Paper deep dive
Is Multimodal Speculative Decoding Ready for Diffusion-Based Parallel Drafting? A Survey and Empirical Diagnosis
Yantao Li, Huanlin Gao, Fang Zhao, Chao Tan, Qiang Hui, Shuting Liu, Fuyuan Shi, Ting Lu, Shaoan Zhao, Xueqiang Guo, Xinpei Su, Jianbing Zhang, Xinyu Dai, Kai Wang, Shiguo Lian
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/24/2026, 5:11:17 AM
Summary
This paper surveys and empirically diagnoses the readiness of multimodal speculative decoding (SD) for diffusion-based parallel drafting (L2). It introduces an L0-L2 taxonomy for drafter-side parallelism, analyzes cross-modal information interaction, and evaluates methods across Vision-Language, Video-Language, Audio, and VLA architectures. The study finds that while L2 drafting (e.g., DFlash, DSpark) offers significant speedups in text-only LLMs, its application to multimodal models faces challenges regarding draft-target agreement and conditioning overhead, though newer multimodal variants show promise.
Entities (15)
Relation Signals (9)
Speculative Decoding → uses → Lightweight Drafter
confidence 95% · Speculative decoding accelerates autoregressive generation by allowing a lightweight drafter to propose future tokens
DFlash → implements → L2
confidence 92% · DFlash (Chen et al., 2026) formulates future-token generation as a block-level diffusion process... L2 treats the future block as a jointly generated or refined unit
DSpark → implements → L2
confidence 92% · DSpark (Cheng et al., 2026) introduces lightweight intra-block dependency modeling... L2 treats the future block as a jointly generated or refined unit
EAGLE → implements → L0
confidence 90% · L0 advances one future token per forward step, covering autoregressive drafters and feature-based approaches such as EAGLE.
Medusa → implements → L1
confidence 90% · L1 predicts multiple predefined future positions from a shared prefix, including methods such as Medusa
Fast-dVLM → implements → L2
confidence 88% · Block-parallel VLMs provide direct L2 examples. Fast-dVLM converts an autoregressive VLM into a KV-cache-compatible block-diffusion model
Nemotron-Labs-Diffusion → implements → L2
confidence 88% · Nemotron-Labs-Diffusion jointly trains autoregressive and diffusion objectives... L2 examples
→ →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Speculative decoding accelerates autoregressive generation by allowing a lightweight drafter to propose future tokens while a target model verifies them in parallel. Its lossless guarantee has motivated a line of work that pushes the drafter itself toward parallel generation. The most recent paradigm is block-parallel generative drafting, including diffusion-based methods such as DFlash and DSpark, achieving up to 3.6x speedup on common daily chatting tasks. While this transition is well studied in text-only LLMs, its applicability to multimodal models remains an open question. Existing multimodal speculative decoding efforts focus on input compression, adapter alignment, candidate coverage, or modality-specific verification; however, block-parallel generative drafting remains largely unexplored. To bridge this gap, this paper combines a modality-centered survey with a cross-architecture empirical study to ask: Is multimodal speculative decoding ready for diffusion-based parallel drafting? In this survey, we systematically analyze a wide spectrum of multimodal models, spanning Vision-Language, Video-Language, Audio, and Vision-Language-Action (VLA) architectures, from the dual perspectives of drafting parallelism and cross-modal information interaction. We introduce a unified taxonomy that isolates drafter-side parallelism from orthogonal design choices such as tree construction and verification strategies. Furthermore, we provide a comprehensive empirical comparison of existing methods under varying degrees of parallelism across standardized multimodal benchmarks, including OCR, VQA, visual reasoning, and image captioning. Finally, we summarize the limitations of current approaches, discuss open challenges, and outline promising future directions for this rapidly evolving field.
Tags
Links
- Source: https://arxiv.org/abs/2608.20743v1
- Canonical: https://arxiv.org/abs/2608.20743v1
Trouble viewing inline? Open PDF directly →
Full Text
103,606 characters extracted from source content.
Expand or collapse full text
IS MULTIMODAL SPECULATIVE DECODING READY FOR DIFFUSION-BASED PARALLEL DRAFTING?A SURVEY AND EMPIRICAL DIAGNOSIS Yantao Li 1,2,3,∗ , Huanlin Gao 2,3,∗ , Fang Zhao 2,3,∗ , Chao Tan 2,3 , Qiang Hui 2,3 , Shuting Liu 2,3 , Fuyuan Shi 2,3 , Ting Lu 2,3 , Shaoan Zhao 2,3 , Xueqiang Guo 2,3 , Xinpei Su 2,3 , Jianbing Zhang 1 , Xinyu Dai 1,† , Kai Wang 2,3,‡ , Shiguo Lian 2,3,† 1 National Key Laboratory for Novel Software Technology, Nanjing University 2 Data Science & Artificial Intelligence Research Institute, China Unicom 3 Unicom Data Intelligence, China Unicom ABSTRACT Speculative decoding accelerates autoregressive generation by allowing a lightweight drafter to propose future tokens while a target model verifies them in parallel. Its lossless guarantee has motivated a line of work that pushes the drafter itself into parallel generation—the most recent enorous method is block- parallel generative approaches, including diffusion-based drafting methods such as DFlash and DSpark, achieving a 3.6x speedup on common daily chatting tasks. While this transition is well studied in text-only LLMs, its applicability to multi- modal models remains an open question. Existing multimodal speculative decod- ing efforts focus on input compression, adapter alignment, candidate coverage, or modality-specific verification; however, block-parallel generative drafting remains largely unexplored. To bridge this gap, this paper combines a modality-centered survey with a cross-architecture empirical study to ask: Is multimodal specula- tive decoding ready for diffusion-based parallel drafting? In this survey, we sys- tematically analyze a wide spectrum of multimodal models—spanning Vision- Language, Video-Language, Audio, and Vision-Language-Action (VLA) archi- tectures—from the dual perspectives of drafting parallelism and cross-modal in- formation interaction. We introduce a unified taxonomy that isolates drafter-side parallelism from orthogonal design choices such as tree construction and verifica- tion strategies. Furthermore, we provide a comprehensive empirical comparison of existing methods under varying degrees of parallelism across standardized mul- timodal benchmarks, including OCR, VQA, visual reasoning, and image caption- ing. Finally, we summarize the limitations of current approaches, discuss open challenges, and outline promising future directions for this rapidly evolving field. 1INTRODUCTION Multimodal generative models have achieved remarkable progress in a wide range of applications, including visual question answering, document understanding, chart reasoning, video analysis, speech interaction, and embodied control. Despite these advances, inference remains dominated by autoregressive decoding. After multimodal context encoding and prefill, the decoder still gener- ates tokens sequentially, requiring one expensive target model execution for each output token. This limitation becomes increasingly severe for long-form reasoning, OCR-heavy generation, structured documents, and interactive multimodal agents. Speculative decoding reduces this autoregressive bottleneck by introducing a lightweight drafter that proposes future tokens, which are then verified by the target model in parallel (Leviathan et al., 2023; Chen et al., 2023). The achievable acceleration depends on two complementary factors: how * Equal contribution. † Corresponding author. ‡ Project leader. 1 arXiv:2608.20743v1 [cs.AI] 21 Aug 2026 many draft tokens can be accepted by the target model and how efficiently the draft candidates can be generated. Existing speculative decoding methods have therefore evolved along two directions. The first direction improves verification efficiency by increasing candidate coverage. Tree-based speculative decoding methods, including SpecInfer (Miao et al., 2024), Sequoia (Chen et al., 2024), DDTree (Ringel & Romano, 2026), and TAPS (Wang et al., 2026a), organize multiple candidate continuations into tree structures and verify them in parallel, improving the expected accepted tokens under a fixed verification budget. These approaches mainly optimize the utilization of each target model forward pass while retaining sequential or weakly parallel drafting. The second direction improves drafter-side parallelism. Traditional autoregressive drafters generate a block of K future tokens through K serial draft steps. Feature-based approaches such as EA- GLE (Li et al., 2024a;b; 2025b), multi-head prediction methods such as Medusa (Cai et al., 2024), and multi-token prediction (Gloeckle et al., 2024) reduce drafting latency by predicting multiple future positions in parallel. More recently, diffusion-based block-parallel drafting has emerged as a stronger form of output-side parallelism. DFlash (Chen et al., 2026) formulates future-token gener- ation as a block-level diffusion process, while DSpark (Cheng et al., 2026) introduces lightweight intra-block dependency modeling and confidence-aware scheduling. These methods shift specula- tive decoding from token-wise prediction toward jointly generating future blocks. Together, these advances change the central question of speculative decoding from “Can the target verify multiple tokens simultaneously?” to “Can the drafter generate sufficiently accurate future blocks without paying the cost of serial drafting?” Multimodal speculative decoding introduces additional challenges beyond text-only generation. Ex- isting approaches mainly investigate how a drafter should access visual information, how redundant visual tokens should be compressed, how target representations can improve draft quality, and how verification should account for perceptual or functional equivalence (Gagrani et al., 2024; Ganesan et al., 2025; Hu et al., 2025b; Huang et al., 2025; Wang et al., 2025a). These studies demonstrate that multimodal conditioning is critical for maintaining draft–target agreement. However, most existing methods remain based on token-wise or semi-parallel drafting. Whether block-parallel drafting can provide practical acceleration for frozen, off-the-shelf multimodal targets under exact verification remains largely unexplored. To analyze this question systematically, we introduce a three-level taxonomy based on drafter-side parallelism. L0 advances one future token per forward step, covering autoregressive drafters and feature-based approaches such as EAGLE. L1 predicts multiple predefined future positions from a shared prefix, including methods such as Medusa and multi-token prediction. L2 treats the future block as a jointly generated or refined unit, including recent block-parallel methods such as DFlash and DSpark. Candidate organization strategies, such as tree branching and verification rules, are orthogonal to this taxonomy. Under this view, most existing multimodal speculative decoding meth- ods operate at L0 or L1. Existing L2 multimodal systems either modify the target architecture for self-speculation (Wu et al., 2026; Fu et al., 2026) or relax exact verification (Hu & Zhang, 2026; Niu et al., 2026). The general applicability of L2 drafting for frozen multimodal targets therefore remains an open question. We investigate this question through a controlled empirical study of diffusion-based parallel drafting on multimodal models. Interestingly, directly transferring an L2 drafter trained on Qwen3 leads to limited draft–target agreement, while the same block-parallel drafting paradigm achieves substan- tially stronger alignment on newer multimodal variants. This observation suggests that multimodal- ity itself is not the fundamental obstacle; instead, the effectiveness of L2 drafting depends on how multimodal information is represented, aligned, and utilized during generation. Therefore, rather than asking whether L2 drafting works for multimodal models, we study under what model, task, and system conditions block-parallel drafting can translate draft quality improvements into practical acceleration. In this paper, we present a systematic empirical study of multimodal L2 speculative decoding readi- ness. Our contributions are: 1. A unified taxonomy and empirical study of multimodal L2 drafting. We establish an L0–L2 taxonomy based on drafter-side parallelism and evaluate block-parallel drafting across representative multimodal models with different scales and architectures. 2 2. Analysis of multimodal conditioning bottlenecks. We investigate how visual information affects draft–target agreement and end-to-end efficiency, revealing the trade-off between multimodal awareness and additional conditioning overhead. 