PaperlayerView exact revision on arXiv

Source-linked AI summary

Loud and Clear: Making Visual Evidence Count in Multimodal Document RAG

Eunchan Jo, Seounghyun Jang, Sangwoo Mo

arXiv:2610.21318v1cs.CVcs.AI

TL;DR

Multimodal document RAG has focused largely on retrieval, leaving comparatively underexplored how retrieved pages are given to the generator. The paper combines full-page scaling and selected-region OCR in LoC; it outperforms eight prior methods across three benchmarks, with gains over ColQwen3 using the same retriever and generator.

  • Problem

    How retrieved pages are given to the generator remains comparatively underexplored, despite the generator needing to use evidence conveyed through page text, layout, tables, charts, figures, and typography.

    Source: §1 · p. 1

  • Method

    LoC combines full-page scaling with OCR from query-relevant regions scored using the retriever’s token embeddings, without ground-truth annotations or additional training.

    Source: §1 · p. 2, §3.4 · p. 7

  • Results

    LoC improves over ColQwen3 by 8.88 points on SciEGQA, 3.09 on ViDoRe-v3, and 4.22 on MMLongBench-Doc, using the same retriever and generator.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings highlight how retrieved pages are presented to the generator alongside which pages are retrieved.

    Source: §5 · p. 9

    Evaluation focuses on English documents, Qwen-family generators, and a single reference-based judge; larger visual-token budgets also trade accuracy against inference latency.

    Source: §5 · p. 9

Abstract

from arXiv · show

Multimodal document retrieval-augmented generation (RAG) has focused largely on retrieval, yet its answers depend on how well a vision-language model reads the retrieved pages. We find that evidence in retrieved pages remains underused even when the correct pages are retrieved. On SciEGQA oracle pages, five prior methods give the generator 0.7K to 2.7K visual tokens per query on average under a common image processor, yet raising the page-only budget from 1K to 32K improves accuracy by 15.22 points. At a fixed 16K budget, adding ground-truth crops improves accuracy by 3.75 points over page-only input. We study what the generator needs from retrieved pages through two requirements: Loud, making full use of visual token capacity, and Clear, refining the page so that its evidence becomes prominent. For Loud, increasing either the visual token budget or visual fidelity, the image resolution those tokens preserve, improves accuracy with the other held fixed, and full-page scaling outperforms repetition and tiling at matched budgets. For Clear, suppressing surrounding content improves accuracy without enlarging selected regions, and adding OCR of these regions to page images outperforms either input alone. We implement both as Loud and Clear (LoC), which combines higher-resolution pages with OCR from query-relevant regions without ground-truth annotations or additional training. LoC outperforms eight prior methods on SciEGQA, ViDoRe-v3, and MMLongBench-Doc, improving over the strongest, ColQwen3, by 8.88, 3.09, and 4.22 points with the same retriever and generator. It also outperforms its page-only variant under the same visual token budget. These results locate a bottleneck beyond retrieval: how the retrieved pages are given to the generator.

1 INTRODUCTION

Multimodal document RAG has emphasized retrieval, while how retrieved pages are presented to the generator remains comparatively underexplored. The paper diagnoses this input bottleneck and introduces Loud and Clear designs, combined in LoC, to help the generator use evidence already present on retrieved pages.

1 INTRODUCTION

  • With pages and generator fixed, increasing the page-only budget from 1K to 32K improves accuracy by 15.22 percentage points, revealing underused evidence after retrieval succeeds.Five prior methods use 0.7K–2.7K visual tokens per query on average under a common image processor.Source: §1 · p. 1
  • At a fixed 16K budget, pairing pages with ground-truth crops adds 3.75 points over pages alone, motivating content refinement.The paper presents this as an oracle diagnosis of the value of making evidence prominent.Source: §1 · p. 1
  • Loud requires sufficient visual token budget and high visual fidelity, while Clear refines page content so its evidence becomes prominent.The controlled comparisons hold retrieved pages and generator fixed; ground-truth region localization is used for oracle diagnosis, not for LoC.Source: §1 · p. 2
  • Full-page scaling improves accuracy as both visual token budget and fidelity increase, while page images with selected-region OCR outperform either input alone at the same budget.Whiting out content outside selected regions also improves accuracy; extracting or regrouping regions gives no consistent advantage.Source: §1 · p. 2
  • LoC outperforms eight prior methods, improving over ColQwen3 by 8.88 points on SciEGQA, 3.09 on ViDoRe-v3, and 4.22 on MMLongBench-Doc.It also beats its page-only variant under the same visual token budget, without ground-truth annotations or additional training.Source: §1 · p. 2

2 RELATED WORK

Related work has advanced visual retrieval, token allocation, and content refinement, but leaves open how well generators use evidence in the same retrieved pages. This paper positions controlled comparisons of generator inputs as addressing that gap.

