms existing approaches, fully validating its effectiveness.