3. A readiness characterization for multimodal block-parallel decoding. We analyze the impact of model characteristics, task and input conditions, and framework support on prac- tical L2 acceleration, and summarize remaining challenges toward broader multimodal speculative decoding deployment. The remainder of this paper is organized as follows. Section 2 introduces multimodal speculative de- coding and the information-preservation requirements behind draft verification. Section 3 presents our L0–L2 taxonomy of drafter-side parallelism. Section 4 evaluates the readiness of block-parallel drafting for multimodal targets through cross-model evaluation, conditioning analysis, and frame- work support analysis. Section 5 synthesizes the empirical findings and discusses future directions before concluding in Section 6. 2MULTIMODAL SPECULATIVE DECODING 2.1FOUNDATIONS Standard speculative decoding. Speculative decoding accelerates autoregressive generation with a draft-and-verify loop. It uses a lightweight drafter to propose multiple future tokens, allowing the target model to verify them in a single forward pass instead of decoding them one by one. Given a prefix x <t , the drafter proposes K future tokens, and the target scores the drafted positions in par- allel while also producing the next-token distribution after the draft block. Under exact speculative sampling, for i = 0, . . . , K− 1, a drafted token ̃x t+i is accepted with probability a i ( ̃x t+i ) = min 1, p( ̃x t+i | x <t , ̃x t:t+i−1 ) q( ̃x t+i | x <t , ̃x t:t+i−1 ) ,(1) where ̃x t:t−1 is empty for i = 0. At the first rejection, the accepted prefix is retained and a correction token is sampled from the normalized positive residual between the corresponding target and draft distributions; after a full acceptance, the target contributes one additional token. This procedure preserves the target sampling distribution (Leviathan et al., 2023; Chen et al., 2023). Its benefit depends on whether the accepted output amortizes the cost of drafting and verification. Moreover, although the target evaluates multiple drafted positions in parallel, a conventional drafter produces them in causal order, so draft generation can remain a serial bottleneck. Sequential and parallel drafting. Parallel drafting reduces this draft-side dependence rather than merely parallelizing target verification. It can take different forms, from predicting multiple future positions in one pass to generating or refining a future block as a coupled state (Stern et al., 2018; Cai et al., 2024; Gloeckle et al., 2024). DFlash exemplifies diffusion-based parallel drafting by using a lightweight block-diffusion model to generate a future block for verification by an autoregressive target (Chen et al., 2026). Section 3 formalizes the progression from sequential to increasingly par- allel draft generation. These mechanisms are commonly studied for text generation, but multimodal settings introduce additional constraints on both input processing and output verification. Multimodal speculative decoding. Unlike text-only SD, a multimodal drafter may need to pro- cess image, video, audio, or environment information that also conditions the target. Access to this information can improve target–draft agreement, but introduces additional modality-specific computation and memory cost. Some methods instead reuse target-derived features or draft with- out directly processing the multimodal input. Multimodal applications may also generate actions, acoustic or codec tokens, and visual tokens or latents rather than text tokens. For these outputs, ex- act token identity is not always the only useful notion of validity, and some methods adopt task- or representation-specific acceptance rules. The next subsection introduces the analytical dimensions used to compare these design choices across application domains. 3 2.2MULTIMODAL SPECULATIVE DECODING LANDSCAPE We organize the multimodal speculative decoding literature considered in this survey into five ap- plication domains: vision–language models, video–language models, vision–language–action mod- els, speech and audio models, and autoregressive visual generation. These domains differ in the multimodal information available during drafting, the outputs being generated, and their task and deployment constraints. Such differences affect both how speculative candidates are produced and how they can be verified, motivating application domain as the primary organization of the review. Within each domain, we examine how candidate generation varies in its serial dependence across future positions. We annotate this property using the L0–L2 scale, which is formalized in Sec- tion 3. Methods with the same P-level can still differ substantially in how the drafter accesses mul- timodal information, how candidates are organized or obtained, how they are accepted or verified, and whether speculative computation adapts to the current input or state. We track these proper- ties separately because they can change the effectiveness of speculative inference without changing future-position dependence. Table 1 summarizes these method-level characteristics, while the dis- cussion below synthesizes the main design patterns and available evidence within each application domain. Vision–language models. In VLMs, visual access can improve target–draft alignment, but its pro- cessing cost can offset the latency saved by accepting more draft tokens. Early work therefore asks whether the drafter needs direct visual input at all. SPD-MLLM shows that a language-only drafter can already accelerate a visually conditioned target, while adding a compact image adaptor produces task-dependent rather than uniform gains (Gagrani et al., 2024). MSD instead develops an explic- itly multimodal drafter: it processes text and visual tokens separately and uses text-only followed by multimodal training to acquire both language-modeling and visual-perception capabilities (Lin et al., 2025a). Both retain an autoregressive draft trajectory, but establish the two endpoints between condition-blind and explicitly multimodal drafting. Subsequent methods seek less redundant or more selective access to visual information. MASSV reuses target vision-encoder outputs through a lightweight projector (Ganesan et al., 2025); Huang et al.’s SpecVLM employs an elastic visual compressor (Huang et al., 2025); and ViSpec combines a compact vision adaptor with persistent global-feature injection (Kang et al., 2025). Other methods increasingly reuse representations already computed by the target. At inference time, DREAM cross-attends to cached target features while compressing the visual context; during training, it uses entropy-guided selection of intermediate target representations (Hu et al., 2025b). HiViS removes visual tokens from the drafter input and transfers visual semantics through target-derived hidden representations (Xie et al., 2026). TIGER makes this access step-dependent by routing visual tokens according to the current textual state and trains the drafter with a verifier-derived accepted-prefix reward (Vo et al., 2026). This progression moves from duplicating explicit visual inputs toward compressed, reused, and selectively exposed visual information. Candidate topology and runtime control provide complementary improvements without necessarily changing future-position dependence. ViSpec also uses a dynamic draft tree, while Spec-LLaVA expands and prunes tree branches according to draft confidence (Huo et al., 2025) and SAGE adapts tree depth and width using online entropy (Tong et al., 2026). ViSkip instead uses vision-aware signals to adapt when speculative decoding is attempted (Shen et al., 2026). Trees broaden the can- didate space, whereas ViSkip controls the allocation of speculative computation; neither mechanism alone removes the serial dependence of the underlying drafter. SpecFLASH changes candidate generation more directly. Its semi-autoregressive drafter processes K placeholder tokens and predicts the next K token distributions in one pass (Wang et al., 2025b). We classify it as L1 because multiple fixed future offsets are predicted in parallel from a shared accepted context, rather than generated through a jointly evolving block state. Block-parallel VLMs provide direct L2 examples. Fast-dVLM converts an autoregressive VLM into a KV-cache-compatible block-diffusion model and supports self-speculative block decoding (Wu et al., 2026). Nemotron-Labs-Diffusion jointly trains autoregressive and diffusion objectives within a shared architecture; its VLM variants can draft through diffusion and verify autoregressively (Fu et al., 2026). In both cases, the future block becomes the draft-side generative object rather than a collection of independently attached offset predictions. 4 Table 1: Taxonomy of representative multimodal speculative decoding methods across application domains, drafter designs, and speculative mechanisms. MethodDrafterMechanism LConditionCandidate formAcceptanceAdaptation Vision–Language Models SPD-MLLM (2024)L0Text / directChainExactStatic MASSV (2025)L0Target feat.ChainExactStatic DREAM (2025b)L0Target feat.ChainExactStatic HiViS (2026)L0Target feat.ChainExactStatic SpecVLM (image) (2025)L0Direct / target feat.Chain / treeExactCondition ViSpec (2025)L0DirectTreeExactCandidate TIGER (2026)L0DirectChainExactCondition Spec-LLaVA / SAGE (2025; 2026)L0DirectTreeExactCandidate ViSkip (2026)–Target feat.InheritedInheritedExecution SpecFLASH (2025b)L1Direct / target feat.Parallel offsetsExactStatic Fast-dVLM (2026)L2DirectJoint blockExact (greedy)Static Nemotron-Labs-Diffusion- VLM (2026) L2DirectJoint blockExact (greedy)Static Video–Language Models Sparse-to-Dense (2025)L0DirectChainExactCandidate SpecVLM (video) (2025)L0Direct / target feat.TreeExactCondition Sparrow (2026b)L0Target feat.TreeExactStatic HIPPO / ParallelVLM (2026; 2026)L0Direct / target feat.ChainExactCondition / execution LVSpec (2026)L0DirectChainRelaxed (position)Verification Axis 1: Vision–Language–Action Models Spec-VLA (2025a)L0DirectChainRelaxed (distance)Static HeiSD (2026b)L0DirectChain / retrievalRelaxed (sequence)Candidate Reasoning-aware SD (diff.) (2026)L2Target feat.Reasoning blockExact (greedy)Static Realtime-VLA FLASH (2026)L2DirectAction blockConsistencyExecution Speech and Audio Models Speech SD (2025b)L0Text / directChainRelaxed (tolerance)Static SpecASR (2025)L0DirectTreeExactCandidate ParaASR (2026)L1DirectParallel offsetsExactStatic VADUSA (2025a)L1Text / directOffsets / treeRelaxed (tolerance)Static CTC-SSD (2026)L1 ∗ DirectFrame hypothesisConfidence-gatedVerification / execution PCG (2026)L0Text / directChainGroup-exactStatic Autoregressive Visual Generation LANTERN (2025a)L0Text / directChainRelaxed (latent)Static LANTERN++ (2025b)L0Text / directTreeRelaxed (latent)Static Spatial SD (2026)L1TextParallel rowsCorrection-basedStatic SDVG (2026)L2TextVideo blockQuality-gatedExecution Speculative Jacobi (2025b)L2 ∗ Text / directIterative blockConvergenceStatic Notation. “Direct” denotes direct non-text conditioning; “target feat.” denotes reused target features. P-level measures dependence across future draft positions, independent of candidate form: L0 = sequential, L1 = multi-position parallel prediction, L2 = block-parallel generative. L1 ∗ parallelizes frame labels; L2 ∗ denotes target self-refinement (block-parallel but not an independent drafter). “Static” indicates no runtime adaptation. ViSkip inherits the wrapped method’s attributes; “group-exact” preserves the induced group variable. Rows are representative, not exhaustive. The VLM literature thus spans L0 systems centered on visual access and candidate coverage, L1 multi-position prediction, and L2 block-parallel generative drafting. Current L2 evidence primar- ily comes from co-designed models in which the same backbone is explicitly trained to support block-parallel drafting and autoregressive verification. It therefore demonstrates block-level self- speculation in VLMs, while remaining distinct from an independent, low-cost parallel drafter that can be paired with heterogeneous pretrained VLM targets. Video–language models. Video-language models amplify the condition-cost problem because long visual prefixes enlarge draft-side prefill, attention, and KV-cache costs. A small drafter may also lose accuracy when relevant evidence is diluted across the video context. Video speculative decoding has consequently focused on reducing condition-processing cost and overlapping system execution, while candidate generation remains largely autoregressive. Sparse-to-Dense (StD) reduces draft computation without changing the input representation: it uses sparse top-K attention during drafting and restores dense attention for target verification (Zhang et al., 2025). SpecVLM instead sparsifies the condition presented to the drafter through verifier- 5 A. Vision-Language Models B. Video-Language Models C. Vision-Language- Action Models D. Speech and Audio Models E.Autoregressive Visual Generation 2026 2025 2024 L2 Nemotron-Labs-Diffusion-VLM L2 Fast-dVLM - ViSkip L0 HiViS L0 TIGET L1 SpecFLASH L0 Spec-LLaVA/SAGE L0 ViSpec L0 SpecVLM(image) L0 DREAM L0 MASSV L0 SPD-MLLM L0 LVSpec L0 HIPPO/ParallelVLM L0 Sparrow L0 SpecVLM(Video) L0 Sparse-to-Dense L2 Realtime-VLA FLASH L2 Reasoning-aware SD(diff) L2 HeiSD L2 Spec-VLA L1* CTC-SSD L1 ParaASR L0 PCG L1 VADUSA L1 Speech SD L1 SpeecSpecASRh SD L1 Spatial SD L2 SDVG