Overview

Speculative decoding aims to reduce the time and computation needed to generate text by letting a drafter propose future tokens for a target model to check. Approaches include small auxiliary drafters, early-exit or multi-token prediction within the target model, and tree-based or block-parallel drafting; retrieval and task-specific signals can also supply candidates. Exact verification can preserve the target model’s output distribution, while relaxed acceptance rules trade some fidelity for speed. A central challenge is making gains in accepted tokens translate into real latency or throughput improvements across different workloads, hardware, and serving loads.

Latest papers

343 papers58 papers in the last four weeks, down 6% on the four weeks before. 0.4% of all new papers.

Nov 13, 2026

MMFlash: Faster Speculative Decoding for Vision-Language Models

Speculative decoding accelerates generation by letting a lightweight drafter propose tokens for parallel verification by a target model. Diffusion drafters such as DFlash propose an entire token block in a single forward pass, but mainly focus on text-only tasks, leaving multimodal drafting underexplored. To study this setting, we first introduce a metric of token-level visual dependence based on the KL divergence between next-token distributions with and without visual input. We find that acceptance rates decrease from 46% in the lowest visual-dependence quartile to 22% in the highest, with this disparity persisting after multimodal training. Motivated by this observation, we propose MMFlash, a simple and effective framework that adaptively upweights visually dependent tokens that the drafter predicts poorly. For long video inputs, MMFlash further employs sparse attention to reduce redundant attention computation and improve drafting quality. Across various multimodal benchmarks, MMFlash achieves lossless decoding speedups of up to 3.58×3.58\times, outperforming EAGLE-3 and DFlash by up to 195.4% and 17.1%, respectively.
Nov 13, 2026

Callo: Serving Cache-Miss Requests with Speculative Decoding on Invalidated KV under Prefill-Decode Disaggregation

Agent sessions may involve frequent context edits, which effectively invalidate cached KV state and force requests to wait for re-prefill before decoding. We present Callo, a disaggregated serving system that leverages invalidated KV for speculative drafts to reduce this latency, while verifying drafts against the exact KV after prefill before committing tokens. Callo makes effective use of decode resources by scheduling speculations during periods of underutilization and preempting speculations when drafts are likely to not be accepted. We evaluated Callo on a two-GPU testbed using SWE-agent sessions across three model families, observing all-request speedups of 1.05-1.19x over vLLM at moderate load. We demonstrate that these gains are also realized at scale through cluster-level simulations driven by real-world traces.
Nov 12, 2026

Towards Green Text Generation: Memory-Based Speculative Decoding

Speculative decoding accelerates text generation with large language models by letting a cheap draft model propose tokens that a large verifier checks in a single forward pass. We introduce Olifant, a draft model built on a memory-based classifier. It predicts the next token from a four-token window in under 0.3 ms on CPU, requiring no GPU, no backpropagation, and no accelerator memory. We measure throughput and energy under two verifier regimes on consumer hardware. Without a KV cache, drafting is nearly free: a domain-matched Olifant model reaches 38.1% acceptance on EU legal text, delivering 2.43x throughput and 2.51x lower energy per token than the neural verifier model alone. With a KV-cached verifier, itself a ~10x energy saver, drafting costs matter: a TinyLlama draft model halves throughput and triples energy despite high acceptance, whereas Olifant still yields 1.25x throughput and the lowest energy in the study (93 J per 100 tokens, 14x below the uncached baseline). We further show that KV caching decouples time from energy: verification of longer drafts raises power draw, shifting the energy-optimal draft length below the speed-optimal one. Speculative decoding is not dead in the age of fast verifiers, provided drafting is nearly free.
Nov 10, 2026

REAL Loss: Optimizing Real Expected Acceptance Length for Speculative Decoding

Speculative decoding accelerates LLM inference by having a draft model propose multiple tokens for the target model to verify in one pass. The speedup is governed by the expected acceptance length (EAL), yet standard training objectives such as cross-entropy, KL divergence, and total variation focus only on distribution matching or single-step overlap. Recent studies instead propose block-level training objectives, aiming to optimize EAL directly. However, these objectives remain surrogates rather than exact formulations of EAL. For example, some average later acceptance over target-distributed continuations rather than the verification branches selected by accepted tokens, introducing a covariance gap that can reverse the gradient direction. We analyze this gap and derive REAL, a real EAL training loss that uses path weights for unbiased estimation and Rao-Blackwellization for variance reduction. In a controlled loss-level comparison with a Qwen3.6-35B-A3B target and a Domino drafter, REAL outperforms WTV, VSD, DBLast, and the static LK baseline on probabilistic serving acceptance length across 9 task subsets and 5 evaluation seeds. These gains require only a change to the training loss, with no modification to the drafter architecture or inference procedure, demonstrating the practicality of real EAL training. A reference implementation of the REAL loss is available at https://github.com/King-ty/speculators/tree/real-loss.
Nov 10, 2026

