id
string
sources
list
title
string
abstract
string
authors
list
categories
list
fields_of_study
list
published_date
timestamp[s]
url
string
pdf_url
string
arxiv_id
string
doi
string
citation_count
int64
influential_citation_count
int64
has_code
bool
code_url
string
venue
string
quality_score
float64
2637d50b286df8bbee11e4a10a0a06661adf804b34461dba537faed835ea5169
[ "arxiv", "semantic_scholar" ]
EAGLE-Pangu: Accelerator-Safe Tree Speculative Decoding on Ascend NPUs
Autoregressive decoding remains a primary bottleneck in large language model (LLM) serving, motivating speculative decoding methods that reduce expensive teacher-model invocations by verifying multiple candidate tokens per step. Tree-structured speculation further increases parallelism, but is often brittle when ported...
[ "Chang Han", "Yijie Hu", "Jingling Liu" ]
[ "cs.LG", "cs.PL" ]
[ "Computer Science" ]
2026-03-09T00:00:00
https://arxiv.org/abs/2603.08088
https://arxiv.org/pdf/2603.08088v1
2603.08088
10.48550/arXiv.2603.08088
0
0
false
null
arXiv.org
0.5019
abaf72d9dcb8237f170e98af3b8e9308c2e965d7c8234d464348ebbfec8e115b
[ "arxiv", "semantic_scholar" ]
Speculating Experts Accelerates Inference for Mixture-of-Experts
Mixture-of-Experts (MoE) models have gained popularity as a means of scaling the capacity of large language models (LLMs) while maintaining sparse activations and reduced per-token compute. However, in memory-constrained inference settings, expert weights must be offloaded to CPU, creating a performance bottleneck from...
[ "Vivan Madan", "Prajwal Singhania", "Abhinav Bhatele", "Tom Goldstein", "Ashwinee Panda" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2026-03-09T00:00:00
https://arxiv.org/abs/2603.19289
https://arxiv.org/pdf/2603.19289v1
2603.19289
10.48550/arXiv.2603.19289
0
0
true
https://github.com/axonn-ai/yalis/tree/offload_prefetch
arXiv.org
0.7756
940d2378569b310049174fb9a710aa7b4bf69280e958af18c8a3bc4f4319269d
[ "arxiv", "semantic_scholar" ]
ConFu: Contemplate the Future for Better Speculative Sampling
Speculative decoding has emerged as a powerful approach to accelerate large language model (LLM) inference by employing lightweight draft models to propose candidate tokens that are subsequently verified by the target model. The effectiveness of this paradigm critically depends on the quality of the draft model. While ...
[ "Zongyue Qin", "Raghavv Goel", "Mukul Gagrani", "Risheek Garrepalli", "Mingu Lee", "Yizhou Sun" ]
[ "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2026-03-09T00:00:00
https://arxiv.org/abs/2603.08899
https://arxiv.org/pdf/2603.08899v3
2603.08899
10.48550/arXiv.2603.08899
0
0
false
null
arXiv.org
0.5019
85ab3f800c4a10936be523e43141f048f35767e726ca2642a30d4b8cd256376b
[ "arxiv", "semantic_scholar" ]
Balancing Coverage and Draft Latency in Vocabulary Trimming for Faster Speculative Decoding
Speculative decoding accelerates inference for Large Language Models by using a lightweight draft model to propose candidate tokens that are verified in parallel by a larger target model. Prior work shows that the draft model often dominates speculative decoding latency, since it generates tokens sequentially and incur...
[ "Ofir Ben Shoham" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2026-03-05T00:00:00
https://arxiv.org/abs/2603.05210
https://arxiv.org/pdf/2603.05210v1
2603.05210
10.48550/arXiv.2603.05210
1
0
false
null
arXiv.org
0.4973
9965bae6110495f0a86782639cc6e909e3b5d6df537f588e4dc5832c4fc4fc5d
[ "arxiv", "semantic_scholar" ]
SLO-Aware Compute Resource Allocation for Prefill-Decode Disaggregated LLM Inference
Prefill-Decode (P/D) disaggregation has emerged as a widely adopted optimization strategy for Large Language Model (LLM) inference. However, there currently exists no well-established methodology for determining the optimal number of P/D hardware resources, subject to constraints on total throughput, service level obje...
[ "Luchang Li", "Dongfang Li", "Bozhao Gong", "Yu Zhang" ]
[ "cs.DC", "cs.IT", "cs.LG" ]
[ "Computer Science", "Mathematics" ]
2026-03-05T00:00:00
https://arxiv.org/abs/2603.04716
https://arxiv.org/pdf/2603.04716v1
2603.04716
10.48550/arXiv.2603.04716
0
0
false
null
arXiv.org
0.4973
9f9e27c8313a7425ad13a0442ab2169c45f549374fe6e4b1aa88ce20b0ed42e7
[ "arxiv", "semantic_scholar" ]
Speculative Speculative Decoding
Autoregressive decoding is bottlenecked by its sequential nature. Speculative decoding has become a standard way to accelerate inference by using a fast draft model to predict upcoming tokens from a slower target model, and then verifying them in parallel with a single target model forward pass. However, speculative de...
[ "Tanishq Kumar", "Tri Dao", "Avner May" ]
[ "cs.LG" ]
[ "Computer Science" ]
2026-03-03T00:00:00
https://arxiv.org/abs/2603.03251
https://arxiv.org/pdf/2603.03251v3
2603.03251
10.48550/arXiv.2603.03251
8
0
true
null
arXiv.org
0.765
c88adb913b99bb02f7b5e4f49fd9a2d7173101b964065942f0a52a2bf37500dd
[ "arxiv", "semantic_scholar" ]
Accelerating OpenPangu Inference on NPU via Speculative Decoding
To mitigate the Memory Wall bottleneck encountered by Large Language Models (LLMs) during inference on \textbf{NPU} hardware, and addressing the scarcity of native support for mainstream speculative decoding algorithms on domestic infrastructure, this study presents an end-to-end speculative inference acceleration sche...
[ "Yuntao Dai", "Jing Wu", "Hang Gu", "Teng Wang" ]
[ "cs.DC" ]
[ "Computer Science" ]
2026-03-03T00:00:00
https://arxiv.org/abs/2603.03383
https://arxiv.org/pdf/2603.03383v1
2603.03383
10.48550/arXiv.2603.03383
0
0
false
null
arXiv.org
0.495
d8e14c6dab3bff9aa086aa6ad24ffafbf19d0230d1470b6a7f4d7c8a10e82bfc
[ "arxiv", "semantic_scholar" ]
Learning to Draft: Adaptive Speculative Decoding with Reinforcement Learning
Speculative decoding accelerates large language model (LLM) inference by using a small draft model to generate candidate tokens for a larger target model to verify. The efficacy of this technique hinges on the trade-off between the time spent on drafting candidates and verifying them. However, current state-of-the-art ...
[ "Jiebin Zhang", "Zhenghan Yu", "Liang Wang", "Nan Yang", "Eugene J. Yu", "Zheng Li", "Yifan Song", "Dawei Zhu", "Xingxing Zhang", "Furu Wei", "Sujian Li" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-03-02T00:00:00
https://arxiv.org/abs/2603.01639
https://arxiv.org/pdf/2603.01639v1
2603.01639
10.48550/arXiv.2603.01639
0
0
false
null
arXiv.org
0.4939
69cfaa5a10370e697fe383985d3b7d207f7a18f1057f0789847a2a3d3cfd7cd4
[ "arxiv", "semantic_scholar" ]
Quasar: Quantized Self-Speculative Acceleration for Rapid Inference via Memory-Efficient Verification
Speculative Decoding (SD) has emerged as a premier technique for accelerating Large Language Model (LLM) inference by decoupling token generation into rapid drafting and parallel verification. While recent advancements in self-speculation and lookahead decoding have successfully minimized drafting overhead, they have s...
[ "Guang Huang", "Zeyi Wen" ]
[ "cs.DC", "cs.LG" ]
[ "Computer Science" ]
2026-03-02T00:00:00
https://arxiv.org/abs/2603.01399
https://arxiv.org/pdf/2603.01399v1
2603.01399
10.48550/arXiv.2603.01399
1
0
true
https://github.com/Tom-HG/Quasar
arXiv.org
0.7632
3fc5790597f01c3d465898874dbabfe2e0af582688e4cb6ae667c8ccf46d118a
[ "arxiv", "semantic_scholar" ]
KERV: Kinematic-Rectified Speculative Decoding for Embodied VLA Models
Vision-Language-Action (VLA) models build a token-domain robot control paradigm, yet suffer from low speed. Speculative Decoding (SD) is an optimization strategy that can boost inference speed. Two key issues emerge when integrating VLA and SD: first, SD relies on re-inference to address token errors, which is computat...
[ "Zihao Zheng", "Zhihao Mao", "Maoliang Li", "Jiayu Chen", "Xinhao Sun", "Zhaobo Zhang", "Donggang Cao", "Hong Mei", "Xiang Chen" ]
[ "cs.RO", "cs.LG" ]
[ "Computer Science" ]
2026-03-02T00:00:00
https://arxiv.org/abs/2603.01581
https://arxiv.org/pdf/2603.01581v2
2603.01581
10.48550/arXiv.2603.01581
6
0
false
null
arXiv.org
0.4939
3c52ce4295805bbf03880da0a8385ba7b764d0c84ceaa0df7970243c6137230c
[ "arxiv", "semantic_scholar" ]
SJD-PV: Speculative Jacobi Decoding with Phrase Verification for Autoregressive Image Generation
Autoregressive (AR) image models have recently demonstrated remarkable generative capability, but their sequential nature results in significant inference latency. Existing training-free acceleration methods typically verify tokens independently, overlooking the strong co-occurrence patterns between adjacent visual tok...
[ "Zhehao Yu", "Baoquan Zhang", "Bingqi Shan", "Xinhao Liu", "Dongliang Zhou", "Guotao Liang", "Guangming Ye", "Yunming Ye" ]
[ "cs.CV" ]
[ "Computer Science" ]
2026-03-02T00:00:00
https://arxiv.org/abs/2603.06666
https://arxiv.org/pdf/2603.06666v1
2603.06666
10.48550/arXiv.2603.06666
0
0
false
null
arXiv.org
0.4939
0f9321eafacdebba715e3ab1b3c36c9ef8568ec76d554dec4501d2d6ac5d1092
[ "arxiv", "semantic_scholar" ]
LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding
Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to propose candidate tokens that are then verified in parallel by the target model. The speedup is significantly determined by the acceptance rate, yet standard training minimizes Kullback-Leibler (KL...
[ "Alexander Samarin", "Sergei Krutikov", "Anton Shevtsov", "Sergei Skvortsov", "Filipp Fisin", "Alexander Golubev" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2026-02-27T00:00:00
https://arxiv.org/abs/2602.23881
https://arxiv.org/pdf/2602.23881v2
2602.23881
10.48550/arXiv.2602.23881
5
0
false
null
arXiv.org
0.4904
a1ec30d62dc75806d0aa475f0ef3030088d533404888c920379eaa673c8aa51d
[ "arxiv", "semantic_scholar" ]
Speculative Decoding Scaling Laws (SDSL): Throughput Optimization Made Simple
Speculative decoding is a technique that uses multiple language models to accelerate infer- ence. Previous works have used an experi- mental approach to optimize the throughput of the inference pipeline, which involves LLM training and can be costly. This study of spec- ulative decoding proposes a theory that ana- lyti...