L2* Specelative Jacobi L0 LANTERN++ L0 LANTERN Application Domains A. Vison-Language Models B. Video-Language Models C. Vision-Language-Action Models D. Speech and Audio Models E. Autoregressive Visual Generation Parallel Drafting Level L0 Sequential Drafting(next-token) L1 Multi-position parallel prediction L2 Block-parallel generative drafting L1* Special variant(L1) L2* Special variant(L2) - Inherited / unspecified Figure 1: Taxonomy and evolution of representative multimodal speculative decoding methods. guided and spatially uniform token pruning, while the target verifies against the full video context (Ji et al., 2025). The two systems thus apply sparsity to different objects—draft attention and draft input—while retaining strict target-token verification. Sparrow questions whether the drafter should process video tokens at all. It attributes long-video degradation to visual KV-cache growth, context mismatch, and attention dilution, and instead reuses target-side text states that have already integrated visual information (Zhang et al., 2026b). Its recursive multi-token prediction exposes the drafter to predicted states during training; the deployed draft trajectory remains autoregressive, optionally organized as a token tree. A separate line reduces mutual waiting between draft and target models. HIPPO combines spatial– temporal relevance pruning with an execution policy that overlaps the next proposal and current ver- ification, adapting between optimistic and conservative schedules from recent outcomes (Lv et al., 2026). ParallelVLM similarly combines verifier-guided pruning with overlapped execution and cor- rects positional bias in attention-based token selection (Kong et al., 2026). This pipeline overlap is systems parallelism rather than parallel generation of future draft positions. LVSpec modifies the acceptance contract instead. It applies strict verification to visually grounded anchors, relaxes acceptance for less visually relevant tokens, and allows nearby-position matching between target and draft sequences (Ji et al., 2026). Because this rule need not reproduce the exact target sequence, its reported preservation of more than 99.8% of target task performance is empirical high-fidelity evidence rather than a formal exactness guarantee. Current video-specific methods are therefore concentrated at L0 but obtain acceleration from three distinct sources: condition sparsification or feature reuse, draft–verify pipeline overlap, and visually informed relaxed acceptance. Block-diffusion VLMs do not yet resolve the long-video setting, where a parallel drafter must avoid recreating the target’s condition cost while preserving temporal grounding and competing with L0 systems that already prune, reuse, or hide draft computation. 6 Vision–language–action models. VLA systems place speculative inference inside a closed-loop control pipeline, which creates two verification problems. Under a fixed observation, exact equality between discretized action tokens may be unnecessarily strict because nearby tokens can encode functionally similar controls. Across time, however, even an initially valid action chunk can become unsafe or obsolete as new observations arrive. Verification must therefore address both candidate– target agreement and the continuing validity of an accepted plan. Spec-VLA studies the first problem with a compact autoregressive VLA drafter and accepts a pro- posed action token when its discretized-bin distance from the target token is below a threshold (Wang et al., 2025a). KERV adds kinematics-aware rectification: a Kalman predictor repairs speculative errors, while a runtime signal adjusts the acceptance threshold (Zheng et al., 2026a). HeiSD in- stead routes geometrically irregular segments to a learned VLA drafter and regular segments to retrieved action sequences, using verify-skip and sequence-wise relaxed acceptance on the retrieval path (Zheng et al., 2026b). The learned draft paths remain autoregressive, while retrieval supplies a different candidate source. Correctness is consequently supported by task success and trajectory quality rather than exact reproduction of the target action sequence. SV-VLA addresses the second problem explicitly. A heavy VLA acts as a low-frequency macro- planner that emits an action chunk and planning context, whereas a lightweight verifier repeatedly compares the planned action with a closed-loop reference conditioned on the latest observation and triggers replanning when needed (Wang et al., 2026b). This is not merely a looser token-matching rule: it tests whether an accepted open-loop plan remains valid after the environment and execution history have changed. Its contribution therefore lies in verification and runtime adaptation rather than in a new candidate-generation P-level. Recent work also changes the object generated in parallel. Reasoning-aware speculative decod- ing accelerates the autoregressive chain of causation that precedes trajectory prediction. Its block- diffusion variant is L2 because it proposes a future reasoning block jointly, although the parallelized object is language reasoning rather than control (Dinh et al., 2026). Realtime-VLA FLASH pro- vides direct action-side evidence: a lightweight model drafts a continuous action chunk, the target Action Expert verifies the chunk in parallel, and a phase-aware mechanism falls back to full infer- ence when consistency is insufficient (Niu et al., 2026). We record this joint action-chunk proposal as L2, with consistency-based rather than exact verification. Fast-dDrive supplies a complementary L2 design for autonomous driving. Its block-diffusion VLA refines tokens within scaffolded seman- tic sections, including a trajectory section, and uses scaffold speculative decoding to accelerate the structured output (Zhang et al., 2026a). L2 evidence within the VLA domain thus spans distinct generated objects: reasoning blocks, con- tinuous action chunks, and structured outputs containing trajectories. The first parallelizes upstream language reasoning, whereas the latter two place speculative generation directly on action or trajec- tory outputs. These control-side settings must additionally validate accepted plans under evolving observations, provide safe fallback and recovery, and establish end-to-end gains across control tasks and VLA architectures. Speech and audio models. Speech speculation spans three output settings. ASR produces a tran- script that is strongly constrained by the acoustic input; audio-language systems generate more open-ended text; and speech synthesis emits long acoustic- or codec-token sequences for which dif- ferent tokenizations may decode to perceptually similar waveforms. These settings offer different sources of predictability and require different notions of verification. For ASR, SpecASR uses a smaller audio-conditioned model with adaptive draft length, sequence recycling, and sparse-tree verification (Wei et al., 2025). Its learned draft trajectory remains L0; adaptation and tree width reduce wasted work without changing its serial depth. ParaASR pro- vides direct L1 evidence by training multiple future-transcript-token branches on an audio-language model. It proposes six transcript tokens per decoding step and commits only the prefix verified by the autoregressive head (Lin et al., 2026). CTC-based proposals parallelize a different unit. CTC-SSD reuses the acoustic encoder to predict frame labels simultaneously, collapses them into a complete transcript hypothesis, and either accepts a low-entropy hypothesis, verifies its token likelihoods in one LLM pass, or resumes autoregressive decoding from the accepted prefix (Saon et al., 2026). We denote this as L1 ∗ : it removes output-length-dependent drafting, as L1 intends, but parallelizes acoustic frames rather than directly predicting future transcript positions. SMUD uses a related CTC 7 preliminary transcript and batched decoder comparison to locate regions that still require causal search (Okabe & Yamamoto, 2025). Speech synthesis shifts attention from transcript alignment to perceptual equivalence. Speech Spec- ulative Decoding (Speech SD) employs a lightweight L0 speech LM and adds a heuristic toler- ance factor to its sampling acceptance probability (Lin et al., 2025b). VADUSA instead attaches Medusa-style heads for parallel future speech-token prediction, forms sparse candidate trees, and uses tolerance-based verification; the heads make it a direct L1 speculative TTS design (Li et al., 2025a). Codec-MTP is also L1, but uses several future-token heads with Viterbi path selection rather than longest-prefix tree verification (Nguyen et al., 2025). These relaxed or path-level procedures are supported by intelligibility, naturalness, and speaker-similarity evidence, not by preservation of the target token distribution. Principled Coarse-Graining (PCG) gives this distinction a more formal treatment: it forms overlapping Acoustic Similarity Groups from the target embedding space and performs rejection sampling over the induced group distribution (Yanuka et al., 2026). Its exactness guarantee applies to the group variable, not to the target’s original token sequence, so perceptual evaluation remains necessary when substituting tokens within a group. For open-ended audio-conditioned text, UGSD drafts speech-emotion captions on an edge model and escalates only uncertain blocks to a cloud verifier, making runtime adaptation trade off latency, communication, and privacy (Xue et al., 2026). Adjacent block-generation work provides enabling evidence for speech-side L2. Chatterbox-Flash is not itself a draft–verify speculative system: it converts an autoregressive TTS decoder into a block-diffusion decoder that generates speech tokens in parallel within streaming blocks (Seo et al., 2026). It therefore demonstrates an enabling block-generation capability rather than a complete L2 speculative pipeline. What remains unresolved is whether such a generator can serve as a sufficiently cheap proposal model for an autoregressive target, and how its blocks should be verified under an exact or perceptually justified contract without erasing the latency saved by parallel drafting. Autoregressive visual generation. In autoregressive visual generation, the speculative object is a spatial or spatiotemporal representation, including discrete image codes, continuous latents, and video blocks. Exact identity may be unnecessarily restrictive when several codes decode to sim- ilar content, yet a local substitution can propagate through later regions. The literature therefore explores both alternative candidate-construction mechanisms and output-aware verification; only some methods parallelize proposal generation. For discrete visual tokens, LANTERN accepts candidates over neighborhoods in the latent codebook and bounds the resulting distributional shift in total variation (Jang et al., 2025a). LANTERN++ combines this relaxed contract with a static tree to avoid shallow drafts under low-confidence visual distributions (Jang et al., 2025b). Both retain a L0 autoregressive proposal trajectory; their main innovations lie in visual equivalence and candidate coverage rather than future-position parallelism. Geometry-aware parallel decoding provides adjacent evidence. ZipAR is a target-side parallel de- coder that predicts row- and column-direction tokens concurrently using adaptive local windows and rejection sampling analogous to speculative decoding (He et al., 2025); it is not a conventional drafter–target speculative system. Spatially Speculative Decoding (Spatial SD), by contrast, learns lightweight horizontal and vertical latent heads that draft subsequent rows in parallel, followed by target-side auto-correction (Xiang et al., 2026). These fixed-offset heads constitute L1: they pre- dict many positions per pass without jointly modeling the entire future block. MuLo-SD supplies a complementary coarse-to-fine design, using a low-resolution drafter and restricting rejection and resampling to local neighborhoods rather than discarding the full raster-order suffix (Peruzzo et al., 2026). Beyond discrete codes, Continuous Speculative Decoding derives acceptance and rejection procedures for the diffusion distributions of continuous-valued visual AR models, with trajectory alignment improving draft–target agreement (Wang et al., 2024a). VVS instead reduces target cost by dynamically skipping selected verification rounds and reusing cached token features (Dong et al., 2026). Speculative Jacobi Decoding (SJD) follows a drafter-free route: the target iteratively updates a fu- ture token window and commits a prefix under a probabilistic convergence criterion (Teng et al., 2025b). We annotate it as L2 ∗ because it refines a parallel block but does not use a separate low- cost drafter (the asterisk denotes block-parallel refinement without an independent drafter). GSD 8 inherits this SJD trajectory and replaces token-level verification with relaxed acceptance over dy- namically constructed groups of visually valid codes (So et al., 2025); its contribution therefore lies primarily in the acceptance contract. SJD2 and SJD++ respectively