ForeSpark: Joint Draft and Verification Budget Scheduling for Parallel Speculative Decoding

Parallel drafters accelerate LLM decoding by proposing a whole block of draft tokens in a single forward pass, making them attractive for LLM serving. Recent systems optimize goodput (i.e., accepted tokens per second) because rejected drafts steal compute from other requests. However, these systems control only how many tokens are verified, so every request still drafts a full block. At high serving load, we find that drafting consumes up to 40% of step time, of which a majority is discarded because it is never verified. Yet, draft-length control from autoregressive drafters does not carry over, because a parallel drafter commits to its length before its single pass. Moreover, draft and verification lengths are coupled across the batch, so they must be selected jointly for every request while keeping the decision computationally cheap at every step. In this paper, we propose work-conserving speculation (WCS), a design that jointly allocates draft and verification budgets per request in LLM serving, and present ForeSpark as its prototype. ForeSpark realizes WCS with (i) a context-aware pre-draft estimator that predicts acceptance for every candidate length from the target model's context features, (ii) a lightweight joint draft-verify scheduler that selects both lengths through a global sort, and (iii) a post-draft budget distribution that spends verification on the most promising tokens. Implemented on vLLM, ForeSpark improves throughput by up to 1.77x over the state-of-the-art DSpark and by 2.49x over autoregressive decoding at a matched per-request generation rate, without changing the output distribution. It also tolerates higher load, sustaining up to 1.51x the request rate of DSpark under latency service level objectives (SLOs).
Nov 9, 2026

Phrase-Level Speculative Decoding on a Frozen Target

Speculative decoders raise throughput by banking more target tokens per forward pass, and Medusa, EAGLE-3, DFlash and DSpark all do it the same way: more candidates that survive verification. All of them draft in the target's own vocabulary, so one node of the candidate tree is one target token, and the tokenizer makes that unit look like a constant. It is not. BPE's pretokenizer splits on whitespace before merges are learned, so no multi-word string can ever become a single token: in Qwen3-4B, "United States of America" is four tokens and no adjacent pair of them exists as one entry, so a phrase is out of reach of any method drafting in the target's vocabulary. We change the unit. A small MLP head reads the frozen target's hidden state and predicts a one- to three-word phrase as one entry in a separate draft vocabulary; greedy longest-match expansion maps it back into the target's own tokens, and N candidates are verified in one forward pass behind a prefix-tree mask, so the change is orthogonal to tree depth and width. The target and its tokenizer stay frozen and a candidate is accepted only while it matches the target's own argmax, so the served output is identical to greedy decoding up to floating-point tie-breaking. On one A5000 at batch 1, against HuggingFace generate(), a 21.94M head (0.545% of Qwen3-4B) reaches 2.391x on GSM8K, 2.444x on MATH-500 and 1.544x on MT-Bench. Because the head's output layer is one row of 257 parameters per entry, the draft vocabulary size is the parameter budget: an 8,000-entry head is 3.434M parameters, 0.0854% of the target and 64x smaller than EAGLE-3's drafter, and keeps 92.3% of the larger head's GSM8K tokens per pass. With EAGLE-3 run on same prompts and baseline, our 21.94M head is ahead of tree-16 on both maths sets and behind tree-60 on every set.
Nov 9, 2026

DForkTree: Accelerating Speculative Decoding with Fork-Aware Block Diffusion Draft Trees

Speculative decoding speeds up language models by having a lightweight drafter propose future tokens that the target model verifies in parallel. Block diffusion drafters such as DFlash predict an entire block in one forward pass, and DDTree builds draft trees directly from their per-position distributions. Each distribution implicitly assumes that the drafter's own greedy continuation is accepted, whereas a draft tree improves over a chain only on branches that leave this greedy path. Once a path forks, the drafter's top-1 accuracy drops from 89% to 18% and its top candidates often miss the true token. Re-running the drafter from the fork token largely restores its accuracy, showing that the errors stem from conditioning on an outdated prefix rather than from limited model capacity. In this paper, we introduce DForkTree, a training-free method that re-anchors each branch on its own tokens. It adjusts the drafter's scores with pointwise mutual information from target-generated n-grams, weighted more heavily after a fork, and a layer-wise GPU search recovers the exact optimal tree under these scores. On Qwen3.5-4B and 9B, DForkTree outperforms all baselines on every benchmark, by up to 41.5% over DFlash and 15.1% over DDTree, with average speedups of up to 4.75x over autoregressive decoding.
Nov 9, 2026

Multi-Future Drafting: Mixture Block Drafters for Tree-Verified Speculative Decoding