2.1 VISUAL RETRIEVAL FOR MULTIMODAL DOCUMENT RAG

  • Visual retrieval preserves page layout through image-native representations, including DSE’s single dense vector and ColPali’s patch-level embeddings with late-interaction scoring [Ma et al., 2024a; Faysse et al., 2025; Khattab & Zaharia, 2020].Evaluation has also expanded from page relevance toward evidence selection, document structure, answer generation, and visual grounding.Source: §2.1 · p. 3

2.2 VISUAL TOKEN CAPACITY IN VISION-LANGUAGE MODELS

  • Visual-token methods select, prune, or allocate tokens to reduce computation or spend a prescribed budget, leaving open how much accuracy low budgets sacrifice in document RAG [Chen et al., 2024a; Zhang et al., 2025; Arif et al., 2025].The paper varies budget and fidelity separately with pages and generator fixed, and compares matched-budget alternatives.Source: §2.2 · p. 3

2.3 CONTENT REFINEMENT IN MULTIMODAL DOCUMENT RAG

  • Generator-side systems concatenate or select pages, curate textual and visual evidence, or replace pages with regions, but their separate pipelines do not isolate which refinement helps on the same pages at a fixed budget.The paper therefore compares content refinements while holding retrieved pages, generator, and visual token budget fixed.Source: §2.3 · pp. 3–4

3 LOUD AND CLEAR: IMPROVING EVIDENCE USE IN RETRIEVED PAGES

The section tests how generators use retrieved pages while holding pages and generator fixed, then develops Loud and Clear as complementary input-design requirements. It combines full-page scaling with OCR from query-relevant regions in LoC.

3 LOUD AND CLEAR: IMPROVING EVIDENCE USE IN RETRIEVED PAGES

  • The diagnosis varies visual token budget and content refinement while fixing retrieved pages, generator, prompt, decoding, and evaluation.Its first three sections use oracle pages containing ground-truth evidence; Section 3.4 combines the Loud and Clear designs.Source: §3.1 · p. 4, §3 · p. 4

3.1 EVIDENCE IN RETRIEVED PAGES REMAINS UNDERUSED

  • 15.22 points: page-only accuracy rises from 65.74% at 1K tokens to 80.96% at 32K on unchanged SciEGQA oracle pages and generator.Five prior methods provide 0.72K–2.68K visual tokens per query on average; at 16K, ground-truth-matched crops raise accuracy to 83.73% versus 79.98% for page-only input.Source: §3.1 · p. 4

3.2 LOUD: MAKING FULL USE OF VISUAL TOKEN CAPACITY

  • Full-page scaling outperforms repetition and spatial tiling at matched 2K–64K budgets on both benchmarks, while multi-scale-view advantages vary by benchmark.Both higher visual fidelity at fixed final budget and a larger final budget at fixed fidelity improve accuracy; full-page scaling increases both.Source: §3.2 · pp. 5–6, Figure 3 · p. 5

3.3 CLEAR: MAKING EVIDENCE PROMINENT

  • Whiteout improves accuracy over Page by 2.03 points on SciEGQA and 0.81 on ViDoRe-v3 without enlarging selected regions.Adding selected-region OCR to page images beats either input alone, improving over Page by 2.89 points on SciEGQA and 0.88 on ViDoRe-v3; the latter gain’s paired bootstrap interval includes zero.Source: §3.3 · pp. 6–7, Figure 4 · p. 7

3.4 THE LOC FRAMEWORK

  • LoC combines scaled retrieved pages with OCR from query-relevant regions selected using the retriever’s token embeddings, without ground-truth annotations or additional training.All retrieved pages remain as visual context; region scores sum query-token maximum similarities to visual tokens within each region.Source: §3.4 · p. 7
  • 8.88, 3.09, and 4.22 percentage points: LoC improves over ColQwen3 on SciEGQA, ViDoRe-v3, and MMLongBench-Doc, respectively.LoC reaches the highest accuracy on all three benchmarks; ColQwen3 and LoC share the retriever and generator.Source: §3.4 · p. 7, Table 2 · p. 8

4 APPLYING LOC TO END-TO-END RAG SYSTEMS

LoC improves end-to-end accuracy over prior systems on all three benchmarks, and controlled comparisons show gains from adding selected-region OCR to scaled page images. The gains carry across evaluated generators, although OCR allocation can hurt at the smallest visual-token budget.