stabilize refinement through next-clean-token prediction and high-confidence token retention (Teng et al., 2025c;a), while Spec- ulative Coupled Decoding couples successive samples to improve stability without changing the target sampling distribution (So et al., 2026). This application domain is also extending to video. SDVG uses a smaller denoising model to propose continuous spatiotemporal blocks for a block-autoregressive video target and verifies them with an image-quality router (Hu & Zhang, 2026). It is a L2 block-proposal design, but uses a quality-gated rather than exact distributional contract. Overall, visual generation spans L0 autoregressive drafting, geometry-aware L1 prediction, L2 block proposals, and L2 ∗ target self-refinement. Across these forms, the output-level correctness claim remains inseparable from the reported efficiency gain. 2.3CROSS-DOMAIN ANALYSIS Table 2 condenses the preceding review along three cross-domain dimensions: candidate-generation parallelism, acceptance semantics, and domain-specific cues for adapting speculative computation. Across domains, multimodal conditioning and output structure change both the cost profile of spec- ulation and the information available for adapting it. Domain-specific cues can act on condition access, candidate construction, verification, or system execution. For example, relevance can de- termine which condition features are exposed, while uncertainty can govern whether speculation or verification should continue. Such adaptations can reduce condition exposure, redundant compu- tation, or wasted verification without changing the dependence among future candidate positions. They are therefore complementary to candidate-generation parallelism rather than evidence of a higher L-level. Acceptance contracts vary more directly with the generated object. Text-output systems can re- tain token-level target equivalence, whereas action, speech, and visual outputs also admit func- tional, sequence-level, group-level, or perceptual notions of validity. Exact contracts provide target- distribution or greedy-output equivalence under their stated assumptions. Relaxed contracts instead require empirical fidelity evidence in addition to any formal bound on their deviation from the tar- get. This distinction becomes more consequential for block candidates, whose usefulness depends on whether they can be checked under an acceptance contract appropriate to the output space. Across the reviewed domains, candidate generation, condition access, verification, and runtime adaptation have developed along different paths. Their effects are coupled in reported systems, leav- ing unresolved what is gained—and what new costs arise—when candidate generation moves from serial proposals to jointly generated or refined blocks. Addressing this question requires separating future-position dependence from domain-specific optimizations. Section 3 therefore formalizes the progression from sequential drafting to L2 block generation and examines block-parallel generative drafting within that progression. Section 4 then evaluates this mechanism empirically in multimodal inference. 9 Table 2: Cross-domain synthesis of multimodal speculative decoding. L-level entries report repre- sentative candidate-generation mechanisms rather than exhaustive coverage within each domain. ApplicationDomain constraintDraft profileAcceptance semantics Domain- specific cue VLMGrounding–cost trade-off L0 / L1 / L2Token exactnessVisual relevance Video-LMLong-context costL0Token exactnessTemporal relevance VLAClosed-loop stateL0 / L2Functional validityControl phase Speech/AudioTemporal dependencies L0 / L1Sequence / perceptual Prediction uncertainty Visual GenerationSpatiotemporal coupling L0 / L1 / L2 (∗) Group / qualitySpatial locality Notation. L2 (∗) indicates that both block-level proposals and L2 ∗ target-side self-refinement are represented. 3EVOLUTION OF DRAFTER-SIDE PARALLELISM 3.1FROM SEQUENTIAL TO PARALLEL DRAFTING Early speculative decoding uses a small autoregressive model to propose future tokens (Leviathan et al., 2023; Chen et al., 2023). The EAGLE series instead predicts target-model features and maps them back to token distributions (Li et al., 2024a;b; 2025b). These methods improve draft quality and can expand multiple candidate branches, but the underlying draft trajectory remains sequential: moving to a deeper future position requires another causally dependent drafter forward. A different line of work reduces this sequential drafting depth. Medusa and multi-token prediction predict several future positions in parallel (Cai et al., 2024; Gloeckle et al., 2024). More generally, parallel drafters allow one drafter forward to advance several future positions rather than only the next one. Block-parallel generative drafting takes a further structural step. DFlash and DSpark organize the main drafting computation around an entire future block, allowing its positions to be processed in parallel (Chen et al., 2026; Cheng et al., 2026). This progression motivates a L0–L2 taxonomy of speculative drafting. The taxonomy characterizes drafter-side parallelism: how far candidate generation can advance along the future sequence within one drafter forward. It is independent of how many alternative candidates are retained or how they are verified. 3.2A L0–L2 TAXONOMY OF SPECULATIVE DRAFTING We first define the terminology used in the taxonomy. Consider one speculation round that proposes K future positions. A candidate trajectory is an ordered sequence of proposed future tokens. Its draft depth denotes the position along this trajectory: depth 1 corresponds to the next token, depth 2 to the following token, and so on. A drafter forward denotes one invocation of the main drafter network. Finally, candidate width denotes the number of alternative candidates maintained at the same draft depth. The L0–L2 taxonomy captures a fundamental progression: token generation → parallel position prediction → block generation. L1 parallelizes prediction across future positions; L2 changes the generative unit from individual future positions to the future block itself. L0: Sequential drafting. L0 advances a candidate trajectory one depth at a time. To propose K consecutive future positions, the drafter therefore performs K causally dependent forwards. Small autoregressive language models and the EAGLE family belong to this regime. EAGLE may produce several alternative candidates at one depth, but these candidates increase width, not depth: reach- 10 Context Tokens Draft Model 풚 풕"ퟏ （a）L0: Sequential drafting 풙 ퟏ 풙 ퟐ 풙 ퟑ ... 풙 풕 TargetModel DraftToken (Verify Draft Token) Append to Context Context Tokens Draft Model （b）L1: Multi-depth parallel drafting 풙 ퟏ 풙 ퟐ 풙 ퟑ ... 풙 풕 Append to Context Shared drafter backbone / multiple heads 풚 풕"ퟏ 풚 풕"ퟐ ... 풚 풕"풌 DraftTokens TargetModel (Verifyinparallel) （c）L2: Block-diffusion drafting Context Tokens 풙 ퟏ 풙 ퟐ 풙 ퟑ ... 풙 풕 M ... M 푲Mask Tokens Diffusion-based Draft Model 풚 풕"ퟏ 풚 풕"ퟐ ... 풚 풕"풌 DraftToken Block TargetModel (Verifyinparallel) 풚 풕"ퟏ 풚 풕"ퟐ ... 풚 풕"풋 AcceptedProposalTokens Jump푗stepforward 풚 풕"풋"ퟏ SingleStepForwardbeyond AcceptedProposalTokens Merge Append to Context Figure 2: Illustration of the L0–L2 taxonomy of drafter-side parallelism. L0 performs sequential drafting with one future position advanced per drafter forward; L1 predicts multiple predefined future positions in parallel from a shared prefix; and L2 treats the future block as a jointly generated state, enabling block-parallel drafting followed by parallel target verification. ing the next depth still requires another drafter forward. Mathematically, L0 produces a sequence through repeated application: p(x t+1 ), p(x t+2 | x t+1 ), . . . , p(x t+K | x t:t+K−1 ). L1: Multi-position parallel drafting. L1 breaks the one-forward–one-depth constraint by pre- dicting several predefined future positions in parallel from shared prefix features. The definition is intentionally architecture-agnostic: parallelism may be implemented with multiple prediction heads, multi-token prediction, grouped prediction, semi-autoregressive chunks, or other parallel drafting mechanisms. What defines L1 is not the specific architecture, but that several future depths are predicted independently from a shared context rather than generated as a jointly evolving block. Mathematically, L1 produces marginals p(x t+1 ), p(x t+2 ), . . . , p(x t+K ) from the accepted prefix, without joint block-level state updates. L2: Block-parallel generative drafting. L2 identifies a block-native form of parallel drafting. Rather than treating future positions only as several predictions to be made in parallel, the drafter takes the future block itself as the basic generative unit. The entire block evolves jointly through refinement steps: X (0) t:t+K → X (1) t:t+K → · → X (S) t:t+K , where S is bounded and does not scale linearly with K. Block diffusion, exemplified by DFlash and DSpark (Chen et al., 2026; Cheng et al., 2026), is the representative realization considered in this paper, but L2 is not limited to diffusion architectures. Lightweight dependency modeling inside the block does not change this classification; L2 is defined by the block-parallel organization of the main drafting computation. The distinction between L1 and L2 is therefore structural rather than a simple count of outputs per forward. Both may advance several future depths in one forward. L1 predicts several predefined future offsets from a shared prefix, whereas L2 treats the future block as a jointly evolving generative state. Candidate width is orthogonal to draft depth. Draft parallelism should not be confused with candidate branching. Candidate width measures how many alternatives are retained at a given depth; draft depth measures how far the candidate trajectory extends into the future. Tree-based speculative decoding increases the former. SpecInfer introduces token-tree verification (Miao et al., 2024), Sequoia optimizes tree topology (Chen et al., 2024), and DDTree constructs a budgeted tree from 11 Table 3: A L0–L2 taxonomy of drafter-side parallelism. K denotes the number of future positions considered in one speculation round. The taxonomy describes how candidate trajectories are gener- ated; candidate width, tree construction, and target-side verification are orthogonal. LevelDrafting regimeCore definitionRepresentative methods L0Sequential draftingOne forward advances one draft depth; reaching depth K requires K causally depen- dent forwards Small AR LM; EAGLE L1Multi-position parallel draftingOne forward predicts several predefined future positions in parallel from a shared pre- fix, but does not treat the fu- ture block as a jointly evolv- ing generative state Medusa; MTP; SpecFLASH L2Block-parallelgenerative drafting The future block is the ba- sic generative state; its po- sitions are jointly generated or refined with bounded serial drafting depth DFlash; DSpark block-parallel predictions (Ringel & Romano, 2026). These mechanisms do not by themselves change the L level of the underlying drafter. An EAGLE drafter remains L0 when its candidates are organized into a tree, while DFlash combined with DDTree remains L2. In short, L0 advances one draft depth per forward, L1 predicts multiple future positions in parallel, and L2 organizes parallel drafting around an entire future block as a jointly generative state. 3.3THE MULTIMODAL L2 GAP Section 2 shows that multimodal speculative decoding has advanced rapidly in condition compres- sion, target–drafter alignment, candidate construction, and modality-aware verification. Yet most existing methods remain within L0 or L1 drafting. Evidence for L2—block-parallel generative draft- ing that treats the future block as the jointly evolving generative state—remains limited. In contrast, text speculative decoding already provides concrete L2 examples such as DFlash and DSpark. Text speculative decoding has entered L2; multimodal speculative decoding largely has not. We call this the multimodal L2 gap. This gap does not imply that multimodal models are inherently unsuitable for block-parallel gener- ative drafting. Multimodal conditioning instead changes the tradeoff. Accessing visual, audio, or other modality features can increase draft cost, while the additional condition may also make fu- ture tokens more predictable. A useful multimodal L2 drafter must therefore exploit the condition well enough to produce longer or higher-quality candidate blocks while preserving the efficiency of block-parallel drafting. This motivates the controlled experiments in Section 4. Following prior speculative-decoding sur- veys (Xia et al., 2024; Hu et al., 2025a), we evaluate both draft quality and system efficiency. We report average accepted length, token acceptance rate, target-token rank, and full-block acceptance, together with stage latency, throughput, decode-only and end-to-end speedup, memory, and energy where available. 4IS MULTIMODAL SPECULATIVE DECODING READY FOR L2? 