Block drafters such as DFlash draft a block of future tokens in one forward pass, and draft trees such as DDTree verify the prefixes ranked highest by the product of the drafter's per-position distributions. Under tree verification, the expected number of accepted tokens is the probability mass the tree covers. We show that a product misallocates it: when the continuation can follow several futures, it splices them into chimeras, prefixes stitched together from different futures that the target never produces, and gives them as much budget as the futures themselves. Barring ties, no product-ranked tree can hold two futures that disagree in mm positions without 2m−1−12^{m-1}-1 chimeras, however well the drafter is trained, and in practice chimeras make up 27-33% of a 64-node tree. We propose Multi-Future Drafting (MFD), which restores the coupling with a latent variable instead of the chain rule: KK learned mode embeddings at the drafter's input draft KK coherent blocks in the same pass, and a mixture DDTree lets the verifier pick the future the target takes, losslessly and with K+1K+1 extra vectors. On four Qwen targets from 4B to 27B parameters, at batch size 1, MFD reaches up to 7.8×\times speedup, 5-8% faster than the best single-mode tree trained identically and, on Qwen3-4B and Qwen3-8B, 9-16% faster than the released DARTree trained on the same prompts.
Nov 8, 2026

D3Tree: Learning to Prune Block-Diffusion Draft Trees for Efficient Verification

Block-diffusion drafters propose a whole block of tokens in one forward pass, and draft trees built from their position-wise marginals increase the expected accepted length of speculative decoding. Every tree node occupies a query row in target verification, so under concurrent serving low-yield branches compete with other requests for verification capacity. Existing tree methods select verified nodes by draft probability under a fixed or cost-selected per-request budget, and thus always retain a prefix of the drafter's ranking. We present D3Tree, which learns which nodes of a block-diffusion draft tree are worth verifying. A tree-size head with 2H+2 parameters predicts each candidate's conditional survival from the draft state, the candidate embedding, and its draft log-probability. Thresholding the product of these scores along parent chains yields an ancestor-closed subtree whose size adapts to each request and decoding round. The head is trained on a frozen drafter with conditional labels on the full candidate tree; an optional joint mode also steers the drafter toward the retained reference path. At serving time, D3Tree verifies whole trees under light load and, under heavy load, allocates one batch-wide node budget by predicted survival, sized to the target's matrix-multiplication tiles. Retained trees share one ancestor-masked target pass, while construction, pruning, and scheduling run on the GPU and overlap with the host. Across six vision-language verifiers and 17 text and multimodal benchmarks, D3Tree is fastest in 70 of 72 load settings, with 3.2-4.0x speedup over autoregressive decoding for one request and 1.3-1.8x at 64 concurrent requests, where fixed-budget trees fall below autoregressive speed. Under a time-varying arrival trace, it has the lowest mean end-to-end latency in every arrival phase.
Nov 7, 2026

D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding

Speculative decoding efficiency depends on the number of consecutively accepted draft tokens, yet token-level training objectives do not directly optimize this quantity. We introduce D-REAL (Drafter Reinforcement of Expected Acceptance Length), a post-training method that directly optimizes expected acceptance length for a fixed decoding context. D-REAL weights position-wise draft-target distributional overlap by prefix survival probabilities along a draft-sampled prefix, faithfully capturing the sequential dependence of verification on earlier proposals. By accounting for the parameter dependence of the proposal distribution, we derive an unbiased gradient estimator for this single-round objective under stochastic verification. We evaluate D-REAL on DSpark and DFlash2 with Qwen3-8B as the target model across ten benchmarks. Starting from baselines whose mean acceptance length has plateaued, one epoch of D-REAL post-training consistently improves the benchmark-averaged mean acceptance length by up to 0.23 tokens across both architectures under greedy (T = 0.0) and stochastic (T = 1.0) sampling, translating into up to a 9.4% end-to-end serving throughput improvement.
Nov 5, 2026

ChainFlow: Few-Step Latent Flow Matching for Speculative Decoding

Speculative decoding has become a standard way to accelerate LLM inference: a lightweight drafter proposes tokens that the target verifies in one parallel pass. Existing drafters face trade-offs that limits their speedups: autoregressive drafters generate tokens one at a time, so drafting cost grows with draft length, while diffusion drafters generate tokens in parallel but predict each position independently, ignoring inter-token dependencies. We introduce ChainFlow, a new speculative decoding paradigm and the first flow-matching drafter conditioned on the target's hidden states. ChainFlow treats drafting as a short transport over hidden states, needing one or two flow steps instead of the many denoising steps of generation from noise. A variational autoencoder compresses these states so drafting cost barely grows with target size, lightweight feedback heads restore inter-token dependencies, and tree drafting raises acceptance. On Qwen3.5 models, at matched drafter size and training data, ChainFlow outperforms EAGLE-3 and DFlash in vLLM by 39% and 11%, respectively. These gains further widen as drafters are adapted to new languages sequentially. Our ablations cover drafter design and concurrent serving, and we open-source our code and drafter models.
Nov 5, 2026