Setup

  • The evaluation covers 1,623 SciEGQA queries, 1,489 ViDoRe-v3 queries, and 1,091 MMLongBench-Doc queries, using Qwen3.6-35B-A3B as the default generator.LoC uses four retrieved pages, five regions per page, and default scale s = 2.25.Source: §4 · p. 8

End-to-end RAG performance

  • 8.88, 3.09, and 4.22 points: LoC improves over ColQwen3 on SciEGQA, ViDoRe-v3, and MMLongBench-Doc, respectively.LoC achieves the highest accuracy on all three benchmarks; paired query-level bootstrap intervals for the gains exclude zero.Source: §4 · p. 8

Effects of visual token budget and content refinement

  • 2.59, 0.54, and 2.29 percentage points: under the same visual token budget, LoC improves over its page-only variant on SciEGQA, ViDoRe-v3, and MMLongBench-Doc, respectively.Both variants use ColQwen3's retrieved pages, page order, and generator.Source: §4 · p. 8
  • LoC w/o Clear rises from 69.19% at s = 1.0 to 75.11% at s = 1.75, then stays near 75%, while LoC adds 2.59–3.76 percentage points at scales s = 1.25–3.0.Figure 5 plots SciEGQA accuracy across reference page scales with retrieved pages, page order, and generator fixed; OCR allocation at the smallest budget may reduce page visual tokens.Source: §4 · p. 9, Figure 5 · p. 9
  • At the smallest budget, LoC scores 65.87%, below 69.19% for LoC w/o Clear, as OCR allocation reduces mean page visual tokens from 3.80K to 1.48K.The authors suggest this may lower page-image fidelity and support keeping visual fidelity high when some budget goes to OCR.Source: §4 · p. 9

Generalization across generators

  • LoC improves over ColQwen3 on all three benchmarks with each of two additional Qwen2.5-VL generators.For each generator, the methods use the same retrieved pages and page order, with LoC at scale s = 2.25.Source: §4 · p. 9, Table 3 · p. 9

5 CONCLUSION AND DISCUSSION

The paper argues that correctly retrieved pages can still be underused and presents LoC as a way to improve how those pages are given to the generator. Its gains come with latency, OCR, and evaluation-scope limitations.

5 CONCLUSION AND DISCUSSION

  • LoC combines full-page scaling with OCR from retriever-scored regions, without ground-truth annotations or additional training.The paper reports that it outperforms eight prior methods on three benchmarks and improves over ColQwen3 across three generators.Source: §5 · p. 9
  • LoC’s larger visual token budgets trade accuracy against inference latency, and Clear depends on document parsing and OCR quality.Allocating tokens to OCR can reduce accuracy under tight budgets.Source: §5 · p. 9
  • Evaluation is limited to English documents, Qwen-family generators, and a single reference-based judge.The authors identify broader evaluation across languages, model families, and judges as future work.Source: §5 · p. 9

A UNDERUSED VISUAL TOKEN CAPACITY IN PRIOR METHODS

The audit measures prior methods’ generator-side visual-token use after adapting their image settings to a shared processor. It characterizes those adapted settings, not the native inputs or performance of the original systems.