[ "Amirhossein Bozorgkhoo", "Igor Molybog" ]
[ "cs.CL", "cs.IT", "cs.LG" ]
[ "Computer Science", "Mathematics" ]
2026-02-25T00:00:00
https://arxiv.org/abs/2603.11053
https://arxiv.org/pdf/2603.11053v1
2603.11053
10.48550/arXiv.2603.11053
0
0
false
null
arXiv.org
0.4881
614457d2f5b19b2841174f379475807c919b6cd0de49039533020870a4263f20
[ "arxiv", "semantic_scholar" ]
MineDraft: A Framework for Batch Parallel Speculative Decoding
Speculative decoding (SD) accelerates large language model inference by using a smaller draft model to propose draft tokens that are subsequently verified by a larger target model. However, the performance of standard SD is often limited by the strictly sequential execution of these drafting and verification stages. To...
[ "Zhenwei Tang", "Arun Verma", "Zijian Zhou", "Zhaoxuan Wu", "Alok Prakash", "Daniela Rus", "Bryan Kian Hsiang Low" ]
[ "cs.CL", "cs.AI", "cs.DC", "cs.LG" ]
[ "Computer Science" ]
2026-02-24T00:00:00
https://arxiv.org/abs/2603.18016
https://arxiv.org/pdf/2603.18016v2
2603.18016
10.48550/arXiv.2603.18016
2
0
false
null
arXiv.org
0.487
3a07305ce7f891698355f95b58fc4c94e26624805574d9a871b4fa5f098db6cd
[ "arxiv", "semantic_scholar" ]
Greedy Multi-Path Block Verification for Faster Decoding in Speculative Sampling
The goal of $L$-step speculative decoding is to accelerate autoregressive decoding of a target model by using a cheaper draft model to generate a candidate path of $L$ tokens. Based on a verification algorithm involving target and draft model probabilities, a prefix of the candidate sequence is accepted, and an additio...
[ "Rahul Thomas", "Arka Pal" ]
[ "cs.IT", "cs.LG" ]
[ "Computer Science", "Mathematics" ]
2026-02-18T00:00:00
https://arxiv.org/abs/2602.16961
https://arxiv.org/pdf/2602.16961v1
2602.16961
10.48550/arXiv.2602.16961
0
0
false
null
arXiv.org
0.4801
7ff93673f3d0de80cf9a5d268e83d2cfd323d63015bd65be696aa45af8421926
[ "arxiv", "semantic_scholar" ]
Privacy-Aware Split Inference with Speculative Decoding for Large Language Models over Wide-Area Networks
We present a practical system for privacy-aware large language model (LLM) inference that splits a transformer between a trusted local GPU and an untrusted cloud GPU, communicating only intermediate activations over the network. Our system addresses the unique challenges of autoregressive LLM decoding over high-latency...
[ "Michael Cunningham" ]
[ "cs.CR", "cs.DC" ]
[ "Computer Science" ]
2026-02-18T00:00:00
https://arxiv.org/abs/2602.16760
https://arxiv.org/pdf/2602.16760v1
2602.16760
10.48550/arXiv.2602.16760
0
0
false
null
arXiv.org
0.4801
f8028e71a8ea62bb19e7f8fc6c259b427149351e6c3097da0bd5378774eca8b0
[ "arxiv", "semantic_scholar" ]
MoE-Spec: Expert Budgeting for Efficient Speculative Decoding
Speculative decoding accelerates Large Language Model (LLM) inference by verifying multiple drafted tokens in parallel. However, for Mixture-of-Experts (MoE) models, this parallelism introduces a severe bottleneck: large draft trees activate many unique experts, significantly increasing memory pressure and diminishing ...
[ "Bradley McDanel", "Steven Li", "Sruthikesh Surineni", "Harshit Khaitan" ]
[ "cs.LG" ]
[ "Computer Science" ]
2026-02-17T00:00:00
https://arxiv.org/abs/2602.16052
https://arxiv.org/pdf/2602.16052v1
2602.16052
10.48550/arXiv.2602.16052
3
0
false
null
arXiv.org
0.479
dae9fa5aba7ff05350cfa8cd1cf97257baf7db1b7dfd4ea5f2a7e23946ab3084
[ "arxiv", "semantic_scholar" ]
Sparrow: Text-Anchored Window Attention with Visual-Semantic Glimpsing for Speculative Decoding in Video LLMs
Although speculative decoding is widely used to accelerate Vision-Language Models (VLMs) inference, it faces severe performance collapse when applied to Video Large Language Models (Vid-LLMs). The draft model typically falls into the trap of attention dilution and negative visual gain due to key-value cache explosion a...
[ "Libo Zhang", "Zhaoning Zhang", "Wangyang Hong", "Peng Qiao", "Dongsheng Li" ]
[ "cs.CV", "cs.AI" ]
[ "Computer Science" ]
2026-02-17T00:00:00
https://arxiv.org/abs/2602.15318
https://arxiv.org/pdf/2602.15318v1
2602.15318
10.48550/arXiv.2602.15318
3
0
false
null
arXiv.org
0.479
d0581e9dc5f171405f2f0f7344445b8c654db77136e29c8ed6317ec5451b93e4
[ "arxiv", "semantic_scholar" ]
Speculative Decoding with a Speculative Vocabulary
Speculative decoding has rapidly emerged as a leading approach for accelerating language model (LM) inference, as it offers substantial speedups while yielding identical outputs. This relies upon a small draft model, tasked with predicting the outputs of the target model. State-of-the-art speculative decoding methods u...
[ "Miles Williams", "Young D. Kwon", "Rui Li", "Alexandros Kouris", "Stylianos I. Venieris" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-02-14T00:00:00
https://arxiv.org/abs/2602.13836
https://arxiv.org/pdf/2602.13836v1
2602.13836
10.48550/arXiv.2602.13836
2
1
false
null
arXiv.org
0.4755
065d91c968177c81423660a305f9818dde4293c39bac8fe49ed564c09c76f506
[ "arxiv", "semantic_scholar" ]
MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenarios
Mixture-of-Experts (MoE) models enable scalable performance but face severe memory constraints on edge devices. Existing offloading strategies struggle with I/O bottlenecks due to the dynamic, low-information nature of autoregressive expert activation. In this paper, we propose to repurpose Speculative Decoding (SD) no...
[ "Shuhuai Li", "Jianghao Lin", "Dongdong Ge", "Yinyu Ye" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2026-02-12T00:00:00
https://arxiv.org/abs/2603.09983
https://arxiv.org/pdf/2603.09983v1
2603.09983
10.48550/arXiv.2603.09983
1
0
true
https://github.com/lshAlgorithm/MoE-SpAc
arXiv.org
0.7314
07d03f57bc1ae6cdd9f65c6895083a47037543009da6a609916b37ce9f736c39
[ "arxiv", "semantic_scholar" ]
Training-free Dropout Sampling for Semantic Token Acceptance in Speculative Decoding
Speculative decoding accelerates large language model inference by proposing tokens with a lightweight draft model and selectively accepting them using a target model. This work introduces DropMatch, a novel approach that matches draft tokens to the predictive distribution of the target model via Monte Carlo dropout ap...
[ "Jeongtae Lee", "Minjung Jo", "Hyunjoon Jeong", "Gunho Park", "Sunghyeon Woo", "Joonghoon Kim", "Se Jung Kwon", "Dongsoo Lee" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-02-11T00:00:00
https://arxiv.org/abs/2603.03333
https://arxiv.org/pdf/2603.03333v1
2603.03333
10.48550/arXiv.2603.03333
0
0
false
null
arXiv.org
0.4721
44af42eb5fe0a5d159f7cdf5e9750863cf446e41b32b72954094fcb96ff61429
[ "arxiv", "semantic_scholar" ]
StreamServe: Adaptive Speculative Flows for Low-Latency Disaggregated LLM Serving
Efficient LLM serving must balance throughput and latency across diverse, bursty workloads. We introduce StreamServe, a disaggregated prefill decode serving architecture that combines metric aware routing across compute lanes with adaptive speculative decoding that tunes speculation depth online from runtime signals. S...
[ "Satyam Kumar", "Arpit Singh Gautam", "Kailash Talreja", "Saurabh Jha" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2026-02-11T00:00:00
https://arxiv.org/abs/2604.09562
https://arxiv.org/pdf/2604.09562v1
2604.09562
10.48550/arXiv.2604.09562
0
0
false
null
arXiv.org
0.4721
f41985dc055efb96a1291d4e6a3cced3615908215c8532e7e515d27ae5eaad48
[ "arxiv", "semantic_scholar" ]
SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding
Speculative Decoding (SD) has emerged as a critical technique for accelerating Large Language Model (LLM) inference. Unlike deterministic system optimizations, SD performance is inherently data-dependent, meaning that diverse and representative workloads are essential for accurately measuring its effectiveness. Existin...
[ "Talor Abramovich", "Maor Ashkenazi", "Izzy Putterman", "Benjamin Chislett", "Tiyasa Mitra", "Bita Darvish Rouhani", "Ran Zilberstein", "Yonatan Geifman" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2026-02-10T00:00:00
https://arxiv.org/abs/2604.09557
https://arxiv.org/pdf/2604.09557v2
2604.09557
10.48550/arXiv.2604.09557
2
1
false
null
arXiv.org
0.4709
91d76cd55b98f485a8d9d828dc0284cb4bc1d2f4892450aff3fc42d234b84fbe
[ "arxiv", "semantic_scholar" ]
Benchmarking the Energy Savings with Speculative Decoding Strategies
Speculative decoding has emerged as an effective method to reduce latency and inference cost of LLM inferences. However, there has been inadequate attention towards the energy requirements of these models. To address this gap, this paper presents a comprehensive survey of energy requirements of speculative decoding str...
[ "Rohit Dutta", "Paramita Koley", "Soham Poddar", "Janardan Misra", "Sanjay Podder", "Naveen Balani", "Saptarshi Ghosh", "Niloy Ganguly" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2026-02-09T00:00:00
https://arxiv.org/abs/2602.09113
https://arxiv.org/pdf/2602.09113v1
2602.09113
10.48550/arXiv.2602.09113
0
0
false
null
Conference of the European Chapter of the Association for Computational Linguistics
0.4698
754971f7a2e1001cff3b1343c3a7c0299b9876bd6e6f05f2f92eea981d717fb1
[ "arxiv", "semantic_scholar" ]
Compiler-Assisted Speculative Sampling for Accelerated LLM Inference on Heterogeneous Edge Devices
LLM deployment on resource-constrained edge devices faces severe latency constraints, particularly in real-time applications where delayed responses can compromise safety or usability. Among many approaches to mitigate the inefficiencies of sequential token-by-token generation, Speculative Decoding (SD) has emerged as ...
[ "Alejandro Ruiz y Mesa", "Guilherme Korol", "Moritz Riesterer", "João Paulo Cardoso de Lima", "Jeronimo Castrillon" ]
[ "cs.LG" ]
[ "Computer Science" ]
2026-02-08T00:00:00
https://arxiv.org/abs/2602.08060
https://arxiv.org/pdf/2602.08060v2
2602.08060
10.48550/arXiv.2602.08060
0
0
false
null
arXiv.org
0.4686
5e54c53680c7efd6fbb43b4542d5b82996063a9b9f92b6f1b4aba13739fa942f
[ "arxiv", "semantic_scholar" ]
Vegas: Self-Speculative Decoding with Verification-Guided Sparse Attention
Long-context large language model (LLM) inference has become the norm for today's AI applications. However, it is severely bottlenecked by the increasing memory demands of its KV cache. Previous works have shown that self-speculative decoding with sparse attention, where tokens are drafted using a subset of the KV cach...