HyFlash: A Prefix-Memory Prior for On-Device Speculative Decoding

Parallel block drafting accelerates speculative decoding by predicting multiple future positions in a single backbone pass, but on-device deployment constrains both drafter capacity and context storage. Under tight memory budgets, drafters must model long-range dependencies and fine-grained draft-position interactions without the growing KV-cache overhead of full-context attention. We propose HyFlash, a compact parallel drafter that introduces a prefix-memory prior to separate long-range context modeling from fine-grained draft-position interactions with limited capacity. Local attention captures recent context and draft-position interactions, while recurrent memory compresses historical target features into fixed-size states. This inductive bias enables effective local and long-range modeling with bounded draft-side memory. Experiments show that HyFlash achieves higher acceptance than full-context attention with substantially less KV storage and with larger gains for shallow drafters and long contexts.
Nov 4, 2026

Decode Before You Hear: Speculative Listening for Chunk-Interleaved Full-Duplex Speech Language Models

Chunk-interleaved full-duplex speech language models listen and speak in fixed time chunks, and each chunk must be processed before the next one arrives. Their workload is asymmetric across chunks: listen chunks leave compute idle while awaiting audio, and speak and tool-call chunks require serial generation, which delays responses and tool calls. Using the idle compute for the next chunk requires decoding it before its audio arrives, under an assumed input. Taking silence as the assumed input and comparing the output with the output under real audio on four models, we observe two properties. First, the next audio chunk does not change the output in 95.3% of chunks. Second, 88% of changes begin at the first output token. Building on these properties, we propose Speculative Listening, a training-free, lossless framework for input-time self-speculative decoding. During idle time, the model drafts the next chunk under silence and, for changes at the first token, a second branch from the runner-up first token. It keeps all drafts in a cross-chunk hypothesis pool and verifies them when the audio arrives. On our eight-task benchmark TickTock-Bench, Speculative Listening speeds up speak chunks by 2.6-2.8x and tool-call chunks by up to 23.4x. On average, it reduces the decoding delay before a response starts by 73%, to 74-241 ms, close to the 200 ms gap between human turns, and the maximum lag behind real time by 94.4%.
Nov 4, 2026

D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00×\times and 5.71×\times decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03×\times and 3.67×\times for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48×\times. Code is publicly available at https://github.com/fudan-generative-vision/D3D.
Nov 4, 2026

RECAP: Context-Aware Prediction Recycling with Request-Local and Persistent Memories for Multimodal Speculative Decoding

Recycling a target model's own predictions offers a training-free source of proposals for speculative decoding. In vision-language models, however, existing recycling indexes predictions by the last token alone, conflating predictions computed under different images and prefixes; we find that such reuse fails most often at image-sensitive decoding states. Longer keys reduce this aliasing but match fewer states. We introduce \method, which stores target predictions under one-, two-, and three-token keys: contextual memories live only for the current request and are thus implicitly conditioned on its image, while token-level memory persists across requests to retain coverage. At each lookup, entropy-confidence weighting fuses the retrieved predictions according to their key specificity and concentration. Best-first search composes the fused transitions into a candidate tree, supplemented by suffix-matched continuations, and one target pass verifies the tree under greedy decoding. Without training or an auxiliary draft model, \method\ achieves mean speedups over autoregressive decoding of 1.66\speed1.66\speed and 1.56\speed1.56\speed across seven single-turn benchmarks on Qwen2.5-VL-7B and LLaVA-1.6-7B, exceeding the strongest training-free baseline, Token Recycling (1.51\speed1.51\speed and 1.37\speed1.37\speed). Our implementation is available in https://github.com/Senmo996/RECAP-official-code.
Nov 3, 2026

Give the Drafter a Second Chance to Guess: Parallel Reconsideration for Speculative Decoding

Speculative decoding accelerates large language model inference with a small drafter, whose per-guess acceptance rate determines the achievable speedup. Feature-level drafters share a design choice: every token, easy or hard, is drafted under a single forward pass. Many of the tokens the target rejects are ones the drafter is itself unsure about yet the target answers decisively. To address this, we introduce Parallel Reconsideration: guided by its own uncertainty, the drafter reconsiders its hesitant guesses with extra passes that run in parallel with the one-pass draft, adding negligible latency. A co-designed training recipe and serving-time drafting algorithm append these reconsidered guesses to the original draft tree as refined candidates. Across models and benchmarks, on average, Parallel Reconsideration improves over the baselines by 5.3% to 10.2% on acceptance length and by 3.5% to 9.2% on speedup.
Nov 3, 2026

Hybrid Spark: Adaptive Speculative Decoding That Lets the Content Choose the Drafter