4.1EVALUATION QUESTIONS AND EXPERIMENTAL PROTOCOL Sections 2 and 3 show that multimodal speculative decoding has made substantial progress in con- dition compression, candidate construction, and verification strategies, while the transition from se- quential or multi-position drafting to L2 block-parallel generative drafting remains unclear. Unlike text-only models, multimodal systems introduce additional factors, including heterogeneous archi- 12 tectures, expensive modality conditioning, task-dependent predictability, and system-level overhead. Therefore, we study L2 readiness through a diagnostic perspective: rather than asking whether a block-parallel drafter can be constructed, we ask under which model, task, condition, and system configurations it can translate draft quality improvements into practical end-to-end acceleration. We organize our empirical study around four evaluation questions: Q1: Model compatibility. Can existing multimodal architectures support effective L2 block- parallel drafting? We evaluate whether lightweight L2 drafters can achieve sufficient draft–target agreement across different multimodal model families, scales, and training paradigms. This analy- sis examines whether L2 readiness is a general property of multimodal models or depends on specific target architectures and drafter training strategies. Q2: Conditioning efficiency. Does multimodal conditioning remain the dominant bottleneck after draft generation becomes parallel? Although L2 drafting substantially reduces the cost of generating future tokens, multimodal drafters may require additional computation to access visual or other modality information. We analyze the latency contribution of conditioning, and investigate whether reducing or removing explicit modality access can preserve draft quality while improving end-to-end efficiency. Q3: Task and input dependence. When do improved draft quality and accepted length translate into practical acceleration? Multimodal generation workloads differ significantly in output uncer- tainty, grounding requirements, and input complexity. We therefore evaluate L2 drafting across diverse tasks and input conditions to identify when the accepted-token gain is sufficient to amortize additional speculative overheads. Q4: System readiness. Is the current inference ecosystem mature enough to support practical mul- timodal L2 deployment? Beyond algorithmic effectiveness, practical adoption requires integrated support for drafter training, candidate generation, verification, and serving frameworks. We summa- rize the current ecosystem support for representative L2 methods and discuss remaining deployment barriers. Based on these questions, we conduct a cross-model evaluation covering different multimodal archi- tectures and task categories. We report both draft quality and system efficiency metrics, including mean accepted tokens (MAT), token acceptance rate, target-token rank, block acceptance, stage-level latency, throughput, memory consumption, and end-to-end speedup. Unless otherwise specified, all experiments use matched decoding configurations and backend-specific autoregressive baselines to ensure fair comparison. We use a common protocol to evaluate the L2 transition across model fam- ilies and tasks. Numerical cells are deliberately marked “TBD” until the final reproducible export; we do not infer or estimate missing measurements. 4.2EXPERIMENTAL SETUP Models and implementation. Our evaluation suite comprises four vision-language models se- lected to span diverse configurations along three orthogonal axes: model scale, architectural family, and pretraining regime. Concretely, we include Qwen3-VL-4B and Qwen3-VL-8B (Bai et al., 2025), which are natively pretrained as multimodal dense models; Qwen3.6-27B (Qwen Team, 2026a), a dense model initialized from text-only pretraining and subsequently aligned to multimodal inputs; and Qwen3.6-35A3B (Qwen Team, 2026b), a Mixture-of-Experts (MoE) model that likewise un- dergoes text pretraining followed by multimodal alignment. This selection spans parameter counts from 4B to 27B (active parameters for MoE) and covers both end-to-end multimodal pretraining and post-hoc multimodal alignment, mitigating the risk that conclusions are confounded by a sin- gle architectural choice or training recipe. For targets with publicly released speculative decod- ing checkpoints—most notably Qwen3.6-27B—we compare up to five decoding configurations: (1) target-only autoregressive (AR) decoding, which serves as the speed and quality baseline; (2) EAGLE-3 (Li et al., 2025b) with the PRISM-EAGLE3 drafter checkpoint released by Ex0bit (L0); (3) the native multi-token prediction (MTP) heads (Gloeckle et al., 2024) shipped with Qwen3.6 (L1); (4) DFlash (Chen et al., 2026) with the publicly released z-lab checkpoint (L2); and (5) https://huggingface.co/Ex0bit/Qwen3.6-27B-PRISM-EAGLE3 https://huggingface.co/z-lab/Qwen3.6-27B-DFlash 13 DSpark (Cheng et al., 2026) with the publicly released satgeze checkpoint (L2). For targets without publicly available L2 checkpoints—in particular Qwen3-VL-8B—we additionally train a DFlash- style block-parallel drafter using the SpecForge training framework (Li et al., 2026), ensuring that our evaluation is not limited to targets for which pre-trained drafters happen to be publicly released. Unless otherwise noted (see † in Table 4), all experiments are conducted on the SGLang backend under matched greedy decoding settings, with a maximum generation length of 2,048 new tokens and a fixed random seed of 42. All methods employ identical prompts, tokenizers, stopping criteria, and batch sizes. We report end-to-end wall-clock speedup, which encompasses vision encoding, prefilling, drafter candidate generation, tree construction, and target model verification. Tasks and evaluation matrix. We construct a 600-sample evaluation set following the task bal- ance of MMSpec (Shen et al., 2026), a dedicated benchmark for multimodal speculative decod- ing. Concretely, we randomly sample 100 instances from each of six task categories using a fixed random seed of 42: general VQA from GQA (Ainslie et al., 2023), image captioning from Flickr30K (Plummer et al., 2015), text-focused VQA from TextVQA (Singh et al., 2019), chart un- derstanding from CharXiv (Wang et al., 2024b), complex multimodal reasoning from MMMU (Yue et al., 2024), and multi-turn conversation from the combined ConvBench (Liu et al., 2024) and M- MT-Bench (Agrawal et al., 2024) splits. The only deviation from MMSpec is that we replace the COCO captioning split with Flickr30K to reduce lexical overlap with common VLM pre-training corpora. Baselines and metrics. The primary baseline is target-only autoregressive decoding executed within the same inference backend and under identical decoding settings. Candidate quality is mea- sured by mean accepted tokens (MAT), defined as the average number of consecutive draft tokens accepted by the target per speculation round. The primary efficiency metric is end-to-end wall-clock speedup, which encompasses vision encoding, target prefill, drafter prefill, candidate generation, tree construction, target verification, and token sampling. Speedup is computed against the matched target-only baseline: Speedup = T AR T method .(2) The experiments that follow answer the four questions through three empirical analyses: cross- model effectiveness evaluation, conditioning and input-scaling analysis, and ecosystem support analysis. 4.3DOES L2 BLOCK-PARALLEL DRAFTING DELIVER MULTIMODAL SPEEDUPS? We begin with the foundational effectiveness question: does L2 block-parallel drafting brings the acceptance-length and latency gains observed in text-only speculative decoding to multimodal au- toregressive models, or does the additional draft-side computation and multimodal conditioning overhead negate these benefits? To guard against conclusions that are artifacts of a single archi- tecture, we evaluate across the four-model suite described above, which jointly varies model scale (4B–27B total parameters), architectural family (dense vs. MoE), and pretraining regime (native multimodal vs. text-pretrained then multimodal aligned). For Qwen3-VL-8B, where no public L2 drafter checkpoint is available, we train a DFlash-style block-parallel drafter via SpecForge (Li et al., 2026) to avoid selection bias toward targets with pre-existing third-party checkpoints. We comple- ment these head-to-head comparisons with a general bench collected in 4.2 —to identify the specific conditions under which L2 drafting succeeds or fails. Table 4 extends the Qwen3.6-27B comparison to the Qwen3-VL family. On Qwen3.6-27B, under a matched SGLang backend, evaluation examples, and decoding protocol, DFlash achieves the highest MAT and speedup on every benchmark, with a task-equal MAT of 4.38 and an average speedup of 2.60× over AR; DSpark ranks second with 3.52 MAT and 2.04× speedup. On the matched HuggingFace (HF) subsets † , DFlash reaches 2.50 MAT and 1.95× speedup for Qwen3-VL-4B, and 2.75 MAT and 2.14× speedup for Qwen3-VL-8B. By contrast, the SGLang EAGLE-3 runs remain below their matched SGLang AR baselines, at 0.71× and 0.88× overall, respectively. Because the https://huggingface.co/satgeze/Qwen3.6-27B-DSpark 14 Table 4: Speculative decoding performance with a maximum of 2,048 new tokens. MAT denotes the mean number of output tokens advanced per speculative step. Overall results are task-equal means. The best and second-best speculative results for each target are shown in bold and underlined, respectively. † Results measured on the HuggingFace Transformers backend; all other results use the SGLang backend. Each speedup is normalized to its own backend-matched autoregressive baseline. We abbreviate target models Q denotes Qwen.zheli TargetMethod GQAF30KTextVQACharXivMMMUMTOverall MAT Speedup MAT Speedup MAT Speedup MAT Speedup MAT Speedup MAT Speedup MAT Speedup Q3VL-4B AR–1.00×–1.00×–1.00×–1.00×–1.00×–1.00×–1.00× EAGLE-3 1.950.71×1.960.67×1.980.71×2.280.76×2.410.81×2.140.60×2.120.71× DFlash † 3.052.28×2.842.36×2.051.57×2.632.06×2.281.74×2.131.70×2.501.95× Q3VL-8B AR–1.00×–1.00×–1.00×–1.00×–1.00×–1.00×–1.00× EAGLE-3 2.220.85×2.200.85×2.050.81×2.660.99×2.480.95×2.170.85×2.300.88× DFlash † 2.451.78×2.892.18×2.531.87×3.452.73×3.092.55×2.081.72×2.752.14× Q3.6-27B AR–1.00×–1.00×–1.00×–1.00×–1.00×–1.00×–1.00× EAGLE-3 2.371.74×2.191.64×2.381.70×2.501.83×2.231.64×2.201.62×2.311.69× MTP3.221.64×3.161.63×3.261.64×3.391.72×3.321.71×2.991.50×3.221.64× DFlash4.192.42×4.302.12×4.162.53×5.123.19×4.353.01×4.172.30×4.382.60× DSpark3.271.92×3.261.69×3.361.98×4.422.53×3.442.34×3.351.76×3.522.04× Qwen3-VL EAGLE-3 and DFlash measurements use different backends, their absolute latency and throughput values are not compared directly; each reported speedup is normalized only to its own backend-matched AR run. Four observations stand out. First, drafter capacity appears to raise the attainable acceptance ceiling: DFlash uses approxi- mately 1.73B parameters, compared with approximately 0.43B for Qwen3.6’s native MTP module and 0.60B for EAGLE-3, and produces substantially longer accepted sequences. However, EAGLE- 3 does not outperform the smaller MTP module, showing that scale alone is insufficient and that drafting architecture and training objective also matter. Second, checkpoint provenance matters. Despite using a larger drafter, the third-party DSpark checkpoint trails DFlash on all six multimodal tasks. This gap contrasts with the strong text-only results reported in DSpark (Cheng et al., 2026), suggesting that text performance does not reliably predict multimodal acceptance quality. Third, natively multimodal-trained models better support speculative decoding. Qwen3.6-27B, which is natively trained with multimodal data, consistently benefits from speculative decoding across different methods, achieving 1.69× with EAGLE-3, 2.04× with DSpark, and 2.60× with DFlash on the SGLang backend. In contrast, on the Qwen3-VL family, EAGLE-3 fails to provide speedup on SGLang (0.71× for 4B and 0.88× for 8B). Even when we explicitly train a DFlash- style block-parallel drafter for Qwen3-VL-8B with multimodal data using SpecForge, it reaches only 2.14× speedup on HF, below the 2.60× achieved by the off-the-shelf z-lab DFlash checkpoint on Qwen3.6-27B. These results suggest that native multimodal training may induce a more uni- fied feature distribution and better alignment between multimodal representations and future-token prediction, making the target model inherently easier for speculative drafters to approximate than adding multimodal signals only during post-hoc drafter training. Fourth, larger target models benefit more from speculative decoding. We observe a clear im- provement as the target model grows. For DFlash, speedup increases from 1.95× on Qwen3-VL-4B to 2.14× on Qwen3-VL-8B, and reaches 2.60× on Qwen3.6-27B. EAGLE-3 shows an even sharper transition, from no net acceleration on 4B (0.71×) and 8B (0.88×) to 1.69× on 27B. Two fac- tors likely contribute. First, DFlash uses a shallow, roughly fixed-cost 5-layer drafter, while the target-model forward cost grows substantially with model size; the drafting overhead is therefore increasingly amortized by each accepted block. Second, larger targets also appear easier to specu- late: DFlash MAT rises from 2.50 and 2.75 on 4B and 8B to 4.38 on 27B. This suggests that larger 15 models may expose more stable and predictable feature distributions, allowing the drafter to better match future-token representations. Larger targets thus benefit from both a more favorable compute ratio and higher draft predictability. 