[ "Yikang Yue", "Yuqi Xue", "Jian Huang" ]
[ "cs.LG" ]
[ "Computer Science" ]
2026-02-06T00:00:00
https://arxiv.org/abs/2602.07223
https://arxiv.org/pdf/2602.07223v2
2602.07223
null
0
0
true
https://github.com/platformxlab/vegas
null
0.5511
092017e2097c57adce599a86e9bb7123673ae9cc3667d737dfd01c581e146d1b
[ "arxiv", "semantic_scholar" ]
Anchored Decoding: Provably Reducing Copyright Risk for Any Language Model
Language models (LMs) tend to memorize portions of their training data and emit verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play in...
[ "Jacqueline He", "Jonathan Hayase", "Wen-tau Yih", "Sewoong Oh", "Luke Zettlemoyer", "Pang Wei Koh" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-02-06T00:00:00
https://arxiv.org/abs/2602.07120
https://arxiv.org/pdf/2602.07120v2
2602.07120
10.48550/arXiv.2602.07120
0
0
true
https://github.com/jacqueline-he/anchored-decoding
arXiv.org
0.7207
fa131ba8f9cc9b5f3c200aadceabdbf6675f23d692788be68da85b430825700d
[ "arxiv", "semantic_scholar" ]
DFlash: Block Diffusion for Flash Speculative Decoding
Autoregressive large language models (LLMs) deliver strong performance but require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM; however, exi...
[ "Jian Chen", "Yesheng Liang", "Zhijian Liu" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-02-05T00:00:00
https://arxiv.org/abs/2602.06036
https://arxiv.org/pdf/2602.06036v2
2602.06036
10.48550/arXiv.2602.06036
33
14
true
https://github.com/z-lab/dflash
arXiv.org
0.719
efd442573bfe236afe3518764162863d47ee0a03543753f6fe9a95538d2c2342
[ "arxiv", "semantic_scholar" ]
SDFP: Speculative Decoding with FIT-Pruned Models for Training-Free and Plug-and-Play LLM Acceleration
Large language models (LLMs) underpin interactive multimedia applications such as captioning, retrieval, recommendation, and creative content generation, yet their autoregressive decoding incurs substantial latency. Speculative decoding reduces latency using a lightweight draft model, but deployment is often limited by...
[ "Hanyu Wei", "Zunhai Su", "Peng Lu", "Chao Li", "Spandan Tiwari", "Ashish Sirasao", "Yuhan Dong" ]
[ "cs.AI" ]
[ "Computer Science" ]
2026-02-05T00:00:00
https://arxiv.org/abs/2602.05499
https://arxiv.org/pdf/2602.05499v1
2602.05499
10.48550/arXiv.2602.05499
0
0
false
null
arXiv.org
0.4652
b4513f6f0d82b39b39430b8c68903a284d0eddfbd3884947b569a90266aeeb4a
[ "arxiv", "semantic_scholar" ]
Variational Speculative Decoding: Rethinking Draft Training from Token Likelihood to Sequence Acceptance
Speculative decoding accelerates inference for (M)LLMs, yet a training-decoding discrepancy persists: while existing methods optimize single greedy trajectories, decoding involves verifying and ranking multiple sampled draft paths. We propose Variational Speculative Decoding (VSD), formulating draft training as variati...
[ "Xiandong Zou", "Jianshu Li", "Jing Huang", "Pan Zhou" ]
[ "cs.LG", "cs.AI", "math.PR" ]
[ "Computer Science", "Mathematics" ]
2026-02-05T00:00:00
https://arxiv.org/abs/2602.05774
https://arxiv.org/pdf/2602.05774v4
2602.05774
null
0
0
false
null
null
0.296
1f5da4d1db6c0e6343c3dd91b420c3358309f3bca24067be9d30d3d1bbaad33f
[ "arxiv", "semantic_scholar" ]
TIDE: Temporal Incremental Draft Engine for Self-Improving LLM Inference
Speculative decoding can substantially accelerate LLM inference, but realizing its benefits in practice is challenging due to evolving workloads and system-level constraints. We present TIDE (Temporal Incremental Draft Engine), a serving-engine-native framework that integrates online draft adaptation directly into high...
[ "Jiyoung Park", "Hankyu Jang", "Changseok Song", "Wookeun Jung" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2026-02-05T00:00:00
https://arxiv.org/abs/2602.05145
https://arxiv.org/pdf/2602.05145v1
2602.05145
10.48550/arXiv.2602.05145
1
0
false
null
arXiv.org
0.4652
b0f2c8b875e126495561622b98c904e33b28b54775b4984e682f473f252452aa
[ "arxiv", "semantic_scholar" ]
LycheeDecode: Accelerating Long-Context LLM Inference via Hybrid-Head Sparse Decoding
The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sha...
[ "Gang Lin", "Dongfang Li", "Zhuoen Chen", "Yukun Shi", "Xuhui Chen", "Baotian Hu", "Min Zhang" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2026-02-04T00:00:00
https://arxiv.org/abs/2602.04541
https://arxiv.org/pdf/2602.04541v1
2602.04541
10.48550/arXiv.2602.04541
3
0
false
null
arXiv.org
0.4641
56ab42c904e0b27e7cd5a547cc0b176f6ea976631f9580d89704d71bd3bfd8dc
[ "arxiv", "semantic_scholar" ]
Beyond Tokens: Semantic-Aware Speculative Decoding for Efficient Inference by Probing Internal States
Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying...
[ "Ximing Dong", "Shaowei Wang", "Dayi Lin", "Boyuan Chen", "Ahmed E. Hassan" ]
[ "cs.CL", "cs.PF" ]
[ "Computer Science" ]
2026-02-03T00:00:00
https://arxiv.org/abs/2602.03708
https://arxiv.org/pdf/2602.03708v2
2602.03708
10.48550/arXiv.2602.03708
0
0
false
null
arXiv.org
0.4629
61cc828d9569c50d6d74243c810300b60afc6d63d8278674a18a2478358f09ac
[ "arxiv", "semantic_scholar" ]
Make Every Draft Count: Hidden State based Speculative Decoding
Speculative decoding has emerged as a pivotal technique to accelerate LLM inference by employing a lightweight draft model to generate candidate tokens that are subsequently verified by the target model in parallel. However, while this paradigm successfully increases the arithmetic intensity of memory-bound inference, ...
[ "Yuetao Chen", "Xuliang Wang", "Xinzhou Zheng", "Ming Li", "Peng Wang", "Hong Xu" ]
[ "cs.CL", "cs.AI", "cs.DC", "cs.LG" ]
[ "Computer Science" ]
2026-02-02T00:00:00
https://arxiv.org/abs/2602.21224
https://arxiv.org/pdf/2602.21224v1
2602.21224
10.48550/arXiv.2602.21224
1
0
false
null
arXiv.org
0.4618
b4f398f7a2d4c0ae59e38cc0ea5a6ce47816c77d0a7e11e49ef64e1e9ad1fe0c
[ "arxiv", "semantic_scholar" ]
PRISM: Parametrically Refactoring Inference for Speculative Sampling Draft Models
Large Language Models (LLMs), constrained by their auto-regressive nature, suffer from slow decoding. Speculative decoding methods have emerged as a promising solution to accelerate LLM decoding, attracting attention from both systems and AI research communities. Recently, the pursuit of better draft quality has driven...
[ "Xuliang Wang", "Yuetao Chen", "Maochan Zhen", "Fang Liu", "Xinzhou Zheng", "Xingwu Liu", "Hong Xu", "Ming Li" ]
[ "cs.AI", "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2026-02-02T00:00:00
https://arxiv.org/abs/2602.01762
https://arxiv.org/pdf/2602.01762v1
2602.01762
10.48550/arXiv.2602.01762
0
0
false
null
arXiv.org
0.4618
769e41cd1d22874d58fb8564839543c8a6179b06ca1fbc86a685124cb1e6449f
[ "arxiv", "semantic_scholar" ]
PACER: Blockwise Pre-verification for Speculative Decoding with Adaptive Length
Speculative decoding (SD) is a powerful technique for accelerating the inference process of large language models (LLMs) without sacrificing accuracy. Typically, SD employs a small draft model to generate a fixed number of draft tokens, which are then verified in parallel by the target model. However, our experiments r...
[ "Situo Zhang", "Yifan Zhang", "Zichen Zhu", "Hankun Wang", "Da Ma", "Danyang Zhang", "Lu Chen", "Kai Yu" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2026-02-01T00:00:00
https://arxiv.org/abs/2602.01274
https://arxiv.org/pdf/2602.01274v1
2602.01274
10.48550/arXiv.2602.01274
0
0
false
null
arXiv.org
0.4606
89d1e1fc975cced6084cc06ce3ffba8e4ba20a8b94bfd5fd49714bbf63298b44
[ "arxiv", "semantic_scholar" ]
SAGE: Accelerating Vision-Language Models via Entropy-Guided Adaptive Speculative Decoding
Speculative decoding has emerged as a promising approach to accelerate inference in vision-language models (VLMs) by enabling parallel verification of multiple draft tokens. However, existing methods rely on static tree structures that remain fixed throughout the decoding process, failing to adapt to the varying predic...
[ "Yujia Tong", "Tian Zhang", "Yunyang Wan", "Kaiwei Lin", "Jingling Yuan", "Chuang Hu" ]
[ "cs.CV" ]
[ "Computer Science" ]
2026-01-31T00:00:00
https://arxiv.org/abs/2602.00523
https://arxiv.org/pdf/2602.00523v1
2602.00523
10.48550/arXiv.2602.00523
1
0
false
null
arXiv.org
0.4595
9a5688d2f937cce7542260fd24d7eadb2ff8471d3cb861d072d4477441b2eb39
[ "arxiv", "semantic_scholar" ]
LLMs as High-Dimensional Nonlinear Autoregressive Models with Attention: Training, Alignment and Inference
Large language models (LLMs) based on transformer architectures are typically described through collections of architectural components and training procedures, obscuring their underlying computational structure. This review article provides a concise mathematical reference for researchers seeking an explicit, equation...
[ "Vikram Krishnamurthy" ]
[ "cs.LG", "cs.AI", "cs.CL", "eess.SP" ]
[ "Computer Science", "Engineering" ]
2026-01-31T00:00:00
https://arxiv.org/abs/2602.00426
https://arxiv.org/pdf/2602.00426v1
2602.00426
10.48550/arXiv.2602.00426
1
0
false
null
arXiv.org
0.4595
f7a595a710135729e69c3c16bd32c089b67df85f11132127cf0243363a1df9a4
[ "arxiv", "semantic_scholar" ]
TriSpec: Ternary Speculative Decoding via Lightweight Proxy Verification
Inference efficiency in Large Language Models (LLMs) is fundamentally limited by their serial, autoregressive generation, especially as reasoning becomes a key capability and response sequences grow longer. Speculative decoding (SD) offers a powerful solution, providing significant speed-ups through its lightweight dra...