A UNDERUSED VISUAL TOKEN CAPACITY IN PRIOR METHODS

  • Five prior generator-side methods use a mean of 0.72K–2.68K visual tokens per query on the 1,623 SciEGQA queries under a common image processor.The audit sums views across retrieved pages; ColPali is a retriever-side setting and is excluded from this range.Source: Table 4 · p. 16, A UNDERUSED VISUAL TOKEN CAPACITY IN… · p. 16
  • The audit fixes queries, retrieved page identities and order, original images, and the final processor while adapting each method’s image settings.It applies settings to all retrieved pages per query without the original retrieval top-k or a matched query-level budget from the controlled sweep.Source: A UNDERUSED VISUAL TOKEN CAPACITY IN… · p. 16
  • The common processor counts (H/32)(W/32) visual tokens for an image view of height H and width W, summing across views and pages per query.The method counts expanded image tokens and checks them against image grids.Source: A UNDERUSED VISUAL TOKEN CAPACITY IN… · p. 16
  • Adapted settings mapped to a common token grid do not preserve native resolution, token count, and information content simultaneously, especially when learned compression is replaced by an area constraint.The audit establishes input levels on these pages; separate controlled results assess the utility of greater visual-token capacity.Source: A UNDERUSED VISUAL TOKEN CAPACITY IN… · p. 17

B EXTENDED ANALYSIS OF LOUD

The extended Loud analysis compares page-image representations at fixed oracle pages and target visual-token budgets. Its conditions include full-page scaling, repeated views, multi-scale views, and spatial tiling.

B.1 CONDITIONS

  • The analysis fixes oracle pages and varies their visual representation, assigning each query a total visual-token target from 1K to 64K using page images only.The conditions add no crops, OCR text, or other inputs.Source: §B.1 · p. 18
  • Full-page resolution scaling uses one full-page view per page, resized directly from the original image to match the query-level target budget.It does not scale from a lower-budget input.Source: §B.1 · p. 18
  • Repeated-page and repeated-blank-page conditions compare multiple page views or added blank images with full-page scaling at matched query-level budgets.The target budget covers all image views for a query.Source: §B.1 · p. 18
  • Multi-scale views provide three independently constructed full-page views per oracle page at relative scales 1, 2, and 4, with budget shares in a 1:4:16 ratio.The corresponding allocations are 1/21, 4/21, and 16/21 per page; views are ordered coarse to fine.Source: §B.1 · p. 18
  • Spatial tiling supplies four uniform 2 × 2 page tiles without retaining a full-page view, sharing the query budget across all tiles and pages.Each tile’s resolution decreases as the number of pages grows.Source: §B.1 · p. 18

B.2 VISUAL TOKEN ACCOUNTING AND RESOLUTION ALLOCATION

The appendix specifies how visual-token budgets are allocated and tests how fidelity, budget, and input construction affect accuracy. Its comparisons also document how selected-region OCR and visual refinements are constructed and evaluated.

B.2 VISUAL TOKEN ACCOUNTING AND RESOLUTION ALLOCATION

  • Mean realized visual-token counts stay within 1.32% of target across Loud constructions; the allocator scales page dimensions jointly and maps them to a 32-pixel grid.At query level, four constructions stay within 5% of target; multi-scale views have larger deviations on ViDoRe-v3 at the smallest budgets.Source: Table 5 · p. 19, §B.2 · p. 19

B.3 SEPARATING INTERMEDIATE RESOLUTION FROM FINAL BUDGET

  • Visual fidelity and final visual-token budget each improve accuracy when the other is held fixed.The two-stage resizing comparisons vary intermediate resolution at fixed final budget, or final budget at fixed intermediate resolution; conditions with intermediate resolution above the final budget are not evaluated.Source: Table 6 · p. 20, §B.3 · p. 20

B.4 COMPLETE BUDGET SWEEP

  • Multi-scale views lead on SciEGQA at 2K–64K, while full-page resolution scaling leads on ViDoRe-v3 at 4K–64K.Full-page scaling reaches its observed maximum at 32K on SciEGQA and 8K on ViDoRe-v3; the authors use a single scaled full-page view as a common choice across both benchmarks.Source: §B.4 · p. 20, Table 7 · p. 20

C.1 CONSTRUCTION

  • Selected MinerU regions are padded after matching; whiteout preserves page dimensions, while crops are presented as individual images.OCR strings are concatenated in detection order by page and crop, without reconstructing table structure or correcting text against reference answers.Source: §C.1 · p. 21

C.2 ADDITIONAL COMPARISONS

  • +2.09 and 0.61 percentage points in accuracy over Page when crop OCR is added while preserving page images and their visual tokens exactly.With identical crop OCR under a shared OCR-plus-visual-token budget, no visual refinement leads on both benchmarks.Source: §C.2 · p. 21, Table 8 · p. 21, Table 9 · p. 22

C.3 REGION SELECTION AND ANSWER CONTENT

The section examines region suppression and OCR/crop variants, then uses examples to show how page context can preserve heading associations and captions. It also introduces the implementation-details section.

