arXiv paper2026-08-31· machine learning — efficient inference for LLMs/VLMs· method paper

Vision Is Not Overhead: One-Pass Block Drafting for Lossless Speculative Decoding in Vision-Language Models

Jungseob Lee, Seongtae Hong, Dongyub Jude Lee, Chanjun Park, Jaehyung Seo, Sugyeong Eo, Heuiseok Lim

cs.AI · cs.CL · cs.CV

Quick verdict

Claims first lossless one-pass block drafter for frozen VLMs; strong engineering and stats, but capacity confound and 'entropy law' framing warrant scrutiny.

Summary

GLANCE is a speculative-decoding drafter for vision-language models that targets two constraints the authors argue have stalled VLM drafting: per-token sequential depth and degraded vision access. A 1.05-billion-parameter block-diffusion head (DFlash family) reads the frozen Qwen3-VL-8B target's fused vision-language hidden states from five layers and fills a 16-token block in one forward pass; the highest-scoring prefixes form a wide tree verified in a single ancestor-masked target pass. The authors prove greedy-path equivalence (Theorem 1), gate runs on bitwise reproduction, and report 60/60 exact fp32 matches. Under SGLang 0.5.6 with an equal 32-draft-token budget, GLANCE decodes up to 2.93x faster than autoregression (ChartQA) and 5.4-7.6% faster than the production EAGLE3-VL head on image-grounded tasks, but 16-19% slower on captioning and TextVQA. With matched training it accepts 2.7x longer blocks (4.0x on ChartQA). Section 6 fits an affine-logit entropy law (per-task curve R2 0.45-0.95, round-level only 7-14% against a disclosed noise ceiling), and a mismatched-image probe shows vision acts through the entropy channel. Stated limitations: batch-one decode-only, an fp32-stated guarantee, ordering that inverts on Llama-3.2-11B, small samples in several harnesses, and a 2.6x larger head without capacity-matched control.

Review complete

Assessment at a glance

machine learning — efficient inference for LLMs/VLMs

risk flags identified

● high 0

● medium 2

● low 2

Reproducibility readinessmedium

Code & data availableyes

Review confidencehigh

AI text signal0/20 passages flagged

Deep research viableCode-reproducibleNeeds GPU

4 strengths · 4 weaknesses · 5 author questions

AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.

Levels reflect verification complexity — not a quality score. Evolve never rates your paper.

Key claims under review

GLANCE is the first one-pass block drafter that is lossless on an unmodified (frozen) vision-language target, combining a block-diffusion head over fused vision-language states with single-pass wide-tree verification (Theorem 1, Table 7 fp32 audit 60/60).
Under one engine (SGLang 0.5.6) and an equal 32-draft-token round budget, GLANCE decodes up to 2.93x faster than autoregression (ChartQA) and 5.4-7.6% faster than the production EAGLE3-VL head on image-grounded tasks, while being 16-19% slower on captioning and TextVQA (Table 1).
With matched training on the same corpus, schedule and framework, GLANCE accepts 2.7x longer blocks pooled across tasks and 4.0x on ChartQA versus an EAGLE-3 head (Table 2).
Accepted block length is governed by the target's next-token entropy through an affine-logit law whose slope steepens with grounding (pooled b1=0.63, per-task 0.46-1.71), and a mismatched-image probe shows the image acts through the entropy channel (Section 6).
A verbatim-copy two-state regime lifts the low-entropy floor above the single-hazard cap (DocVQA floor 5.23 vs cap 3.41, Theorem 2), explaining why one-pass drafting harvests grounded runs whole.

Strengths

Real architectural exit from a measured bottleneck
Section 3.2 and Table 11 quantify the cost asymmetry (one-pass draft = 0.19 target forwards; EAGLE3-VL pays 1.82 ms per sequential draft pass, ~38% of a round from 2 to 8 passes). The DFlash head inherits fused states at layers {1,9,17,25,33}, so vision costs the drafter nothing — a clean resolution of the cycle diagnosed in Section 1.
Losslessness is measured, not asserted
Theorem 1's proof is simple, but the audit discipline is the value: fp32 gating, a tie-aware bf16 gate, and Table 7 showing drafter-free bf16 decoding itself matches only 25/60, correctly relocating bf16 disagreement to arithmetic rather than the drafter. ViSpec/Medusa show 0/63 exact matches under any tree setting.
Attribution ablations isolate the mechanism
Causal and falsification tests rare in this literature

Weaknesses

Capacity confound in the headline comparison
Deployment envelope is narrow and underweighted in headlines
'Entropy law' presentation outruns fit quality
Cross-table comparability and thin samples

Full Review Report

$1.99 · one-time

Unlock the complete preliminary analysis Evolve produced for this paper.