[ "Haoyun Jiang", "Junqi He", "Feng Hong", "Xinlong Yang", "Jianwei Zhang", "Zheng Li", "Zhengyang Zhuge", "Zhiyong Chen", "Bo Han", "Junyang Lin", "Jiangchao Yao" ]
[ "cs.LG" ]
[ "Computer Science" ]
2026-01-30T00:00:00
https://arxiv.org/abs/2601.23180
https://arxiv.org/pdf/2601.23180v1
2601.23180
10.48550/arXiv.2601.23180
1
0
false
null
arXiv.org
0.4583
e653377ac773be3448a90ab49ee5a50810cfdcb3c85d636dfe531e5ae857cf2a
[ "arxiv", "semantic_scholar" ]
StarSD: One-for-Many Speculative Decoding
Speculative decoding accelerates autoregressive generation by separating token proposal from verification, but most existing approaches are designed for single-node execution and do not scale well to multi-accelerator clusters used for serving modern Large Language Models (LLMs). We present StarSD, a one-for-many specu...
[ "Junhao He", "Feiran You", "Hongyang Du" ]
[ "eess.SY" ]
[ "Engineering", "Computer Science" ]
2026-01-29T00:00:00
https://arxiv.org/abs/2601.21622
https://arxiv.org/pdf/2601.21622v1
2601.21622
10.48550/arXiv.2601.21622
0
0
false
null
arXiv.org
0.4572
ce489f83ca124233083ae64ef35a607f69743de5cee2a10a19c3535b29f0d3ac
[ "arxiv", "semantic_scholar" ]
TABED: Test-Time Adaptive Ensemble Drafting for Robust Speculative Decoding in LVLMs
Speculative decoding (SD) has proven effective for accelerating LLM inference by quickly generating draft tokens and verifying them in parallel. However, SD remains largely unexplored for Large Vision-Language Models (LVLMs), which extend LLMs to process both image and text prompts. To address this gap, we benchmark ex...
[ "Minjae Lee", "Wonjun Kang", "Byeongkeun Ahn", "Christian Classen", "Kevin Galim", "Seunghyuk Oh", "Minghao Yan", "Hyung Il Koo", "Kangwook Lee" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2026-01-28T00:00:00
https://arxiv.org/abs/2601.20357
https://arxiv.org/pdf/2601.20357v1
2601.20357
10.48550/arXiv.2601.20357
0
0
true
https://github.com/furiosa-ai/TABED
Conference of the European Chapter of the Association for Computational Linguistics
0.7048
307d3d85a4d5d45ddcfa6a0e380c060eeca2db027fe36c0aa0d6a19e4db304f8
[ "arxiv", "semantic_scholar" ]
DART: Diffusion-Inspired Speculative Decoding for Fast LLM Inference
Speculative decoding is an effective and lossless approach for accelerating LLM inference. However, existing widely adopted model-based draft designs, such as EAGLE3, improve accuracy at the cost of multi-step autoregressive inference, resulting in high drafting latency and ultimately rendering the drafting stage itsel...
[ "Fuliang Liu", "Xue Li", "Ketai Zhao", "Yinxi Gao", "Ziyan Zhou", "Zhonghui Zhang", "Zhibin Wang", "Wanchun Dou", "Sheng Zhong", "Chen Tian" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-01-27T00:00:00
https://arxiv.org/abs/2601.19278
https://arxiv.org/pdf/2601.19278v1
2601.19278
10.48550/arXiv.2601.19278
11
2
true
https://github.com/fvliang/DART
arXiv.org
0.703
ec1cffae932cd51406ec96fbfe04ce78dd9a2df67f2b282cc6fde30174c0a006
[ "arxiv", "semantic_scholar" ]
LLM-42: Enabling Determinism in LLM Inference with Verified Speculation
In LLM inference, the same prompt may yield different outputs across different runs. At the system level, this non-determinism arises from floating-point non-associativity combined with dynamic batching and GPU kernels whose reduction orders vary with batch size. A straightforward way to eliminate non-determinism is to...
[ "Raja Gond", "Aditya K Kamath", "Ramachandran Ramjee", "Ashish Panwar" ]
[ "cs.LG", "cs.AI", "cs.DC" ]
[ "Computer Science" ]
2026-01-25T00:00:00
https://arxiv.org/abs/2601.17768
https://arxiv.org/pdf/2601.17768v2
2601.17768
10.48550/arXiv.2601.17768
7
0
true
https://github.com/microsoft/llm-42
arXiv.org
0.6995
440d3337d6a8a10ae8eabd7da0d34d6b7c5575cc4fa33589931467ba36e790fd
[ "arxiv", "semantic_scholar" ]
MARS: Unleashing the Power of Speculative Decoding via Margin-Aware Verification
Speculative Decoding (SD) accelerates autoregressive large language model (LLM) inference by decoupling generation and verification. While recent methods improve draft quality by tightly coupling the drafter with the target model, the verification mechanism itself remains largely unchanged, relying on strict token-leve...
[ "Jingwei Song", "Xinyu Wang", "Hanbin Wang", "Xiaoxuan Lei", "Bill Shi", "Shixin Han", "Eric Yang", "Xiao-Wen Chang", "Lynn Ai" ]
[ "cs.LG" ]
[ "Computer Science" ]
2026-01-21T00:00:00
https://arxiv.org/abs/2601.15498
https://arxiv.org/pdf/2601.15498v2
2601.15498
10.48550/arXiv.2601.15498
1
0
true
https://github.com/5SSjw/MARS
arXiv.org
0.6924
d389a309b0d387da23092e00ff5101132f9df70d303a178d43587fb6f595264b
[ "arxiv", "semantic_scholar" ]
Stabilizer-Assisted Inactivation Decoding of Quantum Error-Correcting Codes with Erasures
In this work, we develop a reduced complexity maximum likelihood (ML) decoder for quantum low-density parity-check (QLDPC) codes over erasures. Our decoder combines classical inactivation decoding, which integrates peeling with symbolic guessing, with a new dual peeling procedure. In the dual peeling stage, we perform ...
[ "Giulio Pech", "Mert Gökduman", "Hanwen Yao", "Henry D. Pfister" ]
[ "cs.IT" ]
[ "Computer Science", "Mathematics" ]
2026-01-20T00:00:00
https://arxiv.org/abs/2601.14236
https://arxiv.org/pdf/2601.14236v1
2601.14236
10.48550/arXiv.2601.14236
0
0
false
null
arXiv.org
0.4469
aff354abaf6986106576eaed09db112bc85046fcf08c4a59ba54d6b5d51574aa
[ "arxiv", "semantic_scholar" ]
WISP: Waste- and Interference-Suppressed Distributed Speculative LLM Serving at the Edge via Dynamic Drafting and SLO-Aware Batching
As Large Language Models (LLMs) become increasingly accessible to end users, an ever-growing number of inference requests are initiated from edge devices and computed on centralized GPU clusters. However, the resulting exponential growth in computation workload is placing significant strain on data centers, while edge ...
[ "Xiangchen Li", "Jiakun Fan", "Qingyuan Wang", "Dimitrios Spatharakis", "Saeid Ghafouri", "Hans Vandierendonck", "Deepu John", "Bo Ji", "Ali R. Butt", "Dimitrios S. Nikolopoulos" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2026-01-15T00:00:00
https://arxiv.org/abs/2601.11652
https://arxiv.org/pdf/2601.11652v2
2601.11652
10.48550/arXiv.2601.11652
0
0
false
null
null
0.2807
005c13f4225924a16d6ce92bd7dbd5537faca06c2ed425b20777f98bb15d6ff5
[ "arxiv", "semantic_scholar" ]
Annealed Relaxation of Speculative Decoding for Faster Autoregressive Image Generation
Despite significant progress in autoregressive image generation, inference remains slow due to the sequential nature of AR models and the ambiguity of image tokens, even when using speculative decoding. Recent works attempt to address this with relaxed speculative decoding but lack theoretical grounding. In this paper,...
[ "Xingyao Li", "Fengzhuo Zhang", "Cunxiao Du", "Hui Ji" ]
[ "cs.CV", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2026-01-14T00:00:00
https://arxiv.org/abs/2601.09212
https://arxiv.org/pdf/2601.09212v1
2601.09212
10.48550/arXiv.2601.09212
0
0
false
null
AAAI Conference on Artificial Intelligence
0.44
95ee200f10728e73faea9428052361304de5bd18298a44e8e342faf36c033172
[ "arxiv", "semantic_scholar" ]
HIPPO: Accelerating Video Large Language Models Inference via Holistic-aware Parallel Speculative Decoding
Speculative decoding (SD) has emerged as a promising approach to accelerate LLM inference without sacrificing output quality. Existing SD methods tailored for video-LLMs primarily focus on pruning redundant visual tokens to mitigate the computational burden of massive visual inputs. However, existing methods do not ach...
[ "Qitan Lv", "Tianyu Liu", "Wen Wu", "Xuenan Xu", "Bowen Zhou", "Feng Wu", "Chao Zhang" ]
[ "cs.CV", "cs.AI" ]
[ "Computer Science" ]
2026-01-13T00:00:00
https://arxiv.org/abs/2601.08273
https://arxiv.org/pdf/2601.08273v1
2601.08273
10.48550/arXiv.2601.08273
2
0
false
null
arXiv.org
0.4389
6d78267baef039f0be8c9a0a4c9abdc16090a78de7e7488b42ba5562f36586cd
[ "arxiv", "semantic_scholar" ]
TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees
Speculative decoding (SD) has become a standard technique for accelerating LLM inference without sacrificing output quality. Recent advances in speculative decoding have shifted from sequential chain-based drafting to tree-structured generation, where the draft model constructs a tree of candidate tokens to explore mul...
[ "Tianyu Liu", "Qitan Lv", "Yuhao Shen", "Xiao Sun", "Xiaoyan Sun" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-01-12T00:00:00
https://arxiv.org/abs/2601.07353
https://arxiv.org/pdf/2601.07353v1
2601.07353
10.48550/arXiv.2601.07353
8
0
false
null
arXiv.org
0.4377
5351116365b1b6f228e89daffd95a548d2b485d68f4a15f38abca686c6cf73b9
[ "arxiv", "semantic_scholar" ]
Overcoming Joint Intractability with Lossless Hierarchical Speculative Decoding
Verification is a key bottleneck in improving inference speed while maintaining distribution fidelity in Speculative Decoding. Recent work has shown that sequence-level verification leads to a higher number of accepted tokens compared to token-wise verification. However, existing solutions often rely on surrogate appro...
[ "Yuxuan Zhou", "Fei Huang", "Heng Li", "Fengyi Wu", "Tianyu Wang", "Jianwei Zhang", "Junyang Lin", "Zhi-Qi Cheng" ]
[ "cs.AI" ]
[ "Computer Science" ]
2026-01-09T00:00:00
https://arxiv.org/abs/2601.05724
https://arxiv.org/pdf/2601.05724v2
2601.05724
10.48550/arXiv.2601.05724
1
0
true
https://github.com/ZhouYuxuanYX/Hierarchical-Speculative-Decoding
arXiv.org
0.6711
06f5a3e1424f3adfa47cfcf8890829cbc82a3d8d5ff2f3aebf5909ec5beaf65a
[ "arxiv", "semantic_scholar" ]
Multi-Scale Local Speculative Decoding for Image Generation
Autoregressive (AR) models have achieved remarkable success in image synthesis, yet their sequential nature imposes significant latency constraints. Speculative Decoding offers a promising avenue for acceleration, but existing approaches are limited by token-level ambiguity and lack of spatial awareness. In this work, ...