Speculative decoding (SD) speeds up large language model (LLM) serving, but its gain depends on the workload's shape, such as batch size and context length, and on its content. A model-based drafter proposes a short block on any content, while a model-free drafter such as n-gram lookup proposes long copies at almost no cost when the output repeats its input, as in code editing. Adaptive verification spends a batch-wide verification budget on the draft tokens most likely to be accepted, but it supports only one model-based drafter. Model-free drafters provide no per-token confidence, so they cannot take part in budget allocation, and deployments must either fix one drafter or split their GPUs between drafters. We present \sys, a framework that runs a model-based and a model-free drafter on every step and gives each request the better of their chains, within adaptive verification and entirely on the GPU, which keeps asynchronous scheduling intact. The key idea is a price per verified draft token: the goodput that slot would earn elsewhere in the batch. Once the model-free drafter's tokens carry calibrated acceptance estimates, fit once on general text, this price puts both drafters' chains on one scale. Instantiated with \dspark and an n-gram drafter in vLLM, \sys achieves 1.10--1.62\x the output throughput of \dspark on code editing at batch size4 and 1.04--1.27\x at batch size128, while staying at 0.97--1.01\x of it on chat at temperature~0. A single deployment can therefore serve both workloads.
Nov 3, 2026

Adversarial Speculative Reasoning Improves Long Horizon Agents

Speculative decoding has been treated purely as a systems optimization that speeds up large language model decoding without changing the model behavior. Under this view, the growing effort to train better draft models has been focused solely on acceptance rate. We argue that this view does not hold true for long-horizon agent reasoning. Measuring frontier open-weight models on complex agentic tasks with different speculative decoding approaches, we found that numerical noise grows within verification windows, and the draft acceptance results of the thinking process can change the output distribution of the model after 30 steps, thus noticeably influencing the reasoning accuracy of long-horizon agents in random directions. In this work, we improve the reasoning capability of target models by controlling such noise. While we cannot change the overall token distribution verified by the target model, we can encode critical reasoning with highest precision by training the draft model to be rejected. We identify these points in agent traces and train the draft model to favor adversarial preference. At inference the trained draft and the target part ways precisely at these decision points, verification restarts, so that critical thinking tokens are sampled from the target model and encoded under its least noisy conditions. This approach makes critical reasoning better indexed and more durable over a long horizon. On complex agentic coding benchmarks, our approach raises end-to-end success rate with no change to the target model or its output distribution. Our results recast the speculative decoding mechanism from a pure accelerator into an active allocator for context precision that improves the accuracy of long-horizon agents by training 0.06% to 0.28% of hosted parameters.
Nov 2, 2026

HybridSpec: Throughput-Aware Speculative Decoding for Hybrid GDN-MoE Models

In speculative decoding, parallel drafters such as DFlash make drafting cheap, so target verification dominates each round. A central choice is the verification length, the number of tokens that the target processes in each round. Verifying more draft tokens can accept more of them, but work after the first rejection is wasted. For hybrid target models that combine Gated DeltaNet (GDN) and mixture-of-experts (MoE) layers, verification time grows much faster with this length than for dense models. In a single-request profile of Qwen3.6-35B-A3B with DFlash, it grew by 71% from length 4 to 16, almost entirely in GDN and MoE layers, compared with 4%-6% for a dense Qwen3-4B target. Each GDN layer must update a recurrent state at every token and recover it at the end of the accepted prefix. The MoE cost depends on the experts that the tokens activate, which are known only once verification runs. We present HybridSpec, which selects the verification length on the GPU from draft-side acceptance estimates and measured target latency. Verification then runs this length plus one extra token up to the first MoE router, before any expert runs. There, a lightweight routing correction uses the observed expert assignments to decide whether to keep that token. The rest of verification runs only on the selected tokens. Block-parallel GDN kernels avoid storing a state after every token and reconstruct the accepted state of every GDN layer in one launch. In isolated benchmarks at verification lengths 4 to 16 with every draft token accepted, these kernels, including accepted-state reconstruction, are 1.54×1.54\times-4.04×4.04\times faster than the default kernels. With three GDN-MoE target models in single-request greedy decoding, HybridSpec is 1.75×1.75\times-1.88×1.88\times faster than autoregressive decoding and 1.30×1.30\times-1.33×1.33\times faster than DFlash.
Nov 2, 2026