C.3 REGION SELECTION AND ANSWER CONTENT

  • Whiteout gains are larger for queries with smaller selected-region fractions, but the unmatched groups do not establish an area effect or an area-based input rule.The groups differ in question type, document structure, and difficulty.Source: §C.3 · p. 22, Table 10 · p. 22
  • Page answers 3 economic concepts and 3 data sources, while Whiteout answers 4 and 3, matching the selected passage and reference.The abstract names three concepts; the selected introduction passage lists four, including the Poverty Gap Index.Source: §C.3 · p. 22
  • Crop OCR and Page + Crop OCR return Jacobabad’s N28°17′ latitude for the highest qualifying MLAT, while Page answers with Bhownuggur’s latitude.The selected OCR includes both rows’ sites, coordinates, MLAT values, and references; the combined input uses 15,012 VT versus 15,984 for Page.Source: §C.3 · pp. 22–23

C.4 EXAMPLES OF PAGE CONTEXT

  • Crop and Canvas retain the requested items but place “Lot-level” under Verification, whereas Page + Crop answers correctly.The page separates Product Tracing and Verification into columns; the example does not isolate layout from image-size or grouping changes.Source: Figure 6 · p. 23, §C.4 · p. 23
  • Page + Crop OCR answers Figure 4 and Figure 5 correctly, while Whiteout with identical OCR misses Figure 4 after its caption is removed.The shared OCR does not contain the caption, which lies outside the selected regions.Source: Figure 6 · p. 23, §C.4 · p. 23

D.1 CONTROLLED CONDITIONS AND EVALUATION

The controlled comparisons fix query cohorts and retrieved views, vary image-token budgets, and hold generation settings constant. Reported accuracy uses the strict evaluation protocol.

D.1 CONTROLLED CONDITIONS AND EVALUATION

  • SciEGQA comparisons use the same 1,623 queries with oracle pages fixed; ViDoRe-v3 uses five English datasets, with 1,488 queries in the controlled comparisons.The end-to-end ViDoRe-v3 evaluation includes all 1,489 queries; query 239 is excluded from controlled comparisons because its ground-truth boxes are missing.Source: §D.1 · p. 24
  • The visual-token budget targets total image-processor tokens across input images, with OCR conditions counting OCR text and visual tokens together.Budgets range from 1K to 64K and are targets, allowing small deviations after image-size quantization.Source: §D.1 · p. 24
  • Generation uses Qwen3.6-35B-A3B with fixed QA instructions, BF16, temperature zero, disabled thinking, and up to 1,024 generated tokens.OCR conditions prepend extracted OCR text to the QA prompt.Source: §D.1 · p. 24
  • Reported text and Figure 2 accuracy is strict accuracy under the shared evaluation protocol.The supplied passage identifies the protocol but does not specify its scoring criteria.Source: §D.1 · p. 24

D.2 ANSWER EVALUATION PROTOCOL

The paper evaluates answers with a shared reference-based judge and compares end-to-end configurations using paired query outcomes. The paired intervals support LoC gains on all three benchmarks, while Clear’s ViDoRe-v3 gain remains uncertain.

D.2 ANSWER EVALUATION PROTOCOL

  • The judge weighs question suitability and reference accuracy equally; only answers fully satisfying both criteria are correct, and partial or incorrect answers receive zero strict-score credit.Equivalent wording and exact numerical or unit representations are accepted, but rounding is accepted only when the question or reference permits it.Source: §D.2 · p. 25

E.1 PAIRED COMPARISONS AND UNCERTAINTY

  • LoC-over-ColQwen3 paired query-bootstrap intervals exclude zero on all three benchmarks; Clear’s intervals exclude zero on SciEGQA and MMLongBench-Doc, but remain uncertain on ViDoRe-v3.Intervals describe query-sampling uncertainty conditional on the saved answers and judgments.Source: Table 11 · p. 28, §E.1 · p. 28

D.3 END-TO-END RAG CONFIGURATIONS

  • End-to-end conditions share query sets, questions, and references within each benchmark, and the evaluation includes all 1,489 ViDoRe-v3 queries.Recomputed accuracy reproduces all nine reported values; no query-intersection filtering is needed.Source: §E.1 · p. 28