[ "Elia Peruzzo", "Guillaume Sautière", "Amirhossein Habibian" ]
[ "cs.CV" ]
[ "Computer Science" ]
2026-01-08T00:00:00
https://arxiv.org/abs/2601.05149
https://arxiv.org/pdf/2601.05149v2
2601.05149
10.48550/arXiv.2601.05149
1
0
false
null
arXiv.org
0.4331
d46349153a3709ab81f615bc8c6471466fdbf2a9fdd4931a6860ab4049b92ff2
[ "arxiv", "semantic_scholar" ]
LoRA-Drop: Temporal LoRA Decoding for Efficient LLM Inference
Autoregressive large language models (LLMs) are bottlenecked by sequential decoding, where each new token typically requires executing all transformer layers. Existing dynamic-depth and layer-skipping methods reduce this cost, but often rely on auxiliary routing mechanisms or incur accuracy degradation when bypassed la...
[ "Hossein Rajabzadeh", "Maryam Dialameh", "Chul B. Park", "Il-Min Kim", "Hyock Ju Kwon" ]
[ "cs.CL" ]
[ "Computer Science" ]
2026-01-05T00:00:00
https://arxiv.org/abs/2601.02569
https://arxiv.org/pdf/2601.02569v1
2601.02569
10.48550/arXiv.2601.02569
0
0
true
https://github.com/hosseinbv/LoRA-Drop.git
arXiv.org
0.6641
b9dfa6ab794986bc558b147c423c668cee3f76d8115b4907b4751157b2b1a041
[ "arxiv", "semantic_scholar" ]
FlexSpec: Frozen Drafts Meet Evolving Targets in Edge-Cloud Collaborative LLM Speculative Decoding
Deploying large language models (LLMs) in mobile and edge computing environments is constrained by limited on-device resources, scarce wireless bandwidth, and frequent model evolution. Although edge-cloud collaborative inference with speculative decoding (SD) can reduce end-to-end latency by executing a lightweight dra...
[ "Yuchen Li", "Rui Kong", "Zhonghao Lyu", "Qiyang Li", "Xinran Chen", "Hengyi Cai", "Lingyong Yan", "Shuaiqiang Wang", "Jiashu Zhao", "Guangxu Zhu", "Linghe Kong", "Guihai Chen", "Haoyi Xiong", "Dawei Yin" ]
[ "cs.DC" ]
[ "Computer Science" ]
2026-01-02T00:00:00
https://arxiv.org/abs/2601.00644
https://arxiv.org/pdf/2601.00644v1
2601.00644
10.48550/arXiv.2601.00644
3
0
false
null
IEEE Transactions on Mobile Computing
0.4263
6a37ecf101f70bd44f86bc1ec79283bf4ddba6aaab0277f37f8ea57d09725b37
[ "arxiv", "semantic_scholar" ]
Speculative Decoding: Performance or Illusion?
Speculative decoding (SD) has become a popular technique to accelerate Large Language Model (LLM) inference, yet its real-world effectiveness remains unclear as prior evaluations rely on research prototypes and unrealistically small batch sizes. We present, to our knowledge, the first systematic study of SD on a produc...
[ "Xiaoxuan Liu", "Jiaxiang Yu", "Jongseok Park", "Ion Stoica", "Alvin Cheung" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-12-31T00:00:00
https://arxiv.org/abs/2601.11580
https://arxiv.org/pdf/2601.11580v2
2601.11580
10.48550/arXiv.2601.11580
8
0
false
null
arXiv.org
0.424
f7fa9d9ad5bb6239527942757ebe4d044080c4c6b72e3feb5e53ca0658b899a9
[ "arxiv", "semantic_scholar" ]
Entropy-Aware Speculative Decoding Toward Improved LLM Reasoning
Speculative decoding (SD) accelerates large language model (LLM) reasoning by using a small draft model to generate candidate tokens, which the target LLM either accepts directly or regenerates upon rejection. However, excessive alignment between the draft and target models constrains SD to the performance of the targe...
[ "Tiancheng Su", "Meicong Zhang", "Guoxiu He" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-12-29T00:00:00
https://arxiv.org/abs/2512.23765
https://arxiv.org/pdf/2512.23765v1
2512.23765
10.48550/arXiv.2512.23765
2
0
false
null
arXiv.org
0.4217
fea2f8148499c044a32dcd8138ca3c35c6bc6a8a61306f1861540e27e9d0e9a2
[ "arxiv", "semantic_scholar" ]
Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding
Speculative decoding improves LLM inference by generating and verifying multiple tokens in parallel, but existing systems suffer from suboptimal performance due to a mismatch between dynamic speculation and static runtime assumptions. We present Yggdrasil, a co-designed system that enables latency-optimal speculative d...
[ "Yue Guan", "Changming Yu", "Shihan Fang", "Weiming Hu", "Zaifeng Pan", "Zheng Wang", "Zihan Liu", "Yangjie Zhou", "Yufei Ding", "Minyi Guo", "Jingwen Leng" ]
[ "cs.LG", "cs.PL" ]
[ "Computer Science" ]
2025-12-29T00:00:00
https://arxiv.org/abs/2512.23858
https://arxiv.org/pdf/2512.23858v1
2512.23858
10.48550/arXiv.2512.23858
3
1
false
null
arXiv.org
0.4217
ec7402bcbd806d51eeaa663fb73a85956c2e5fde1105175c8b980600fc939c97
[ "arxiv", "semantic_scholar" ]
Nightjar: Dynamic Adaptive Speculative Decoding for Large Language Models Serving
Speculative decoding (SD) accelerates LLM inference by verifying draft tokens in parallel. However, this method presents a critical trade-off: it improves throughput in low-load, memory-bound systems but degrades performance in high-load, compute-bound environments due to verification overhead. Existing speculative dec...
[ "Rui Li", "Zhaoning Zhang", "Libo Zhang", "Huaimin Wang", "Xiang Fu", "Zhiquan Lai" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2025-12-27T00:00:00
https://arxiv.org/abs/2512.22420
https://arxiv.org/pdf/2512.22420v5
2512.22420
10.48550/arXiv.2512.22420
3
0
false
null
Journal of systems architecture
0.4194
eedb793a0825b82ee5f0476441917a0bf7f9d4bf4543530592c1d806d8bafcca
[ "arxiv", "semantic_scholar" ]
Accelerate Speculative Decoding with Sparse Computation in Verification
Speculative decoding accelerates autoregressive language model inference by verifying multiple draft tokens in parallel. However, the verification stage often becomes the dominant computational bottleneck, especially for long-context inputs and mixture-of-experts (MoE) models. Existing sparsification methods are design...
[ "Jikai Wang", "Jianchao Tan", "Yuxuan Hu", "Jiayu Qin", "Yerui Sun", "Yuchen Xie", "Xunliang Cai", "Juntao Li", "Min Zhang" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-12-26T00:00:00
https://arxiv.org/abs/2512.21911
https://arxiv.org/pdf/2512.21911v1
2512.21911
10.48550/arXiv.2512.21911
0
0
false
null
arXiv.org
0.4182
74a85d93819e9e43dbac2ac476bb113996d249e0cf4fdfe7e9e202e9e91bc4ba
[ "arxiv", "semantic_scholar" ]
Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMs
Diffusion Large Language Models (dLLMs) offer fast, parallel token generation, but their standalone use is plagued by an inherent efficiency-quality tradeoff. We show that, if carefully applied, the attributes of dLLMs can actually be a strength for drafters in speculative decoding with autoregressive (AR) verifiers. O...
[ "Rui Pan", "Zhuofu Chen", "Hongyi Liu", "Arvind Krishnamurthy", "Ravi Netravali" ]
[ "cs.LG", "cs.AI", "cs.DC" ]
[ "Computer Science" ]
2025-12-23T00:00:00
https://arxiv.org/abs/2512.20573
https://arxiv.org/pdf/2512.20573v3
2512.20573
10.48550/arXiv.2512.20573
4
0
true
https://github.com/ruipeterpan/failfast
arXiv.org
0.641
c8c14eaa0db49706c434427c1ad212c0b3bdb8f610474766adc2115a0846be5a
[ "arxiv", "semantic_scholar" ]
Fast Collaborative Inference via Distributed Speculative Decoding
Speculative decoding accelerates large language model (LLM) inference by allowing a small draft model to predict multiple future tokens for verification by a larger target model. In AI-native radio access networks (AI-RAN), this enables device-edge collaborative inference but introduces significant uplink overhead, as ...
[ "Ce Zheng", "Ke Zhang", "Chen Sun", "Wenqi Zhang", "Qiong Liu", "Angesom Ataklity Tesfay" ]
[ "eess.SP" ]
[ "Engineering" ]
2025-12-18T00:00:00
https://arxiv.org/abs/2512.16273
https://arxiv.org/pdf/2512.16273v2
2512.16273
10.1016/j.jiixd.2025.12.008
3
0
false
null
Journal of Information and Intelligence
0.4091
499684e1a913c91000964b15c55641304481c53ac2a6d836d928a4bc0ca45e41
[ "arxiv", "semantic_scholar" ]
Efficient Adaptive Rejection Sampling for Accelerating Speculative Decoding in Large Language Models
Speculative Decoding is a prominent technique for accelerating the autoregressive inference of large language models (LLMs) by employing a fast draft model to propose candidate token sequences and a large target model to verify them in parallel. However, its core component -- the rejection sampling mechanism -- relies ...
[ "Chendong Sun", "Ali Mao", "Lei Xu", "mingmin Chen" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-12-15T00:00:00
https://arxiv.org/abs/2512.13194
https://arxiv.org/pdf/2512.13194v3
2512.13194
10.48550/arXiv.2512.13194
1
0
false
null
arXiv.org
0.4056
30507c9f738661addfb8d8658d3369800663d30373e2958894fc3615fac4659d
[ "arxiv", "semantic_scholar" ]
TS-DP: Reinforcement Speculative Decoding For Temporal Adaptive Diffusion Policy Acceleration
Diffusion Policy (DP) excels in embodied control but suffers from high inference latency and computational cost due to multiple iterative denoising steps. The temporal complexity of embodied tasks demands a dynamic and adaptable computation mode. Static and lossy acceleration methods, such as quantization, fail to hand...
[ "Ye Li", "Jiahe Feng", "Yuan Meng", "Kangye Ji", "Chen Tang", "Xinwan Wen", "Shutao Xia", "Zhi Wang", "Wenwu Zhu" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-12-13T00:00:00
https://arxiv.org/abs/2512.15773
https://arxiv.org/pdf/2512.15773v1
2512.15773
10.48550/arXiv.2512.15773
4
0
false
null
arXiv.org
0.4033
59a03ddf2e16daeaa14dc6cc178590dca8c3d72eca7829aa1b8c19168da8b458
[ "arxiv", "semantic_scholar" ]
AdaSD: Adaptive Speculative Decoding for Efficient Language Model Inference
Large language models (LLMs) have achieved remarkable performance across a wide range of tasks, but their increasing parameter sizes significantly slow down inference. Speculative decoding mitigates this issue by leveraging a smaller draft model to predict candidate tokens, which are then verified by a larger target mo...