Testing Processor Contracts in Speculative Decoding Engines

Speculative sampling can accelerate decoding without changing the target distribution. Its standard guarantee assumes that logits processors, which enforce constraints or modify token scores, receive the appropriate token prefix and state at each verification position. We formalize this processor-prefix contract and present BADC (branch-aware differential conformance testing). BADC combines enumerated references on tiny random models, checks of processor inputs and state, and simultaneous confidence bounds on output-event probabilities. Preregistered, source-informed experiments on selected Transformers and llama-cpp-python releases confirm two causes: processors read unverified drafts or advance state without rollback. For one pilot-selected prompt on a trained checkpoint, fresh samples give a 60.26-percentage-point lower bound on the difference in the probability of a predefined first-token event between ordinary and prompt-lookup decoding, with nominal simultaneous 95% coverage under within-arm i.i.d. sampling. Controlled comparisons establish no cause-recovery advantage of directed selection over random selection, while permutation testing detects each tested family at least as often as BADC. These findings connect processor-interface failures to observable decoding differences without establishing diagnostic superiority or workload prevalence.
Nov 2, 2026

SpecPilot: Autonomous Algorithm Discovery and Systems Optimization for Speculative Decoding

Speculative decoding accelerates memory-bound LLM decoding with a lightweight drafter, and parallel drafters extend this speedup by drafting a whole block in one forward pass. However, it remains unclear whether their released systems (stock) attain the full speedup that parallel drafting permits. We identify three challenges. First, with one request, each step accepts few tokens but reads every target weight. Second, with batching, per-request verification work roughly halves the speedup over autoregressive (AR) decoding. Third, speculative decoding is deployment-specific, and no universal optimization fits every serving regime, drafter, target model and GPU. To address these challenges, we present SpecPilot, an autonomous agentic loop that discovers algorithms and systems optimizations for each deployment. In each round, a cost model locates the largest gaps between the current system and its ceiling. Parallel agents explore directions and integrate the successful ones. Token forcing checks every output token against stock to detect reward hacking. A structured memory carries validated mechanisms to new rounds and deployments. SpecPilot discovered a general lossless 12-bit weight format, tree verification for hybrid recurrent targets, two-level verification and three batching mechanisms. We evaluate it across four target models, three drafters, two GPU platforms and five benchmarks. Without training or quality loss, it raises the speedup over AR from stock's 4.46×\times to 10.87×\times with one request, and from 1.94×\times to 3.93×\times with batching.
Nov 1, 2026

FactorSpec: Serving Exact Relational Constraints as Speculative Drafts

Speculative decoding speeds up LLM serving by drafting tokens that the target model verifies in one pass, and its gain grows with the drafts the target accepts. Structured outputs should be an ideal case, because much of a record follows from relations among its fields: an itinerary's days follow from its cities' stays and flights. Serving engines, however, draft such text by proposing it from the prompt or a learned head, and their grammar masks enforce only syntax, so the model must satisfy the relations, which it violates: on NATURAL PLAN's trip problems, Qwen3-8B writes a valid itinerary for 10%, and retries that name the violated relation raise this to 14%. We present FactorSpec, which serves exact relational constraints as speculative drafts. FactorSpec compiles a JSON Schema and its relations into a factor graph with an exact feasibility check, drafts every token the constraint determines as certain, verifies each request at a width chosen for it, and computes the masks in worker processes outside the scheduler while the GPU runs. In SGLang, its drafts yield up to 10.5 tokens per verify step, up to 1.9x those of the fastest n-gram drafter, and it decodes up to 1.43x faster than speculators tuned per load, at a grammar mask's cost; it also leads in vLLM. Because every record is also valid, it serves up to 13.9x the valid records per GPU-second of syntax-only baselines.
Nov 1, 2026

ReFrame: Silence-Aware Compute Reclamation for Full-Duplex Speech Models

Full-duplex speech models execute dense model inference for every audio frame throughout a live interaction. This uniform framewise schedule creates sustained GPU load across dialogue states whose computation has strong, reusable temporal structure. ReFrame formulates full-duplex inference as a silence-aware computation problem. Mutual Silence Speculation (MSS) operates when both participants are silent: it evaluates multi-frame Temporal continuations, verifies text and hidden-state stability, and applies batched Mimi audio-token correction. Listening Frame Folding (LFF) operates while the user speaks and the agent listens: it processes already-arrived user frames through parallel Temporal execution and frame-batched audio-token generation while accepting only the verified prefix. Frame-Budget Reinvestment (FBR) assigns the released frame intervals to resumable external retrieval and language-model workers, then injects timely references through Moshi's conditioning interface. We instantiate ReFrame on five full-duplex models spanning Moshi-family, phase-structured, and flattened autoregressive generation. Across the complete 1,225-conversation FDBench, ReFrame reduces measured GPU execution time by 18.80-31.01% and retains 96.92-100% non-PAD text agreement on five different models. On HaluEval tests, FBR increases the number of correct generated spoken answers from 86 to 205 under the same auxiliary-service budget. ReFrame turns recurring dialogue structure into lower inference cost and higher end-to-end service capability.
Nov 1, 2026

Seeing What Matters, Verifying What Pays: Load-Adaptive Speculative Decoding for Vision-Language Models