4.4DOES MULTIMODAL CONDITIONING REMAIN A BOTTLENECK UNDER BLOCK PARALLELISM? 0510152025303540 Latency (ms) EAGLE3 (L0) MTP (L1) DSpark (L2) DFlash (L2) 26.28.78 0.81 35.8 26.58.74 1.36 36.6 26.510.18 1.78 38.5 27.49.80 1.14 38.3 Vision ConditioningDraftVerification Figure 3: Stage-level latency decomposition of a representative decoding pass on Qwen3.6-27B. Vision conditioning includes vision encoding, multimodal prefill, and drafter-side target-feature-to- KV construction when applicable. Table 5: HR-Bench results under 4K and 8K resolutions on Qwen3-VL, with a 512-token generation cap on a common 30-sample manifest. MAT denotes the mean number of output tokens advanced per speculative step. Speedup is end-to-end latency speedup over the backend-matched autoregres- sive baseline. Top1 denotes top-1 acceptance; Sampling denotes speculate sampling. Resolution Qwen3-VL-4BQwen3-VL-8B Top1SamplingTop1Sampling MATSpeedupMATSpeedupMATSpeedupMATSpeedup HR-Bench 4K2.901.03×2.450.96×3.221.06×2.881.11× HR-Bench 8K2.860.90×2.700.83×3.090.99×3.220.85× Multimodal prefill, not block drafting, bounds end-to-end speedup. A central tension in mul- timodal speculative decoding is the cost of accessing the visual condition. A condition-light drafter is cheap but poorly aligned with visually grounded continuations; a fully conditioned drafter im- proves alignment but may duplicate part of the target-side multimodal prefill. Block-parallel drafting changes this trade-off, but does not remove it. Figure 3 makes this clear. For L2 drafters, draft generation itself is no longer the bottleneck: DFlash spends only 1.14 ms on drafting and DSpark 1.78 ms. This confirms the main benefit of block parallelism: proposing a longer block is cheap. However, the dominant term is still multimodal conditioning. MTP, DSpark, and DFlash spend 26.45–27.40 ms in the vision-conditioning stage, far larger than the draft step and even larger than target verification. DFlash is especially informative here. Its drafter is fast, but it conditions on target hidden states by projecting them into the drafter KV cache. This target-feature-to-KV construction improves draft quality, yet it also introduces a prefill-like cost that is tied to the visual sequence length. Thus, DFlash removes most of the token- drafting cost, but not the cost of preparing the multimodal condition. We further stress this effect on HR-Bench, using Qwen3-VL at 4K and 8K resolutions. Table 5 shows that the drafter still obtains non-trivial MAT, staying around 2.5–3.2 across models and ver- 16 ification rules. However, this does not translate into end-to-end acceleration. At 4K, speedups are already marginal, ranging from 0.96× to 1.11×. At 8K, all settings fall to or below the autore- gressive baseline, with speedups of 0.90× and 0.83× on Qwen3-VL-4B, and 0.99× and 0.85× on Qwen3-VL-8B. The contrast is especially clear under rejection sampling on Qwen3-VL-8B: MAT increases from 2.88 to 3.22 when moving from 4K to 8K, yet speedup drops from 1.11× to 0.85×. Thus, the failure mode is not simply poor draft quality. Rather, higher-resolution inputs increase vi- sual prefill and drafter-side cache-construction cost, while MAT does not grow enough to amortize this added conditioning overhead. These two measurements point to the same conclusion. Under block-parallel speculative decoding, drafting one more token is almost free; conditioning on one more visual token is not. The end-to- end speedup is therefore bounded by multimodal prefill, especially for DFlash-style drafters that additionally convert target features into drafter-side KV cache. Future gains should thus come less from further accelerating the draft forward pass, and more from reducing or reusing the conditioning path: compressed visual tokens, target-feature reuse, shared drafter caches, or lightweight visual- state adapters. Acceptance is robust to context removal. Table 6 further clarifies what kind of bottleneck multi- modal conditioning creates under block parallelism. DFlash acceptance is only weakly sensitive to explicit visual conditioning in the drafter context: masking visual-token K/Vs reduces Overall MAT by only 4.2% on Qwen3-VL-4B and 3.9% on Qwen3-VL-8B. Even removing the entire original prefill context lowers MAT by only 14.1% and 11.0%, respectively. Since the target model always receives the full multimodal input, this ablation isolates the drafter-side source of acceptance. The results suggest that the diffusion-based drafter does not exploit visual evidence in the same way as a full VLM target during autoregressive generation. Instead, a large fraction of accepted draft tokens can be predicted from local continuation statistics and recent contextual states, rather than from detailed re-access to the full question-image input. The degradation is also task dependent rather than uniform. TextVQA is the most sensitive to context removal, especially on Qwen3-VL-8B (2.750 to 1.953 under no-prefill context), whereas MMMU remains comparatively stable and GQA shows no consistent monotonic drop. This pattern suggests that explicit source conditioning matters most when the next-token distribution is tightly constrained by localized visual or OCR evidence, but contributes less when acceptance is dominated by short- range target-token continuation. A larger target also appears slightly more robust to drafter-side context ablation, which may indicate that stronger models expose more stable local continuation patterns for speculative matching. 4.5TRAINING AND INFERENCE FRAMEWORK SUPPORT Beyond algorithmic effectiveness, the practical adoption of L2 drafting also depends on the avail- ability of training and inference infrastructure. Block-parallel drafting requires dedicated support for drafter training, block-level proposal generation, and speculative verification. We therefore sum- marize the current open-source ecosystem for representative L2 methods, focusing on both training frameworks and inference backends. As summarized in Table 7, support for L2 drafting has expanded rapidly beyond method-specific research implementations. The official DFlash repository provides released checkpoints across mul- tiple model families, including Qwen3, Qwen3.5/3.6, and Gemma, together with inference support through Transformers, MLX, SGLang, and vLLM. DeepSpec provides a unified training and evalu- ation pipeline for DFlash and DSpark on Qwen3 and Gemma models. AngelSpec supports several block-parallel architectures, including DFly, DFlash, DFlare, and DSpark, within a unified training framework using vLLM, SGLang, or HuggingFace-based inference backends. Speculators provides DFlash and DSpark training with direct integration into the vLLM ecosystem, while SpecForge pro- vides corresponding support for DFlash, Domino, and DSpark with direct integration into SGLang. The model families listed in the table indicate representative documented configurations rather than exhaustive model–method combinations. At the inference level, vLLM and SGLang provide the most complete production-oriented support for block-parallel speculative decoding. Both frameworks support DFlash-style drafting, and sup- port for newer L2 methods such as DSpark is also being integrated into their speculative-decoding 17 Table 6: Effect of multimodal target context on DFlash speculative acceptance. Results use top-1 verification with 100 examples per task and a maximum generation length of 2,048 tokens. Overall MAT is the equal-weight arithmetic mean over GQA, TextVQA, and MMMU. Text-only context masks visual-token K/V entries in the DFlash context, whereas no-prefill context masks the complete original prefill context. The target model always receives the full multimodal input. ModelDFlash context settingGQA TextVQA MMMU Overall MAT↑ ∆ Overall vs. full Qwen3-VL-8B Full multimodal context 2.6922.7503.2592.900– Text-only context2.6082.5043.2462.786 −0.114 (−3.9%) No-prefill context2.8451.9532.9462.581 −0.319 (−11.0%) Qwen3-VL-4B Full multimodal context 2.5872.3052.5832.491– Text-only context2.4872.2212.4532.387 −0.104 (−4.2%) No-prefill context2.0951.9042.4172.139 −0.353 (−14.1%) Table 7: Representative framework support for L2 speculative decoding. RepositoryModelsL2 MethodsFramework Support DFlashQwen3/3.5/3.6, Gemma4DFlashMLX, SGLang, vLLM DeepSpecQwen3, Gemma4DFlash, DSparkTrain, Eval AngelSpecQwen3, Hunyuan3DFly, DFlash, DFlare, DSparkTrain, vLLM, SGLang SpeculatorsQwen3/3.6, Gemma4DFlash, DSparkTrain, vLLM SpecForgeQwen3/3.6DFlash, Domino, DSparkTrain, SGLang stacks. Transformers provides a more general assisted-decoding interface and has begun to incor- porate model-specific DFlash support. MLX provides an additional execution path for DFlash on Apple Silicon, although its L2 coverage is currently narrower than that of vLLM and SGLang. Overall, L2 speculative decoding is transitioning from isolated research implementations toward in- tegrated training–serving ecosystems. In particular, the Speculators–vLLM and SpecForge–SGLang stacks provide increasingly complete paths from drafter training to efficient inference. However, support remains uneven across model families and is still substantially less mature for multimodal targets than for text-only models. This ecosystem gap constitutes an additional practical barrier to the broader adoption of multimodal L2 drafting. 5L2 READINESS AND FUTURE DIRECTIONS 5.1WHAT DETERMINES L2 READINESS? Taken together, the results in Section 4 suggest that L2 readiness should not be treated as a fixed property of an entire modality or model family. Instead, it depends on the specific target model, multimodal condition, task and input regime, and inference environment. We therefore interpret readiness at the model–task–condition–system level. Model compatibility. L2 drafting is effective across the evaluated targets, but its benefit varies substantially across model families. DFlash achieves positive end-to-end speedups on both Qwen3- VL-4B and Qwen3-VL-8B, while obtaining substantially higher MAT and speedup on Qwen3.6- 27B. These results indicate that multimodal L2 drafting is feasible, but not uniformly effective across targets. Importantly, the observed differences should not yet be attributed to a single architectural factor, since model scale, drafter checkpoint, training recipe, and inference backend also vary across configurations. The current evidence therefore supports model-dependent compatibility rather than an architecture-agnostic notion of L2 readiness. Multimodal conditioning. The strongest multimodal-specific bottleneck is the cost of condition- ing. The latency decomposition in Figure 3 shows that block-parallel draft generation itself is not the dominant cost in our evaluated configurations, whereas multimodal conditioning contributes a substantial portion of the speculative decoding overhead. Our conditioning analysis further suggests 18 that repeatedly exposing the drafter to the full multimodal context is not always necessary for main- taining useful speculative acceptance. A L2-ready multimodal system should therefore provide a lightweight conditioning path that preserves the information needed for drafting without reproduc- ing the target model’s multimodal processing cost. Task and input dependence. Readiness also varies with the workload. Table 4 shows noticeable task-level variation in MAT and end-to-end speedup, indicating that the predictability of future to- kens differs across captioning, VQA, OCR, and reasoning tasks. More importantly, the HR-Bench results show that relatively high MAT does not guarantee acceleration under expensive visual inputs: increasing the image resolution from 4K to 8K preserves MAT at roughly the same level while reduc- ing end-to-end speedup to around or below the autoregressive baseline. Thus, L2 readiness depends not only on draft quality, but also on whether the accepted-token gain is sufficient to amortize the multimodal input cost. Dynamic routing across P levels or block sizes is a promising consequence of this observation, although we leave generation-stage routing to future work. System and ecosystem support. Practical readiness further depends on whether L2 methods can be trained and deployed using existing infrastructure. As summarized in Table 7, DFlash and DSpark are increasingly supported by dedicated training frameworks such as DeepSpec, AngelSpec, Spec- ulators, and SpecForge, while vLLM and SGLang provide increasingly mature serving paths for block-parallel speculative decoding. Transformers and MLX also provide more limited L2 execu- tion paths. This trend shows that L2 is moving beyond isolated research implementations toward integrated training–serving ecosystems. However, support remains substantially less uniform for multimodal targets than for text-only models. Overall readiness. The evidence therefore supports a qualified conclusion: multimodal specula- tive decoding is partially ready for L2 block-parallel drafting. L2 can already provide substantial gains for favorable target models and workloads, demonstrating that multimodality itself is not a fun- damental barrier. However, these gains remain sensitive to target-model compatibility, multimodal conditioning cost, input characteristics, and framework support. The practical question is there- fore not whether multimodal L2 works in general, but under which model–task–condition–system configurations its accepted-token gains are large enough to translate into end-to-end acceleration. 