[ "Kuan-Wei Lu", "Ding-Yong Hong", "Pangfeng Liu", "Jan-Jan Wu" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-12-12T00:00:00
https://arxiv.org/abs/2512.11280
https://arxiv.org/pdf/2512.11280v2
2512.11280
10.48550/arXiv.2512.11280
1
0
false
null
arXiv.org
0.4022
67e315d660d5f1469f1263b315c90fdec539f557aa3179f8a111ba7052ab0cea
[ "arxiv", "semantic_scholar" ]
Speculative Decoding Speed-of-Light: Optimal Lower Bounds via Branching Random Walks
Speculative generation has emerged as a promising technique to accelerate inference in large language models (LLMs) by leveraging parallelism to verify multiple draft tokens simultaneously. However, the fundamental limits on the achievable speedup remain poorly understood. In this work, we establish the first ``tight''...
[ "Sergey Pankratov", "Dan Alistarh" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-12-12T00:00:00
https://arxiv.org/abs/2512.11718
https://arxiv.org/pdf/2512.11718v1
2512.11718
10.48550/arXiv.2512.11718
0
0
false
null
Conference of the European Chapter of the Association for Computational Linguistics
0.4022
670a39eec13d5f66aa9e877ad2f1d84fbcb03bf2c49eea0cf0014f0004c268c8
[ "arxiv", "semantic_scholar" ]
CXL-SpecKV: A Disaggregated FPGA Speculative KV-Cache for Datacenter LLM Serving
Large Language Models (LLMs) have revolutionized natural language processing tasks, but their deployment in datacenter environments faces significant challenges due to the massive memory requirements of key-value (KV) caches. During the autoregressive decoding process, KV caches consume substantial GPU memory, limiting...
[ "Dong Liu", "Yanxuan Yu" ]
[ "cs.AI" ]
[ "Computer Science" ]
2025-12-11T00:00:00
https://arxiv.org/abs/2512.11920
https://arxiv.org/pdf/2512.11920v1
2512.11920
10.1145/3748173.3779188
7
0
true
https://github.com/FastLM/CXL-SpecKV
Symposium on Field Programmable Gate Arrays
0.6198
2e80d9b6fbf019bf647f2f04a06f3066ac60b83b23be734ec2b284fe908d6065
[ "arxiv", "semantic_scholar" ]
GoodSpeed: Optimizing Fair Goodput with Adaptive Speculative Decoding in Distributed Edge Inference
Large language models (LLMs) have revolutionized natural language processing, yet their high computational demands pose significant challenges for real-time inference, especially in multi-user server speculative decoding and resource-constrained environments. Speculative decoding has emerged as a promising technique to...
[ "Phuong Tran", "Tzu-Hao Liu", "Long Tan Le", "Tung-Anh Nguyen", "Van Quan La", "Eason Yu", "Han Shu", "Choong Seon Hong", "Nguyen H. Tran" ]
[ "cs.DC" ]
[ "Computer Science" ]
2025-12-10T00:00:00
https://arxiv.org/abs/2512.09963
https://arxiv.org/pdf/2512.09963v2
2512.09963
10.48550/arXiv.2512.09963
0
0
false
null
arXiv.org
0.3999
0557546589750fcbeab1a35e6bccf23929d4f4507b4d9454decdd8811ca5aefd
[ "arxiv", "semantic_scholar" ]
A scalable and real-time neural decoder for topological quantum codes
Fault-tolerant quantum computing will require error rates far below those achievable with physical qubits. Quantum error correction (QEC) bridges this gap, but depends on decoders being simultaneously fast, accurate, and scalable. This combination of requirements remains unmet by a machine-learning decoder, nor by any ...
[ "Andrew W. Senior", "Thomas Edlich", "Francisco J. H. Heras", "Lei M. Zhang", "Oscar Higgott", "James S. Spencer", "Taylor Applebaum", "Sam Blackwell", "Justin Ledford", "Akvilė Žemgulytė", "Augustin Žídek", "Noah Shutty", "Andrew Cowie", "Yin Li", "George Holland", "Peter Brooks", "...
[ "quant-ph", "cs.LG" ]
[ "Computer Science", "Physics" ]
2025-12-08T00:00:00
https://arxiv.org/abs/2512.07737
https://arxiv.org/pdf/2512.07737v2
2512.07737
10.48550/arXiv.2512.07737
19
2
false
null
arXiv.org
0.3976
442c163ae3ab78445c11b565f4c68cbc558b01e1bf82b7e0aa51bd34d7a50687
[ "arxiv", "semantic_scholar" ]
SJD++: Improved Speculative Jacobi Decoding for Training-free Acceleration of Discrete Auto-regressive Text-to-Image Generation
Large autoregressive models can generate high-quality, high-resolution images but suffer from slow generation speed, because these models require hundreds to thousands of sequential forward passes for next-token prediction during inference. To accelerate autoregressive text-to-image generation, we propose Speculative J...
[ "Yao Teng", "Zhihuan Jiang", "Han Shi", "Xian Liu", "Xuefei Ning", "Guohao Dai", "Yu Wang", "Zhenguo Li", "Xihui Liu" ]
[ "cs.CV" ]
[ "Medicine", "Computer Science" ]
2025-12-08T00:00:00
https://arxiv.org/abs/2512.07503
https://arxiv.org/pdf/2512.07503v1
2512.07503
10.48550/arXiv.2512.07503
0
0
false
null
IEEE Transactions on Pattern Analysis and Machine Intelligence
0.3976
be5a16eca63b9912582b7752969dbd11d89e3cda9093c17715c66ab5cd39ff9c
[ "arxiv", "semantic_scholar" ]
SpecPV: Improving Self-Speculative Decoding for Long-Context Generation via Partial Verification
Growing demands from tasks like code generation, deep reasoning, and long-document understanding have made long-context generation a crucial capability for large language models (LLMs). Speculative decoding is one of the most direct and effective approaches for accelerating generation. It follows a draft-verify paradig...
[ "Zhendong Tan", "Xingjun Zhang", "Chaoyi Hu", "Junjie Peng", "Kun Xia" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-12-02T00:00:00
https://arxiv.org/abs/2512.02337
https://arxiv.org/pdf/2512.02337v1
2512.02337
10.48550/arXiv.2512.02337
4
0
false
null
arXiv.org
0.3907
bac73db2cf10fc389dccbcb151df7fbb6920d13207327a04aa286b38972c4e3b
[ "arxiv", "semantic_scholar" ]
Accelerating Large-Scale Reasoning Model Inference with Sparse Self-Speculative Decoding
Reasoning language models have demonstrated remarkable capabilities on challenging tasks by generating elaborate chain-of-thought (CoT) solutions. However, such lengthy generation shifts the inference bottleneck from compute-bound to memory-bound. To generate each token, the model applies full attention to all previous...
[ "Yilong Zhao", "Jiaming Tang", "Kan Zhu", "Zihao Ye", "Chi-Chih Chang", "Chaofan Lin", "Jongseok Park", "Guangxuan Xiao", "Mohamed S. Abdelfattah", "Mingyu Gao", "Baris Kasikci", "Song Han", "Ion Stoica" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-12-01T00:00:00
https://arxiv.org/abs/2512.01278
https://arxiv.org/pdf/2512.01278v1
2512.01278
10.48550/arXiv.2512.01278
2
0
false
null
arXiv.org
0.3896
1812932d4c0c2e914ed10c58122b86492a9f2bfc70356a76770aae5a07324c65
[ "arxiv", "semantic_scholar" ]
Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match
Large language models (LLMs) achieve strong performance across diverse tasks but suffer from high inference latency due to their autoregressive generation. Speculative Decoding (SPD) mitigates this issue by verifying candidate tokens in parallel from a smaller draft model, yet its strict exact-match verification discar...
[ "Jinze Li", "Yixing Xu", "Guanchen Li", "Shuo Yang", "Jinfeng Xu", "Xuanwu Yin", "Dong Li", "Edith C. H. Ngai", "Emad Barsoum" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-11-28T00:00:00
https://arxiv.org/abs/2511.22972
https://arxiv.org/pdf/2511.22972v3
2511.22972
10.48550/arXiv.2511.22972
8
2
false
null
arXiv.org
0.3861
7c109b21d12fa51179d73bf274895c52d4ac5a25433627e6e32fe929aa2f80c7
[ "arxiv", "semantic_scholar" ]
DSD: A Distributed Speculative Decoding Solution for Edge-Cloud Agile Large Model Serving
Large language model (LLM) inference often suffers from high decoding latency and limited scalability across heterogeneous edge-cloud environments. Existing speculative decoding (SD) techniques accelerate token generation but remain confined to single-node execution. We propose DSD, a distributed speculative decoding f...
[ "Fengze Yu", "Leshu Li", "Brad McDanel", "Sai Qian Zhang" ]
[ "cs.LG", "cs.DC" ]
[ "Computer Science" ]
2025-11-26T00:00:00
https://arxiv.org/abs/2511.21669
https://arxiv.org/pdf/2511.21669v2
2511.21669
10.48550/arXiv.2511.21669
4
1
false
null
arXiv.org
0.3839
c1e7923edff546035821afb67aea2c6cea26aa2ea99eea9cceceaaef82c18b42
[ "arxiv", "semantic_scholar" ]
Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios
Speculative decoding accelerates LLM inference by utilizing otherwise idle computational resources during memory-to-chip data transfer. Current speculative decoding methods typically assume a considerable amount of available computing power, then generate a complex and massive draft tree using a small autoregressive la...
[ "Luohe Shi", "Zuchao Li", "Lefei Zhang", "Baoyuan Qi", "Guoming Liu", "Hai Zhao" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-11-25T00:00:00
https://arxiv.org/abs/2511.20340
https://arxiv.org/pdf/2511.20340v1
2511.20340
10.48550/arXiv.2511.20340
2
0
false
null
AAAI Conference on Artificial Intelligence
0.3827
4b9f8426458debdbea1d191d159c55a38276c6df91da7a3f3883f67f4fb88a40
[ "arxiv", "semantic_scholar" ]
Reducing Latency of LLM Search Agent via Speculation-based Algorithm-System Co-Design
LLM-based search agents achieve strong performance but suffer from severe latency, as each step requires serialized LLM reasoning followed by action of tool execution. We revisit this bottleneck through the lens of speculation. While traditional predict-verify speculation paradigm can break serial execution, its benefi...
[ "Zixiao Huang", "Wen Zeng", "Tianyu Fu", "Tengxuan Liu", "Yizhou Sun", "Ke Hong", "Xinhao Yang", "Chengchun Liu", "Yan Li", "Quanlu Zhang", "Guohao Dai", "Zhenhua Zhu", "Yu Wang" ]
[ "cs.AI", "cs.LG", "cs.PF" ]
[ "Computer Science" ]
2025-11-25T00:00:00
https://arxiv.org/abs/2511.20048
https://arxiv.org/pdf/2511.20048v1
2511.20048
10.48550/arXiv.2511.20048
1
1
false
null
arXiv.org
0.3827
0669f1c82d8e4cf5424ef9f2ab9e08268c702e0692afc844fa2d2e12c07c7380
[ "arxiv", "semantic_scholar" ]
Accelerating Time Series Foundation Models with Speculative Decoding
Modern web applications--from real-time content recommendation and dynamic pricing to CDN optimization--increasingly rely on time-series forecasting to deliver personalized experiences to billions of users. Large-scale Transformer-based models have achieved state-of-the-art performance in time-series forecasting but su...