Speculative decoding accelerates vision-language model (VLM) inference without altering the output distribution, yet two factors limit its gains. Existing drafters either lack the image evidence required for visually grounded tokens or pay to condition on every visual token, although a short draft block typically depends on only a few image regions. Moreover, the value of verifying additional candidates varies with GPU load, so no fixed verification structure is efficient across serving conditions. We present Uni-LaSpec, a block-parallel speculative decoding framework that addresses both factors. To improve draft quality, Uni-LaSpec reuses the verifier's own attention: the query retained from the preceding verification ranks the cached, int8-quantized visual keys, and a cross-layer vote selects a few visual tokens that the drafter fuses through gated cross-attention. Acceptance-aware training further weights each draft position by its contribution to the accepted prefix. To allocate verification, a marginal-cost criterion selects the structure per request and per round: draft trees exploit idle verifier capacity under light load, whereas a dynamic head truncates linear drafts under heavy load. Across six VLMs from three model families and 17 text and multimodal benchmarks, selecting eight visual tokens retains at least 98.6% of the acceptance obtained from the full image at sub-millisecond selection cost. With greedy decoding at concurrency one, Uni-LaSpec achieves up to 5.11x text and 3.24x multimodal throughput over autoregressive decoding. Across six load levels and three GPU types, it outperforms autoregressive decoding in every setting and matches or exceeds the best fixed verification strategy in 123 of 144 settings; under bursty arrivals, it reduces mean end-to-end latency by 47.6% relative to linear verification.
Oct 31, 2026

TVP: TV is Almost All You Need for Speculative Decoding

Speculative decoding accelerates LLM inference with a drafter whose proposed tokens are verified by the target model in parallel. Under lossless speculative sampling, a draft token is accepted with probability one minus the total variation (TV) distance between the target and draft distributions, yet drafters are often trained with cross-entropy (CE) or KL. We first show that under matched and well-tuned training, TV matches or surpasses several published draft-training objectives when fine-tuning, and a fixed TV and KL mixture does the same when training from scratch. We identify two remaining problems, TV's vanishing gradients on undercovered tokens and the prefix products that starve later positions in acceptance-length objectives. To address both, we derive TVP, a loss family based on expected acceptance length, whose power parameter interpolates between coverage and acceptance oriented objectives. Annealing the power from zero to one gives the highest or tied-highest acceptance length across three drafter architectures and three target models, including longer drafts and block verification. TVP improves acceptance length over CE by 9.8% when fine-tuning, and over the public DFlash2 training recipes by 6.2 to 12.9%.
Oct 30, 2026

TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents

LLM agents execute tasks through repeated model calls over evolving histories of reasoning, actions, and tool feedback. Aggregate speculative-decoding metrics merge these distinct decisions, providing limited guidance on where draft supervision is needed. We present TraceSpec, a framework connecting trajectory evaluation with adaptive draft training. It combines dynamic execution with fixed-prefix replay and introduces TraceSpec-Traj, a corpus of 10,000 training states and 2,500 test states spanning browsing and software engineering. Analysis of five speculative baselines reveals that aggregate acceptance can rise with context length while declining within matched late browsing actions. Workload and phase further distinguish these trends, motivating supervision guided by trajectory conditions. TraceSpec trains on intermediate assistant turns and allocates a fixed prediction budget using state-conditioned estimates of prefix-aware acceptance gaps. Across three target sizes and two temperatures, it improves replay throughput by 12.5% to 22.6% over its DSpark base. Qwen3-4B ablations show that adaptive allocation increases accepted length by 3.3% to 10.0% over fixed full-turn supervision, with additional comparisons substantiating the roles of state conditioning and consecutive-token agreement. The resulting evaluation resources, trajectory findings, and training strategy connect the analysis of agent execution to practical draft improvement.
Oct 29, 2026

Only Accepted Tokens Count: An Energy Accounting of Speculative Decoding Under Latency Targets

Large language model serving is limited by facility power. Energy per token determines how many tokens a datacenter can serve, while a latency target bounds how quickly each user receives them. During decode, every weight is read from memory for every generated token. Batching amortizes this read across users but makes no individual user faster. Speculative decoding does: a drafter proposes several future tokens, and the target model verifies them in a single forward pass that reuses the same weight read. Because every drafted token must be verified whether or not it is accepted, the energy speculation saves depends on how many drafted tokens are accepted. This saving is hard to pin down. It varies across target models from different attention families and across drafting schemes, and it depends on the text being served. Every drafter we evaluate accepts fewer tokens on conversation than on code, and every trained drafter accepts fewer on Romansh, a low-resource language, than on English. We measure speculative decoding under production serving conditions (SGLang, CUDA graphs enabled) on H100 and H200 GPUs, at batch sizes up to 64 across 5 corpora. On dense models, verification, not drafting, dominates the energy cost. The saving shrinks as the batch grows, because batching has already shared the weight read that speculation would divide among the accepted tokens. At large batch sizes, block drafters push verification past the balance point between arithmetic and memory traffic into the compute-bound regime, where the saving becomes a loss, and a drafter with low acceptance raises energy per token. Meeting a tighter latency target with speculation then costs additional energy. Larger batches lower energy per token without speculation. Speculation adds to that saving only when acceptance is high, and on the model we tested it leaves little headroom for latency-preserving DVFS.
Oct 28, 2026