E.1 PAIRED COMPARISONS AND UNCERTAINTY

  • Paired comparisons assess ColQwen3 against LoC w/o Clear, LoC w/o Clear against LoC, and ColQwen3 against LoC; intervals are not corrected for multiple comparisons.The bootstrap resamples paired query outcomes and is conditional on saved answers and judgments, not independent model runs.Source: §E.1 · p. 28

E.2 VIDORE-V3 DOMAIN RESULTS

Clear’s ViDoRe-v3 contribution varies across domains, despite LoC’s higher observed accuracy in each. Document-cluster analyses support positive Clear and Total intervals on SciEGQA and MMLongBench-Doc, with a boundary on ViDoRe-v3 analysis.

E.2 VIDORE-V3 DOMAIN RESULTS

  • Clear’s net ViDoRe-v3 gain is eight correct answers: computer science +12, pharmaceuticals +9, industrial reports +5, offset by HR −12 and finance −6.The pooled row weights each query equally.Source: §E.2 · p. 28, Table 12 · p. 29
  • LoC has higher observed accuracy than ColQwen3 in all five ViDoRe-v3 domains; computer science, finance, and pharmaceuticals account for 44 of 46 additional correct answers.The domain intervals are exploratory and are not corrected for multiple comparisons.Source: §E.2 · p. 28, Table 12 · p. 29

E.3 DOCUMENT-CLUSTER SENSITIVITY

  • Document-cluster 95% intervals for Clear and Total remain above zero on SciEGQA and MMLongBench-Doc; Loud’s MMLongBench-Doc lower bound is approximately zero.SciEGQA intervals are [0.98, 4.15] for Clear and [6.55, 11.31] for Total; MMLongBench-Doc intervals are [0.18, 4.47] and [2.33, 6.19], respectively.Source: §E.3 · pp. 28–29
  • ViDoRe-v3 document-cluster resampling is omitted because 130 queries reference multiple documents, requiring extra assumptions for unique assignments.Source: §E.3 · p. 29

F GENERATION-STAGE LATENCY ACROSS VISUAL TOKEN BUDGETS

The archived runs show a visual-token-budget/accuracy–latency tradeoff, but do not establish a uniform latency advantage for LoC or a latency-optimal scale. Reported latency covers preprocessing and concurrent engine residence, not end-to-end RAG latency.

F GENERATION-STAGE LATENCY ACROSS VISUAL TOKEN BUDGETS

  • Latency sweeps compare LoC and LoC w/o Clear at nine scales on the same 1,623 SciEGQA queries; slow queries and answers reaching the output limit remain included.Source: F GENERATION-STAGE LATENCY ACROSS… · p. 30
  • Measured latency includes preprocessing and concurrent engine residence with queueing, but excludes retrieval, region selection, OCR, model loading, and some preparation and I/O; it is neither end-to-end RAG latency nor isolated GPU time.Source: F GENERATION-STAGE LATENCY ACROSS… · p. 30
  • Runs share GPU and generation settings but workloads differ; host identity and instantaneous resource load were unrecorded, making these descriptive comparisons rather than controlled speedup measurements.Source: F GENERATION-STAGE LATENCY ACROSS… · p. 30
  • Table 14's s = 1.0 page baseline comes from a mixed-benchmark session with different input preparation and no explicit answer warm-up, so it establishes no controlled latency ratio against LoC.Source: Table 14 · p. 31
  • At s = 2.75 to 3.0, LoC accuracy falls from 78.93% to 78.50% as median latency rises from 482.30 to 658.35 seconds; across each dedicated sweep, median latency rises while accuracy gains diminish.For LoC w/o Clear, s = 1.75 to 3.0 changes accuracy from 75.11% to 75.23% while median latency rises from 121.65 to 745.66 seconds.Source: F GENERATION-STAGE LATENCY ACROSS… · p. 31
  • Across seven shared scales, LoC has lower median latency but higher P95 at s = 2.0, 2.5, 2.75, and 3.0; at s = 3.0, median is 658.35 versus 745.66 seconds, while P95 is 1,000.64 versus 792.53 seconds.Source: F GENERATION-STAGE LATENCY ACROSS… · p. 31

Related papers

Loading 2610.21318v1…