[ "Pranav Subbaraman", "Fang Sun", "Yue Yao", "Huacong Tang", "Xiao Luo", "Yizhou Sun" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-11-22T00:00:00
https://arxiv.org/abs/2511.18191
https://arxiv.org/pdf/2511.18191v1
2511.18191
10.48550/arXiv.2511.18191
1
1
true
https://github.com/PranavSubbaraman/STRIDE
arXiv.org
0.5861
42300a828df305c43b757c35800a31c6e2f426c43adb8614194d31602809506c
[ "arxiv", "semantic_scholar" ]
Fast LLM Post-training via Decoupled and Fastest-of-N Speculation
Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. This work, SpecActor, achieves fast rollout with speculative decoding that deploys a fast draft path to accelerate the unparallelizable generation, while the corr...
[ "Rongxin Cheng", "Kai Zhou", "Xingda Wei", "Siyuan Liu", "Mingcong Han", "Mingjing Ai", "Yeju Zhou", "Baoquan Zhong", "Wencong Xiao", "Rong Chen", "Haibo Chen" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2025-11-20T00:00:00
https://arxiv.org/abs/2511.16193
https://arxiv.org/pdf/2511.16193v3
2511.16193
10.48550/arXiv.2511.16193
3
0
false
null
arXiv.org
0.377
4868fcf1f532f7e9defcefad5b634557cd8b43275d1cf72889cc0bfb8b6bd24b
[ "arxiv", "semantic_scholar" ]
Comparative Analysis of Large Language Model Inference Serving Systems: A Performance Study of vLLM and HuggingFace TGI
The deployment of Large Language Models (LLMs) in production environments requires efficient inference serving systems that balance throughput, latency, and resource utilization. This paper presents a comprehensive empirical evaluation of two prominent open-source LLM serving frameworks: vLLM and HuggingFace Text Gener...
[ "Saicharan Kolluru" ]
[ "cs.LG", "cs.DC", "cs.PF" ]
[ "Computer Science" ]
2025-11-17T00:00:00
https://arxiv.org/abs/2511.17593
https://arxiv.org/pdf/2511.17593v1
2511.17593
10.48550/arXiv.2511.17593
1
0
true
null
arXiv.org
0.5773
a47aaa8701ad8543fe841c4badb21ed10e029b487c238c6b4e48cdc2268d739b
[ "arxiv", "semantic_scholar" ]
Beat the long tail: Distribution-Aware Speculative Decoding for RL Training
Reinforcement learning(RL) post-training has become essential for aligning large language models (LLMs), yet its efficiency is increasingly constrained by the rollout phase, where long trajectories are generated token by token. We identify a major bottleneck:the long-tail distribution of rollout lengths, where a small ...
[ "Zelei Shao", "Vikranth Srivatsa", "Sanjana Srivastava", "Qingyang Wu", "Alpay Ariyak", "Xiaoxia Wu", "Ameen Patel", "Jue Wang", "Percy Liang", "Tri Dao", "Ce Zhang", "Yiying Zhang", "Ben Athiwaratkun", "Chenfeng Xu", "Junxiong Wang" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-11-17T00:00:00
https://arxiv.org/abs/2511.13841
https://arxiv.org/pdf/2511.13841v1
2511.13841
10.48550/arXiv.2511.13841
9
0
false
null
arXiv.org
0.3735
0a5f3ed1432b1b847bb5be16f5ba398f6f9f26621ba4b1591441ee34c4518193
[ "arxiv", "semantic_scholar" ]
Cacheback: Speculative Decoding With Nothing But Cache
We present Cacheback Decoding, a training-free and model-agnostic speculative decoding method that exploits the locality in language to accelerate Large Language Model (LLM) inference. Cacheback leverages only Least Recently Used (LRU) cache tables of token n-grams to generate draft sequences. Cacheback achieves state-...
[ "Zhiyao Ma", "In Gim", "Lin Zhong" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-11-15T00:00:00
https://arxiv.org/abs/2511.21699
https://arxiv.org/pdf/2511.21699v1
2511.21699
10.18653/v1/2025.emnlp-main.1581
3
0
false
null
Conference on Empirical Methods in Natural Language Processing
0.3713
5c970964f67e211782934047ac1149eaec7e55165aee8215dcf3bbb1c8ff7eae
[ "arxiv", "semantic_scholar" ]
Striking the Right Balance between Compute and Copy: Improving LLM Inferencing Under Speculative Decoding
With the skyrocketing costs of GPUs and their virtual instances in the cloud, there is a significant desire to use CPUs for large language model (LLM) inference. KV cache update, often implemented as allocation, copying, and in-place strided update for each generated token, incurs significant overhead. As the sequence ...
[ "Arun Ramachandran", "Ramaswamy Govindarajan", "Murali Annavaram", "Prakash Raghavendra", "Hossein Entezari Zarch", "Lei Gao", "Chaoyi Jiang" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2025-11-15T00:00:00
https://arxiv.org/abs/2511.12031
https://arxiv.org/pdf/2511.12031v1
2511.12031
10.48550/arXiv.2511.12031
0
0
false
null
arXiv.org
0.3713
f380d052435b23fff7004248d2e7f09ba440c028ec4121c72a7c957cdaa69fd2
[ "arxiv", "semantic_scholar" ]
Speculative Decoding in Decentralized LLM Inference: Turning Communication Latency into Computation Throughput
Speculative decoding accelerates large language model (LLM) inference by using a lightweight draft model to propose tokens that are later verified by a stronger target model. While effective in centralized systems, its behavior in decentralized settings, where network latency often dominates compute, remains under-char...
[ "Jingwei Song", "Wanyi Chen", "Xinyuan Song", " Max", "Chris Tong", "Gufeng Chen", "Tianyi Zhao", "Eric Yang", "Bill Shi", "Lynn Ai" ]
[ "cs.DC", "cs.AI" ]
[ "Computer Science" ]
2025-11-13T00:00:00
https://arxiv.org/abs/2511.11733
https://arxiv.org/pdf/2511.11733v1
2511.11733
10.48550/arXiv.2511.11733
2
1
false
null
arXiv.org
0.369
876b9d510c25384844b4e76cba4cda89cca09c751648aaefd4de6c661b391d10
[ "arxiv", "semantic_scholar" ]
Steering Pretrained Drafters during Speculative Decoding
Speculative decoding accelerates language model inference by separating generation into fast drafting and parallel verification. Its main limitation is drafter-verifier misalignment, which limits token acceptance and reduces overall effectiveness. While small drafting heads trained from scratch compensate with speed, t...
[ "Frédéric Berdoz", "Peer Rheinboldt", "Roger Wattenhofer" ]
[ "cs.LG", "cs.PF" ]
[ "Computer Science" ]
2025-11-13T00:00:00
https://arxiv.org/abs/2511.09844
https://arxiv.org/pdf/2511.09844v1
2511.09844
10.48550/arXiv.2511.09844
0
0
false
null
AAAI Conference on Artificial Intelligence
0.369
c253e46bf368d665c679cb9f5cee11730d06b8b6bdfdd8a9f7580f5523520550
[ "arxiv", "semantic_scholar" ]
Principled Coarse-Grained Acceptance for Speculative Decoding in Speech
Speculative decoding accelerates autoregressive speech generation by letting a fast draft model propose tokens that a larger target model verifies. However, for speech LLMs that generate acoustic tokens, exact token matching is overly restrictive: many discrete tokens are acoustically or semantically interchangeable, r...
[ "Moran Yanuka", "Paul Dixon", "Eyal Finkelshtein", "Daniel Rotman", "Raja Giryes" ]
[ "eess.AS", "cs.LG" ]
[ "Computer Science", "Engineering" ]
2025-11-05T00:00:00
https://arxiv.org/abs/2511.13732
https://arxiv.org/pdf/2511.13732v4
2511.13732
10.48550/arXiv.2511.13732
2
0
false
null
IEEE International Conference on Acoustics, Speech, and Signal Processing
0.3598
c3f745d68b303decafed4e49057542b350d1f5eb3477a1d7f8fa0e42d08911cf
[ "arxiv", "semantic_scholar" ]
Collaborative Large Language Model Inference via Resource-Aware Parallel Speculative Decoding
The growing demand for on-device large language model (LLM) inference highlights the need for efficient mobile edge computing (MEC) solutions, especially in resource-constrained settings. Speculative decoding offers a promising solution by partitioning token generation between a lightweight draft model on mobile device...
[ "Jungyeon Koh", "Hyun Jong Yang" ]
[ "cs.LG", "eess.SP" ]
[ "Computer Science", "Engineering" ]
2025-11-03T00:00:00
https://arxiv.org/abs/2511.01695
https://arxiv.org/pdf/2511.01695v4
2511.01695
10.48550/arXiv.2511.01695
0
0
false
null
arXiv.org
0.3575
82e7aad187c8d1079c665a2cf8dc154252b87c5d07f3d01b64cb9697c3dc5aa4
[ "arxiv", "semantic_scholar" ]
TapOut: A Bandit-Based Approach to Dynamic Speculative Decoding
Speculative decoding accelerates LLMs by using a lightweight draft model to generate tokens autoregressively before verifying them in parallel with a larger target model. However, determining the optimal number of tokens to draft remains a key challenge limiting the approach's effectiveness. Dynamic speculative decodin...
[ "Aditya Sridhar", "Nish Sinnadurai", "Sean Lie", "Vithursan Thangarasa" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2025-11-03T00:00:00
https://arxiv.org/abs/2511.02017
https://arxiv.org/pdf/2511.02017v1
2511.02017
10.48550/arXiv.2511.02017
1
0
false
null
arXiv.org
0.3575
fb1eb6085b163a354ea9cd43f252586e3dcac20d4d22b4f74bfd3776ae10f2aa
[ "arxiv", "semantic_scholar" ]
When, What, and How: Rethinking Retrieval-Enhanced Speculative Decoding
Speculative decoding (SD) has emerged as an effective technique to accelerate large language model (LLM) inference without compromising output quality. However, the achievable speedup largely depends on the effectiveness of the drafting model. While model-based methods like EAGLE-2 are accurate but costly, retrieval-en...
[ "Min Fang", "Zhihui Fu", "Qibin Zhao", "Jun Wang" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-11-03T00:00:00
https://arxiv.org/abs/2511.01282
https://arxiv.org/pdf/2511.01282v1
2511.01282
10.48550/arXiv.2511.01282
1
0
false
null
arXiv.org
0.3575
8cb9b7be3c81f03ae0f0a4ca5244c7510af10ca872daa5ddfa281a43a50b1d79
[ "arxiv", "semantic_scholar" ]
SpecDiff-2: Scaling Diffusion Drafter Alignment For Faster Speculative Decoding
Speculative decoding has become the standard approach for accelerating Large Language Model (LLM) inference. It exploits a lossless draft-then-verify procedure to circumvent the latency of autoregressive decoding, achieving impressive speed-ups. Yet, current speculative decoding approaches remain limited by two fundame...