Report preview

Scope & Claims

The paper introduces GLANCE, a speculative-decoding drafter for vision-language models, and makes four load-bearing claims. (1) It is the first one-pass block drafter that is lossless on an unmodified (frozen) VLM target: a DFlash-family block-diffusion head (1.05B parameters, 5 layers, block size 16) reads the target's fused vision-language hidden states at layers {1, 9, 17, 25, 33} and fills a 16-token block in one forward pass, after which a prefix-closed candidate tree (budget 31-63) is verified in a single ancestor-masked target pass. (2) Under one engine (SGLang 0.5.6) and an equal 32-draft-token round budget (Table 1), GLANCE decodes up to 2.93x faster than autoregression on ChartQA and 5.4-7.6% faster than the production EAGLE3-VL head on the three image-grounded tasks, while losing 16.2-19.1% on captioning and TextVQA. (3) With matched training (same corpus, batch, epochs, framework; Table 2), it accepts 2.7x longer blocks pooled across tasks (4.0x on ChartQA). (4) Accepted length is governed by the target's next-token entropy through an affine-logit law (Eq. 2), and a mismatched-image probe shows the image acts through the entropy channel.

Methodology Assessment

The core design is sound and the cost model is quantified rather than asserted: one draft pass costs 0.19 target forwards, the packed verify 1.07 (Table 11), and EAGLE3-VL pays 1.82 ms per sequential draft pass, so depth is exactly what one-pass drafting removes. The attribution work is better than typical for this literature. The matched-training block controls corpus, schedule and framework; the tree-vs-chain ablation runs the identical head as a width-1 chain (gives up at least 1.45x on every task; forward-level tree gain 1.36x, Table 9); zeroing the fused conditi

The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.

All 4 strengths & 4 weaknesses, analyzed in detail
4 risk flags with severity, section & reviewer questions
Methods & statistical evidence assessment
5 questions a peer reviewer would ask
Full six-dimension evaluation report & recommendations

One-time purchase · this paper only

One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.

Frequently asked questions

What is GLANCE in one sentence?

GLANCE is a speculative-decoding drafter for vision-language models: a block-diffusion head reads the frozen target model's fused vision-language hidden states and drafts a whole 16-token block in one forward pass, and a wide candidate tree is then verified in a single target pass with provably unchanged (greedy-identical) output.

How much faster is GLANCE than normal decoding?

On Qwen3-VL-8B in SGLang 0.5.6 with equal draft budgets, up to 2.93x versus plain autoregression on ChartQA, and roughly 5-8% faster than the production EAGLE3-VL head on image-grounded QA tasks; on open-ended captioning it is 16-19% slower than that head, so the method targets grounded workloads.

Does GLANCE change the model's outputs?

The paper proves and audits greedy-path equivalence: all 60 audited prompts matched the frozen target's fp32 greedy decoding exactly. In production bf16 arithmetic, a tie-aware gate defers to the target's own token at near-ties, and the target's bf16 decoding itself drifts from fp32 even with no drafter.

Is code or training data available?

The abstract links a public GitHub repository (github.com/js-lee-AI/GLANCE), and Appendix D gives a complete training card: 16K teacher-forced rows of the target's own greedy generations (8K COCO + 8K TextVQA), one epoch on a single GPU.

About this review

This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.

Figures 1-9 appear only as captions and in-text descriptions in the extraction; plotted curve shapes (e.g., Fig. 3a decile curves) could not be visually inspected.
Some table rows are interleave-scrambled by PDF extraction (e.g., Table 1's header line); all numbers cited in this review were cross-checked against the body text and appendix tables.

Go deeper on this paper

Wrote a paper? Stress-test it before reviewers do.

Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.

Scan your paper