Why Speculative Decoding Works Better Than Predicted on Sparse MoE Models

Speculative decoding (SD) accelerates autoregressive inference by leveraging the memory bound nature of decoding, verifying multiple draft tokens in a single forward pass. For mixture-of-experts (MoE) models, prior work predicts a concave speedup-batch size relationship, with gains peaking at intermediate batch sizes due to expert saturation. We evaluate this across 10 sparse MoE models (30B--1T parameters, 8--512 experts) and find recent fine-grained models break this trend, showing higher speedup at low batch size. We find verification cost is governed not only by activated-expert bytes, but also by temporal overlap across draft tokens and GPU execution dynamics. At low batch size, MoE layers in fine-grained models account for a much smaller share of runtime than of weights; on H100, kernel profiling attributes this mainly to higher bandwidth utilization of MoE kernels relative to attention, with L2 absorbing repeated reads of co-routed experts. Across draft tokens, routing overlap further reduces the unique experts activated during verification relative to independence assumptions. We model this effect using a maximum-entropy Iterative Proportional Fitting (IPF) framework that predicts unique expert counts from routing-overlap statistics with less than 4% error, and substantially narrows the gap to real-model curves and captures the qualitative low-batch trend from synthetic routing.
Oct 28, 2026

TableTune: Measured, Self-Correcting Draft-Length Tables for Dynamic Speculative Decoding

Speculative decoding speeds up large language model (LLM) serving only while the GPU has spare compute, so the best draft length, the number of tokens a small drafter guesses per step, shrinks as the batch grows. vLLM lets operators set the draft length for each batch size with a hand-written table, but nothing measures or corrects that table. We present TableTune. Its tuner times the decoding step for every batch size and draft length on vLLM's own serving path and computes, in minutes, the table that maximizes throughput. Its controller, a vLLM plugin, tracks how many guesses are accepted in live traffic and rebuilds the table while serving. We measure TableTune only against other ways to speculate. On an H100, the tuned table delivers at least 94% of the throughput of the best fixed draft length at every load in four of five drafter settings (88% in the fifth) without the operator having to know which length that is. Against reimplementations of TurboSpec- and Nightjar-style control on the same engine, its worst-case loss is never higher than theirs, and with the EAGLE-3 drafter it is 3% against 9% and 12%. On a mixture-of-experts model, vLLM's documented table loses up to 18% and the tuned table at most 3%. Starting from mistuned tables, the controller raises throughput by up to 58% (up to 45% on vLLM's default runner), and after a workload switch with the DFlash drafter it recovers 96% of a stale table's loss. In llm-d's optimized-baseline deployment as shipped, two Qwen3-32B replicas on four GPUs, the tuned table is up to 32% faster than a fixed draft length of 3 and up to 25% faster than vLLM's documented table at high load, and the controller raises the documented table's throughput by up to 29%.
Oct 26, 2026

LOCUS: Locate Once, Check Until Settled. Draft-and-Verify Visual Grounding with a Box Head and Coordinate-Token Verification

Vision-language models (VLMs) ground a referring expression in one of two ways. They decode the box as coordinate tokens, which is precise but takes one sequential step per token, or they regress it from a hidden state with a small head, which takes one step but loses accuracy at high IoU. The two fail differently: a head usually finds the right object but places its edges loosely, while coordinate tokens place edges precisely. We therefore use the regression head as a speculative drafter for the model's own coordinate tokens. Locus writes the head's box as coordinate tokens and checks them with the same backbone by Jacobi iteration over a shared image prefix. The fixed point is the model's greedy coordinate-token answer, reached in about 5.60 sequential passes on average instead of 18.46 decoding steps; head and verifier are one set of weights, trained jointly. On the 8 RefCOCO, RefCOCO+ and RefCOCOg splits, verification raises Locus-2B's [email protected] from 54.87 for its own head to 59.13, and the one-pass head is ahead at [email protected] of every published model with at most 3B parameters except Florence-2-L. Against Jag, the strongest open regression-head model, at its own size and training budget, Locus ties at [email protected] and, with typical acceptance, gains +4.52 points at [email protected], with paired significance tests. Exact answers come 1.5x faster than vLLM greedy-decoding the same coordinate tokens with the same weights; a pure regression head remains faster still.