[ "Jameson Sandler", "Jacob K. Christopher", "Thomas Hartvigsen", "Ferdinando Fioretto" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-11-01T00:00:00
https://arxiv.org/abs/2511.00606
https://arxiv.org/pdf/2511.00606v2
2511.00606
10.48550/arXiv.2511.00606
10
0
false
null
arXiv.org
0.3552
5fa7d9ee24b82123aeeb6c8d91cc117789bdd3d684949b7f811080298a5c308a
[ "arxiv", "semantic_scholar" ]
Reject Only Critical Tokens: Pivot-Aware Speculative Decoding
Speculative Decoding (SD) ensures that the output matches the target model's distribution exactly. However, we argue that this distribution matching requirement is too stringent and results in unnecessarily low acceptance rates, limiting potential speedups. Instead, we advocate a reformulation of the decoding objective...
[ "Amir Ziashahabi", "Yavuz Faruk Bakman", "Duygu Nur Yaldiz", "Mostafa El-Khamy", "Sai Praneeth Karimireddy", "Salman Avestimehr" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2025-11-01T00:00:00
https://arxiv.org/abs/2511.00351
https://arxiv.org/pdf/2511.00351v1
2511.00351
10.48550/arXiv.2511.00351
2
0
true
https://github.com/amir-zsh/PAD
arXiv.org
0.549
8416dc812eaed8f798cbf438e74eadcf685191de4939aa6b78968bd7904f199e
[ "arxiv", "semantic_scholar" ]
SpecAttn: Speculating Sparse Attention
Large Language Models (LLMs) face significant computational bottlenecks during inference due to the quadratic complexity of self-attention mechanisms, particularly as context lengths increase. We introduce SpecAttn, a novel training-free approach that seamlessly integrates with existing speculative decoding techniques ...
[ "Harsh Shah" ]
[ "cs.CL", "cs.LG", "eess.SY" ]
[ "Computer Science", "Engineering" ]
2025-10-31T00:00:00
https://arxiv.org/abs/2510.27641
https://arxiv.org/pdf/2510.27641v1
2510.27641
10.48550/arXiv.2510.27641
0
0
false
null
arXiv.org
0.3541
2e174777c59543ca0d26699225763832092fc478ef07f85e86ed6def7ac342ba
[ "arxiv", "semantic_scholar" ]
CAS-Spec: Cascade Adaptive Self-Speculative Decoding for On-the-Fly Lossless Inference Acceleration of LLMs
Speculative decoding has become a widely adopted as an effective technique for lossless inference acceleration when deploying large language models (LLMs). While on-the-fly self-speculative methods offer seamless integration and broad utility, they often fall short of the speed gains achieved by methods relying on spec...
[ "Zhiyuan Ning", "Jiawei Shao", "Ruge Xu", "Xinfei Guo", "Jun Zhang", "Chi Zhang", "Xuelong Li" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-10-30T00:00:00
https://arxiv.org/abs/2510.26843
https://arxiv.org/pdf/2510.26843v1
2510.26843
10.48550/arXiv.2510.26843
0
0
false
null
arXiv.org
0.3529
a07d6ed6c705aa74327ac24e68fc95aaeab3d5422fecd79dbc0663cd379de51a
[ "arxiv", "semantic_scholar" ]
Polybasic Speculative Decoding Through a Theoretical Perspective
Inference latency stands as a critical bottleneck in the large-scale deployment of Large Language Models (LLMs). Speculative decoding methods have recently shown promise in accelerating inference without compromising the output distribution. However, existing work typically relies on a dualistic draft-verify framework ...
[ "Ruilin Wang", "Huixia Li", "Yuexiao Ma", "Xiawu Zheng", "Fei Chao", "Xuefeng Xiao", "Rongrong Ji" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-10-30T00:00:00
https://arxiv.org/abs/2510.26527
https://arxiv.org/pdf/2510.26527v1
2510.26527
10.48550/arXiv.2510.26527
0
0
false
null
International Conference on Machine Learning
0.3529
e74f829e608057929e2acaa671b6e92e04fb164c8bc6d45cfb82f7ff9ab0ab79
[ "arxiv", "semantic_scholar" ]
ReSpec: Towards Optimizing Speculative Decoding in Reinforcement Learning Systems
Adapting large language models (LLMs) via reinforcement learning (RL) is often bottlenecked by the generation stage, which can consume over 75\% of the training time. Speculative decoding (SD) accelerates autoregressive generation in serving systems, but its behavior under RL training remains largely unexplored. We ide...
[ "Qiaoling Chen", "Zijun Liu", "Peng Sun", "Shenggui Li", "Guoteng Wang", "Ziming Liu", "Yonggang Wen", "Siyuan Feng", "Tianwei Zhang" ]
[ "cs.LG", "cs.DC" ]
[ "Computer Science" ]
2025-10-30T00:00:00
https://arxiv.org/abs/2510.26475
https://arxiv.org/pdf/2510.26475v1
2510.26475
10.48550/arXiv.2510.26475
12
2
false
null
arXiv.org
0.3529
6c5571a9a8c1a70f1d4d0a1a6b41025a1d1f8b78ac0ef4ef36ec34854191e76d
[ "arxiv", "semantic_scholar" ]
Speculative Coupled Decoding for Training-Free Lossless Acceleration of Autoregressive Visual Generation
Autoregressive (AR) modeling has recently emerged as a promising new paradigm in visual generation, but its practical adoption is severely constrained by the slow inference speed of per-token generation, which often requires thousands of steps to produce a single sample. While several Speculative Decoding (SD)-based me...
[ "Junhyuk So", "Hyunho Kook", "Chaeyeon Jang", "Eunhyeok Park" ]
[ "cs.CV" ]
[ "Computer Science" ]
2025-10-28T00:00:00
https://arxiv.org/abs/2510.24211
https://arxiv.org/pdf/2510.24211v2
2510.24211
null
2
0
true
https://github.com/junhyukso/SCD
null
0.4144
3e1f76c61d3fd3eb0e72cb811966a9cc8143dcfe426fef547478ada559e87e39
[ "arxiv", "semantic_scholar" ]
Batch Speculative Decoding Done Right
Speculative decoding must produce outputs distribution identical to standard autoregressive generation-this output equivalence is not an optimization target but the defining criterion of valid speculative decoding. We demonstrate that all existing batch speculative decoding implementations violate this fundamental requ...
[ "Ranran Haoran Zhang", "Soumik Dey", "Ashirbad Mishra", "Hansi Wu", "Binbin Li", "Rui Zhang" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-10-26T00:00:00
https://arxiv.org/abs/2510.22876
https://arxiv.org/pdf/2510.22876v3
2510.22876
10.48550/arXiv.2510.22876
1
0
true
https://github.com/eBay/spec_dec
arXiv.org
0.5383
424cf2d40c53b479a9b54901307ef04f87102ad6f915e42b195b5e96f5c53a9e
[ "arxiv", "semantic_scholar" ]
FastVLM: Self-Speculative Decoding for Fast Vision-Language Model Inference
Vision-language Models (VLMs) have made significant strides in visual understanding and query response generation, but often face challenges of high computational cost and inference latency due to autoregressive decoding. In this work, we introduce an imitation-learning-based Self-Speculative Decoding (SSD) framework, ...
[ "Divya Jyoti Bajpai", "Manjesh Kumar Hanawal" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-10-26T00:00:00
https://arxiv.org/abs/2510.22641
https://arxiv.org/pdf/2510.22641v1
2510.22641
10.48550/arXiv.2510.22641
2
1
false
null
null
0.2217
9f10a6bfd30a7547fe28f913e5dbd921840f2eb08c48a7e6612bd15e92e9dc9b
[ "arxiv", "semantic_scholar" ]
Language Ranker: A Lightweight Ranking framework for LLM Decoding
Conventional research on large language models (LLMs) has primarily focused on refining output distributions, while paying less attention to the decoding process that transforms these distributions into final responses. Recent advances, such as scaling the computation of inference time with reward models, have undersco...
[ "Chenheng Zhang", "Tianqi Du", "Jizhe Zhang", "Mingqing Xiao", "Yifei Wang", "Yisen Wang", "Zhouchen Lin" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-10-23T00:00:00
https://arxiv.org/abs/2510.21883
https://arxiv.org/pdf/2510.21883v1
2510.21883
10.48550/arXiv.2510.21883
1
0
false
null
arXiv.org
0.3449
3fa3ac9526ecf22e5b087fbac6759bdf31765418cf52dd9ce391e9af20dd8678
[ "arxiv", "semantic_scholar" ]
Not-a-Bandit: Provably No-Regret Drafter Selection in Speculative Decoding for LLMs
Speculative decoding is widely used in accelerating large language model (LLM) inference. In this work, we focus on the online draft model selection problem in speculative decoding. We design an algorithm that provably competes with the best draft model in hindsight for each query in terms of either the token acceptanc...
[ "Hongyi Liu", "Jiaji Huang", "Zhen Jia", "Youngsuk Park", "Yu-Xiang Wang" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-10-22T00:00:00
https://arxiv.org/abs/2510.20064
https://arxiv.org/pdf/2510.20064v2
2510.20064
10.48550/arXiv.2510.20064
3
2
true
null
arXiv.org
0.5312
468aba3d5dbe2f421933777ca7432eb3797d420859add9220c4b914c583549fc
[ "arxiv", "semantic_scholar" ]
Fast Inference via Hierarchical Speculative Decoding
Transformer language models generate text autoregressively, making inference latency proportional to the number of tokens generated. Speculative decoding reduces this latency without sacrificing output quality, by leveraging a small draft model to propose tokens that the larger target model verifies in parallel. In pra...
[ "Clara Mohri", "Haim Kaplan", "Tal Schuster", "Yishay Mansour", "Amir Globerson" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-10-22T00:00:00
https://arxiv.org/abs/2510.19705
https://arxiv.org/pdf/2510.19705v2
2510.19705
10.48550/arXiv.2510.19705
1
0
false
null
arXiv.org
0.3438
568a9bc978c86633380376765bd736d91ae693e0ec7d9fb4cc3f9df7ef325065
[ "arxiv", "semantic_scholar" ]
AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders
Speculative Decoding (SD) accelerates large language model inference by employing a small draft model to generate predictions, which are then verified by a larger target model. The effectiveness of SD hinges on the alignment between these models, which is typically enhanced by Knowledge Distillation (KD). However, conv...
[ "Yuezhou Hu", "Jiaxin Guo", "Xinyu Feng", "Tuo Zhao" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-10-22T00:00:00
https://arxiv.org/abs/2510.19779
https://arxiv.org/pdf/2510.19779v1
2510.19779
10.48550/arXiv.2510.19779
7
0
true
https://github.com/yuezhouhu/adaspec
arXiv.org
0.5312
eccaa7e525efb5095c757c9e5121e1353b0ab0a04221e0e590a34958742a4267
[ "arxiv", "semantic_scholar" ]
From Quarter to All: Accelerating Speculative LLM Decoding via Floating-Point Exponent Remapping and Parameter Sharing
Large language models achieve impressive performance across diverse tasks but exhibit high inference latency due to their large parameter sizes. While quantization reduces model size, it often leads to performance degradation compared to the full model. Speculative decoding remains lossless but typically incurs extra o...
[ "Yushu Zhao", "Yubin Qin", "Yang Wang", "Xiaolong Yang", "Huiming Han", "Shaojun Wei", "Yang Hu", "Shouyi Yin" ]
[ "cs.AR" ]
[ "Computer Science" ]
2025-10-21T00:00:00
https://arxiv.org/abs/2510.18525
https://arxiv.org/pdf/2510.18525v1
2510.18525
10.48550/arXiv.2510.18525
1
0
false
null
arXiv.org
0.3426