5.2CHALLENGES AND OPPORTUNITIES Lightweight multimodal condition injection. Future drafters should learn the smallest multi- modal representation sufficient for next-block prediction. Promising routes include target-produced semantic summaries, query-based visual compression, layer-adaptive feature selection, and dynamic routing that enables visual conditioning only when token uncertainty indicates that it is needed. The objective should jointly optimize condition sufficiency and added prefill latency. Dependency modeling for L2 drafting. L2 drafters need enough block-internal structure to pre- serve consistency without returning to K serial steps. Lightweight causal correction, confidence- ordered refinement, local remasking, and cross-position state transfer may offer a better latency– quality balance than either fully independent or fully autoregressive drafting. Dynamic Parallel levels and block sizes. No single drafter is likely to dominate all multimodal tasks. A practical system can route between an autoregressive drafter, multi-token heads (L1), DFlash (L2), DSpark (L2), retrieval, and tree expansion according to task type, current entropy, generation stage, and measured break-even cost. When the predicted gain falls below zero, the system should shorten the block or revert to target autoregressive decoding. Tree-aware parallel-drafter training. Position-wise imitation does not directly optimize the ac- cepted prefix or the utility of a finite verification tree. In block-parallel speculative decoding methods like DFlash, where the drafter is heavily conditioned on target model representations (via multi-layer hidden state injection), the target’s top-1 token is frequently covered within the drafter’s top-k candi- dates. For instance, the top-6 draft tokens can already cover 98–99% of the target probability mass, making exact position-wise matching largely unnecessary. Consequently, training objectives should instead reward recoverable target-consistent paths, calibrate marginal probabilities for budgeted tree 19 construction, and concentrate capacity on frontier positions that limit acceptance. This bridges mul- timodal L2 drafting with recent tree-based verification methods, rather than treating the drafter and verifier as independent modules. Multimodal verification and relaxed acceptance. Exact token verification is appropriate for lossless text generation, but actions, codec tokens, and visual codebooks may admit functionally or perceptually equivalent outputs. Future systems should define output-space-specific acceptance rules together with explicit quality guarantees, and should distinguish relaxed verification from merely skipping target computation. Standardized evaluation and system co-design. Parallel drafting changes the workload pre- sented to the target: longer verification sequences and wider trees can improve arithmetic inten- sity but also increase memory traffic and batch interference. A shared benchmark should span architectures, modalities, input lengths, task uncertainty, output lengths, batch sizes, hardware, and backends. It should report candidate and task quality, stage latency, throughput, memory, energy, and both decode-only and end-to-end speedup. Efficient tree attention, KV-cache reuse, overlapped execution, and scheduling should be evaluated under the same protocol so that readiness becomes a reproducible model–task–system measurement. 6CONCLUSION This paper asks whether multimodal speculative decoding is ready to move from L0/L1 drafting to L2 block-parallel generative drafting. Our modality-centered review shows that VLM, Video- LM, VLA, Speech/Audio, and Visual-AR research has already developed strong mechanisms for condition access, token compression, candidate construction, and task-aware verification. How- ever, systematic evidence for L2 block-parallel drafting across multimodal architectures remains limited. Our cross-architecture study directly examines this transition and reveals a more nuanced picture: neither a universal failure nor a universal success. While direct L2 transfer can be weak on some multimodal architectures, newer architectures demonstrate substantially stronger draft–target compatibility. These findings suggest that multimodal L2 readiness is configuration-dependent, de- termined by architectural compatibility, condition accessibility, task predictability, dependency han- dling, and system-level amortization. Therefore, multimodal speculative decoding is partially ready for diffusion-based parallel drafting, but practical deployment still requires careful co-design of the model architecture, conditioning pathway, and inference system. REFERENCES Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jes- sica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebr ́ on, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head check- points. In Proceedings of the 2023 conference on empirical methods in natural language process- ing, p. 4895–4901, 2023. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, and Tri Dao. MEDUSA: Simple llm inference acceleration framework with multiple decoding heads. In Proceedings of the 41st International Conference on Machine Learning. PMLR, 2024. Charlie Chen, Sebastian Borgeaud, Geoffrey Irving, Jean-Baptiste Lespiau, Laurent Sifre, and John Jumper. Accelerating large language model decoding with speculative sampling. arXiv preprint arXiv:2302.01318, 2023. doi: 10.48550/arXiv.2302.01318. URL https://arxiv.org/ abs/2302.01318. 20 Jian Chen, Yesheng Liang, and Zhijian Liu. DFlash: Block diffusion for flash speculative decoding. arXiv preprint arXiv:2602.06036, 2026. doi: 10.48550/arXiv.2602.06036. URL https:// arxiv.org/abs/2602.06036. Accepted at ICML 2026. Zhuoming Chen,Avner May,Ruslan Svirschevski,Yuhsun Huang,Max Ryabinin, Zhihao Jia, and Beidi Chen.Sequoia:Scalable and robust speculative decoding. In Advances in Neural Information Processing Systems, volume 37, 2024.URL https://proceedings.neurips.c/paper_files/paper/2024/hash/ ea1f5f0878d43f4fb8bf64ef4a2326c-Abstract-Conference.html. Xin Cheng, Xingkai Yu, Chenze Shao, Jiashi Li, Yunfan Xiong, Yi Qian, Jiaqi Zhu, Shirong Ma, Xi- aokang Zhang, Jiasheng Ye, Qinyu Chen, Chengqi Deng, Jiping Yu, Damai Dai, Zhengyan Zhang, Yixuan Wei, Yixuan Tan, Wenkai Yang, Runxin Xu, Yu Wu, Zhean Xu, Xuanyu Wang, Muyang Chen, Rui Tian, Xiao Bi, Zhewen Hao, Shaoyuan Chen, Huanqi Cao, Wentao Zhang, Anyi Xu, Huishuai Zhang, Dongyan Zhao, and Wenfeng Liang. DSpark: Confidence-scheduled specula- tive decoding with semi-autoregressive generation. arXiv preprint arXiv:2607.05147, 2026. doi: 10.48550/arXiv.2607.05147. URL https://arxiv.org/abs/2607.05147. Anh-Dung Dinh, Simon Khan, and Flora D. Salim. Reasoning-aware speculative decoding for efficient vision-language-action models in autonomous driving. arXiv preprint arXiv:2606.31160, 2026. doi: 10.48550/arXiv.2606.31160. URL https://arxiv.org/abs/2606.31160. Haotian Dong, Ye Li, Rongwei Lu, Chen Tang, Shu-Tao Xia, and Zhi Wang. VVS: Accelerating speculative decoding for visual autoregressive generation via partial verification skipping. arXiv preprint arXiv:2511.13587, 2026. doi: 10.48550/arXiv.2511.13587. URL https://arxiv. org/abs/2511.13587. Yonggan Fu, Lexington Whalen, Abhinav Garg, Chengyue Wu, Maksim Khadkevich, Nicolai Os- wald, Enze Xie, Daniel Egert, Sharath Turuvekere Sreenivas, Shizhe Diao, Chenhan Yu, Ye Yu, Weijia Chen, Sajad Norouzi, Jingyu Liu, Shiyi Lan, Ligeng Zhu, Jin Wang, Jindong Jiang, Morteza Mardani, Mehran Maghoumi, Song Han, Ante Juki ́ c, Nima Tajbakhsh, Jan Kautz, and Pavlo Molchanov. Nemotron-Labs-Diffusion: A tri-mode language model unifying autoregres- sive, diffusion, and self-speculation decoding. arXiv preprint arXiv:2607.05722, 2026. doi: 10.48550/arXiv.2607.05722. URL https://arxiv.org/abs/2607.05722. Mukul Gagrani, Raghavv Goel, Wonseok Jeon, Junyoung Park, Mingu Lee, and Christopher Lott. On speculative decoding for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8285–8289, 2024. Mugilan Ganesan, Shane Segal, Ankur Aggarwal, Nish Sinnadurai, Sean Lie, and Vithursan Thangarasa. MASSV: Multimodal adaptation and self-data distillation for speculative decod- ing of vision-language models. arXiv preprint arXiv:2505.10526, 2025. doi: 10.48550/arXiv. 2505.10526. URL https://arxiv.org/abs/2505.10526. Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozi ` ere, David Lopez-Paz, and Gabriel Syn- naeve.Better & faster large language models via multi-token prediction. arXiv preprint arXiv:2404.19737, 2024. doi: 10.48550/arXiv.2404.19737. URL https://arxiv.org/ abs/2404.19737. Yefei He, Feng Chen, Yuanyu He, Shaoxuan He, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. ZipAR: Parallel autoregressive image generation through spatial locality. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, p. 22368–22378, 2025. URL https://proceedings.mlr.press/ v267/he25b.html. Yuezhou Hu and Jintao Zhang. Speculative decoding for autoregressive video generation. arXiv preprint arXiv:2604.17397, 2026. doi: 10.48550/arXiv.2604.17397. URL https://arxiv. org/abs/2604.17397. Yunhai Hu, Zining Liu, Zhenyuan Dong, Tianfan Peng, Bradley McDanel, and Sai Qian Zhang. Mit- igating sequential dependencies: A survey of algorithms and systems for generation-refinement frameworks in autoregressive models. In Findings of the Association for Computational Linguis- tics: EMNLP 2025, p. 13289–13304. Association for Computational Linguistics, 2025a. 21 Yunhai Hu, Tianhua Xia, Zining Liu, Rahul Raman, Xingyu Liu, Bo Bao, Eric Sather, Vithur- san Thangarasa, and Sai Qian Zhang.DREAM: Drafting with refined target features and entropy-adaptive cross-attention fusion for multimodal speculative decoding. arXiv preprint arXiv:2505.19201, 2025b. doi: 10.48550/arXiv.2505.19201. URL https://arxiv.org/ abs/2505.19201. Haiduo Huang, Fuwei Yang, Zhenhua Liu, Xuanwu Yin, Dong Li, Pengju Ren, and Emad Barsoum. SpecVLM: Fast speculative decoding in vision-language models. arXiv preprint arXiv:2509.11815, 2025. doi: 10.48550/arXiv.2509.11815. URL https://arxiv.org/ abs/2509.11815. Mingxiao Huo, Jiayi Zhang, Hewei Wang, Jinfeng Xu, Zheyu Chen, Huilin Tai, and Yijun Chen. Spec-LLaVA: Accelerating vision-language models with dynamic tree-based speculative decod- ing. arXiv preprint arXiv:2509.11961, 2025. doi: 10.48550/arXiv.2509.11961. Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung- Yub Kim, and Eunho Yang. LANTERN: Accelerating visual autoregressive models with relaxed speculative decoding. In The Thirteenth International Conference on Learning Representations, 2025a. URL https://arxiv.org/abs/2410.03355. Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung- Yub Kim, and Eunho Yang. LANTERN++: Enhanced relaxed speculative decoding with static tree drafting for visual auto-regressive models. In ICLR 2025 SCOPE Workshop, 2025b. URL https://arxiv.org/abs/2502.06352. Yicheng Ji, Jun Zhang, Heming Xia, Jinpeng Chen, Lidan Shou, Gang Chen, and Huan Li. SpecVLM: Enhancing speculative decoding of video llms via verifier-guided token pruning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 7216–7230, 2025. Yicheng Ji, Jun Zhang, Jinpeng Chen, Cong Wang, Lidan Shou, Gang Chen, and Huan Li. See the forest for the trees: Loosely speculative decoding via visual-semantic guidance for efficient inference of video llms. In Proceedings of the 64th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), p. 23707–23726, 2026. doi: 10.18653/v1/2026. acl-long.1087. URL https://aclanthology.org/2026.acl-long.1087/. Jialiang Kang, Han Shu, Wenshuo Li, Yingjie Zhai, and Xinghao Chen.ViSpec:Ac- celerating vision-language models with vision-aware speculative decoding. arXiv preprint arXiv:2509.15235, 2025. doi: 10.48550/arXiv.2509.15235. Quan Kong, Yuhao Shen, Yicheng Ji, Huan Li, and Cong Wang. ParallelVLM: Lossless video-llm acceleration with visual alignment aware parallel speculative decoding.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 11392–11402, 2026.URL https://openaccess.thecvf.com/content/CVPR2026/html/ Kong_ParallelVLM_Lossless_Video-LLM_Acceleration_with_Visual_ Alignment_Aware_Parallel_Speculative_CVPR_2026_paper.html. Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, p. 19274–19286. PMLR, 2023. URL https: //proceedings.mlr.press/v202/leviathan23a.html. Bohan Li, Hankun Wang, Situo Zhang, Yiwei Guo, and Kai Yu. Fast and high-quality auto- regressive speech synthesis via speculative decoding. arXiv preprint arXiv:2410.21951, 2025a. doi: 10.48550/arXiv.2410.21951. URL https://arxiv.org/abs/2410.21951. Shenggui Li, Chao Wang, Yikai Zhu, Yubo Wang, Fan Yin, Shuai Shi, Yefei Chen, Xiaomin Dong, Qiaoling Chen, Jin Pan, et al. SpecForge: A flexible and efficient open-source training framework for speculative decoding. arXiv preprint arXiv:2603.18567, 2026. Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. EAGLE: Speculative sampling requires rethinking feature uncertainty. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 28935–28948. PMLR, 2024a. URL https://proceedings.mlr.press/v235/li24bt.html. 22 Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. EAGLE-2: Faster inference of lan- guage models with dynamic draft trees. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p. 7421–7432. Association for Computational Lin- guistics, 2024b. doi: 10.18653/v1/2024.emnlp-main.422. URL https://aclanthology. org/2024.emnlp-main.422/. Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. EAGLE-3: Scaling up inference accel- eration of large language models via training-time test. In Advances in Neural Information Pro- cessing Systems, 2025b. URL https://openreview.net/forum?id=4exx1hUffq. Luxi Lin, Zhihang Lin, Zhanpeng Zeng, and Rongrong Ji. Speculative decoding reimagined for multimodal large language models. arXiv preprint arXiv:2505.14260, 2025a. doi: 10.48550/ arXiv.2505.14260. URL https://arxiv.org/abs/2505.14260. Qingjian Lin, Yuxin Li, Haoyang Zhang, Jun Chen, Yechang Huang, Feng Tian, Xie Li, Xi- angyu Tony Zhang, Daijiao Liu, Yuxin Zhang, Jinglan Gong, Bo Zhao, Fei Tian, Xuerui Yang, Gang Yu, Xiangyu Zhang, and Daxin Jiang. ParaASR: Multi-token prediction for fast and long-context LLM-based speech recognition. arXiv preprint arXiv:2607.29279, 2026. doi: 10.48550/arXiv.2607.29279. URL https://arxiv.org/abs/2607.29279. Zijian Lin, Yang Zhang, Yougen Yuan, Yuming Yan, Jinjiang Liu, Zhiyong Wu, Pengfei Hu, and Qun Yu. Accelerating autoregressive speech synthesis inference with speech speculative decoding. arXiv preprint arXiv:2505.15380, 2025b. doi: 10.48550/arXiv.2505.15380. URL https:// arxiv.org/abs/2505.15380. Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, et al. Convbench: A multi-turn conversation evaluation benchmark with hierarchical capability for large vision-language models. arXiv preprint arXiv:2403.20194, 2024. Qitan Lv, Tianyu Liu, Wen Wu, Xuenan Xu, Bowen Zhou, Feng Wu, and Chao Zhang. HIPPO: Ac- celerating video large language models inference via holistic-aware parallel speculative decoding. arXiv preprint arXiv:2601.08273, 2026. doi: 10.48550/arXiv.2601.08273. Xupeng Miao, Gabriele Oliaro, Zhihao Zhang, Xinhao Cheng, Zeyu Wang, Zhengxin Zhang, Rae Ying Yee Wong, Alan Zhu, Lijie Yang, Xiaoxiang Shi, Chunan Shi, Zhuoming Chen, Daiyaan Arfeen, Reyna Abhyankar, and Zhihao Jia. SpecInfer: Accelerating large language model serving with tree-based speculative inference and verification. In Proceedings of the 29th ACM Interna- tional Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3, p. 932–949. ACM, 2024. doi: 10.1145/3620666.3651335. Tan Dat Nguyen, Ji-Hoon Kim, Jeongsoo Choi, Shukjae Choi, Jinseok Park, Younglo Lee, and Joon Son Chung. Accelerating codec-based speech synthesis with multi-token prediction and speculative decoding. In 2025 IEEE International Conference on Acoustics, Speech and Signal Processing, p. 1–5. IEEE, 2025. Jiahui Niu, Kefan Gu, Yucheng Zhao, Shengwen Liang, Tiancai Wang, Xing Hu, Ying Wang, and Huawei Li. Realtime-VLA FLASH: Speculative inference framework for diffusion-based VLAs. arXiv preprint arXiv:2605.13778, 2026. doi: 10.48550/arXiv.2605.13778. URL https:// arxiv.org/abs/2605.13778. Koji Okabe and Hitoshi Yamamoto. Simultaneous masked and unmasked decoding with specu- lative decoding masking for fast ASR without accuracy loss. In Proceedings of Interspeech 2025, p. 634–637, 2025.doi: 10.21437/Interspeech.2025-382.URL https://w. isca-archive.org/interspeech_2025/okabe25_interspeech.html. Elia Peruzzo, Guillaume Sauti ` ere, and Amirhossein Habibian.Multi-scale local speculative decoding for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. URL https://openaccess.thecvf.com/content/ CVPR2026/html/Peruzzo_Multi-Scale_Local_Speculative_Decoding_ for_Image_Generation_CVPR_2026_paper.html. 23 Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svet- lana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image- to-sentence models. In Proceedings of the IEEE international conference on computer vision, p. 2641–2649, 2015. Qwen Team. Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026a. URL https://qwen.ai/blog?id=qwen3.6-27b. Qwen Team. Qwen3.6-35B-A3B: Agentic coding power, now open to all, April 2026b. URL https://qwen.ai/blog?id=qwen3.6-35b-a3b. Liran Ringel and Yaniv Romano. Accelerating speculative decoding with block diffusion draft trees. arXiv preprint arXiv:2604.12989, 2026. doi: 10.48550/arXiv.2604.12989. URL https: //arxiv.org/abs/2604.12989. George Saon, Samuel Thomas, Takashi Fukuda, Tohru Nagano, Avihu Dekel, and Luis Lastras. Self- speculative decoding for llm-based asr with ctc encoder drafts. arXiv preprint arXiv:2603.11243, 2026. doi: 10.48550/arXiv.2603.11243. Deokjin Seo, Gangin Park, and Kihyun Nam. Chatterbox-flash: Prior-calibrated block diffusion for streaming zero-shot TTS. arXiv preprint arXiv:2605.30748, 2026. doi: 10.48550/arXiv.2605. 30748. URL https://arxiv.org/abs/2605.30748. Hui Shen, Xin Wang, Ping Zhang, Yunta Hsieh, Qi Han, Zhongwei Wan, Ziheng Zhang, Jingxuan Zhang, Jing Xiong, Ziyuan Liu, et al. MMSpec: Benchmarking speculative decoding for vision- language models. arXiv preprint arXiv:2603.14989, 2026. doi: 10.48550/arXiv.2603.14989. URL https://arxiv.org/abs/2603.14989. Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8309–8318. IEEE, 2019. Junhyuk So, Juncheol Shin, Hyunho Kook, and Eunhyeok Park. Grouped speculative decoding for autoregressive image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 15375–15384, 2025.URL https://openaccess.thecvf. com/content/ICCV2025/html/So_Grouped_Speculative_Decoding_for_ Autoregressive_Image_Generation_ICCV_2025_paper.html. Junhyuk So, Hyunho Kook, Chaeyeon Jang, and Eunhyeok Park. Speculative coupled decoding for training-free lossless acceleration of autoregressive visual generation. In Proceedings of the International Conference on Machine Learning, 2026. URL https://arxiv.org/abs/ 2510.24211. Mitchell Stern, Noam Shazeer, and Jakob Uszkoreit. Blockwise parallel decoding for deep autore- gressive models. In Advances in Neural Information Processing Systems, volume 31, 2018. Yao Teng, Zhihuan Jiang, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. SJD++: Improved speculative jacobi decoding for training-free acceleration of discrete autoregressive text-to-image generation. arXiv preprint arXiv:2512.07503, 2025a. doi: 10.48550/arXiv.2512.07503. URL https://arxiv.org/abs/2512.07503. Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, Weiyang Liu, Di- fan Zou, and Xihui Liu. Accelerating auto-regressive text-to-image generation with training-free speculative jacobi decoding. In The Thirteenth International Conference on Learning Represen- tations, 2025b. URL https://arxiv.org/abs/2410.01699. Yao Teng, Fuyun Wang, Xian Liu, Zhekai Chen, Han Shi, Yu Wang, Zhenguo Li, Weiyang Liu, Difan Zou, and Xihui Liu. Speculative jacobi-denoising decoding for accelerating autoregressive text-to-image generation. arXiv preprint arXiv:2510.08994, 2025c. doi: 10.48550/arXiv.2510. 08994. URL https://arxiv.org/abs/2510.08994. Yujia Tong, Tian Zhang, Yunyang Wan, Kaiwei Lin, Jingling Yuan, and Chuang Hu. SAGE: Accel- erating vision-language models via entropy-guided adaptive speculative decoding. arXiv preprint arXiv:2602.00523, 2026. doi: 10.48550/arXiv.2602.00523. 24 Quynh Vo, Cong-Duy Nguyen, Ponhvoan Srey, Luu Anh Tuan, and Thong Nguyen. TIGER: Text- conditioned visual gated routing with acceptance alignment for multimodal speculative decoding. arXiv preprint arXiv:2607.11131, 2026. doi: 10.48550/arXiv.2607.11131. URL https:// arxiv.org/abs/2607.11131. Songsheng Wang, Rucheng Yu, Zhihang Yuan, Chao Yu, Feng Gao, Yu Wang, and Derek F. Wong. Spec-VLA: Speculative decoding for vision-language-action models with relaxed ac- ceptance. arXiv preprint arXiv:2507.22424, 2025a. doi: 10.48550/arXiv.2507.22424. URL https://arxiv.org/abs/2507.22424. Zhuoyu Wang, Junnan Huang, and Xinyu Chen. TAPS: Target-aware prefix tree selection for diffusion-drafted speculative decoding. arXiv preprint arXiv:2606.00487, 2026a. doi: 10.48550/ arXiv.2606.00487. URL https://arxiv.org/abs/2606.00487. Zihua Wang, Ruibo Li, Haozhe Du, Joey Tianyi Zhou, Yu Zhang, and Xu Yang. SpecFLASH: A latent-guided semi-autoregressive speculative decoding framework for efficient multimodal generation. arXiv preprint arXiv:2505.12728, 2025b. doi: 10.48550/arXiv.2505.12728. URL https://arxiv.org/abs/2505.12728. Under review; arXiv v3. Zihua Wang, Zhitao Lin, Ruibo Li, Yu Zhang, Xu Yang, Siya Mi, and Xiu-Shen Wei. Open- loop planning, closed-loop verification: Speculative verification for VLA. arXiv preprint arXiv:2604.02965, 2026b. doi: 10.48550/arXiv.2604.02965. URL https://arxiv.org/ abs/2604.02965. Zili Wang, Robert Zhang, Kun Ding, Qi Yang, Fei Li, and Shiming Xiang. Continuous speculative decoding for autoregressive image generation. arXiv preprint arXiv:2411.11925, 2024a. doi: 10.48550/arXiv.2411.11925. URL https://arxiv.org/abs/2411.11925. Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, et al. Charxiv: Charting gaps in realistic chart understanding in multimodal llms. Advances in Neural Information Processing Systems, 37:113569–113697, 2024b. Linye Wei, Shuzhang Zhong, Songqiang Xu, Runsheng Wang, Ru Huang, and Meng Li. SpecASR: Accelerating llm-based automatic speech recognition via speculative decoding. In Proceedings of the 62nd ACM/IEEE Design Automation Conference, p. 1–7. IEEE, 2025. Chengyue Wu, Shiyi Lan, Yonggan Fu, Sensen Gao, Jin Wang, Jincheng Yu, Jose M. Alvarez, Pavlo Molchanov, Ping Luo, Song Han, Ligeng Zhu, and Enze Xie. Fast-dVLM: Efficient block- diffusion vlm via direct conversion from autoregressive vlm. arXiv preprint arXiv:2604.06832, 2026. doi: 10.48550/arXiv.2604.06832. URL https://arxiv.org/abs/2604.06832. Heming Xia, Zhe Yang, Qingxiu Dong, Peiyi Wang, Yongqi Li, Tao Ge, Tianyu Liu, Wenjie Li, and Zhifang Sui. Unlocking efficiency in large language model inference: A comprehensive survey of speculative decoding. In Findings of the Association for Computational Linguistics: ACL 2024, p. 7655–7671, 2024. Shilong Xiang, Zirui Zhang, Lijun Yu, and Chengzhi Mao. SSD: Spatially speculative decoding accelerates autoregressive image generation. arXiv preprint arXiv:2606.20543, 2026. doi: 10. 48550/arXiv.2606.20543. URL https://arxiv.org/abs/2606.20543. Zhinan Xie, Peisong Wang, Shuang Qiu, and Jian Cheng. HiViS: Hiding visual tokens from the drafter for speculative decoding in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Findings, p. 8952–8961, 2026. URL https://openaccess.thecvf.com/content/CVPR2026F/html/Xie_HiViS_ Hiding_Visual_Tokens_from_the_Drafter_for_Speculative_Decoding_ CVPRF_2026_paper.html. Xiangyuan Xue, Jiajun Lu, Yan Gao, Gongping Huang, Ting Dang, and Hong Jia. Edge-cloud collaborative speech emotion captioning via token-level speculative decoding in audio-language models. arXiv preprint arXiv:2603.11397, 2026.doi: 10.48550/arXiv.2603.11397.URL https://arxiv.org/abs/2603.11397. 25 Moran Yanuka, Paul Dixon, Eyal Finkelshtein, Daniel Rotman, and Raja Giryes. Principled coarse- grained acceptance for speculative decoding in speech. arXiv preprint arXiv:2511.13732, 2026. doi: 10.48550/arXiv.2511.13732. URL https://arxiv.org/abs/2511.13732. Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multi- modal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 9556–9567, 2024. Kewei Zhang, Jin Wang, Sensen Gao, Chengyue Wu, Yulong Cao, Songyang Han, Boris Ivanovic, Langechuan Liu, Marco Pavone, Song Han, Daquan Zhou, and Enze Xie. Fast-ddrive: Efficient block-diffusion VLM for autonomous driving. arXiv preprint arXiv:2605.23163, 2026a. doi: 10.48550/arXiv.2605.23163. URL https://arxiv.org/abs/2605.23163. Libo Zhang, Zhaoning Zhang, Wangyang Hong, Peng Qiao, and Dongsheng Li. Sparrow: Text- anchored window attention with visual-semantic glimpsing for speculative decoding in video llms. arXiv preprint arXiv:2602.15318, 2026b. doi: 10.48550/arXiv.2602.15318. Xuan Zhang, Cunxiao Du, Sicheng Yu, Jiawei Wu, Fengzhuo Zhang, Wei Gao, and Qian Liu. Sparse-to-dense: A free lunch for lossless acceleration of video understanding in llms. In Pro- ceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p. 734–742, 2025. Zihao Zheng, Zhihao Mao, Maoliang Li, Jiayu Chen, Xinhao Sun, Zhaobo Zhang, Donggang Cao, Hong Mei, and Xiang Chen. KERV: Kinematic-rectified speculative decoding for embodied vla models. In Proceedings of the 63rd ACM/IEEE Design Automation Conference, 2026a. Zihao Zheng, Zhihao Mao, Sicheng Tian, Maoliang Li, Jiayu Chen, Xinhao Sun, Zhaobo Zhang, Xuanzhe Liu, Donggang Cao, Hong Mei, et al. HeiSD: Hybrid speculative decoding for embod- ied vision-language-action models with kinematic awareness. arXiv preprint arXiv:2603.17573, 2026b. doi: 10.48550/arXiv.2603.17573. 26