acceptodds
Desk rejected at ICLR 2027

VPCache: Answer-Preserving Cache-View Selection for Repeated-Visual LVLM Inference

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Repeated-visual LVLM inference answers multiple questions about the same image, document, chart, or screenshot. Exact prefix reuse saves repeated encoding but retains the full visual representation, while compression can discard evidence needed by a later question. We propose VPCache, which selects reusable cache views according to paired answer sufficiency. Its Answer-Preserving Cache View Selector (AP-CVS) qualifies candidate views within question cells, prioritizes answer quality, and freezes the resulting routes before evaluation. A learned high- fidelity view, APVC-hifi, refines a deterministic anchor for reuse across questions. On 448 held-out requests from seven workloads with Qwen2.5-VL-7B-Instruct, VPCache adds two soft and two exact matches over the guarded base, with no paired soft losses against either the base or full-prefix inference. Static compression controls incur 25–72 paired soft losses against the base. Across all held-out requests, VPCache reduces injected visual tokens by 4.90% and selected-view memory slots by 2.46% on average, matching the base’s representation budget while improving answer quality. These results support question-cell selection as a way to govern reusable visual compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large vision–language models, Multimodal inference, Visual memory routing, Query-aware compression, KV-cache optimization, Risk-aware inference, Visual question answering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fBryOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes VPCache for repeated-visual LVLM inference. It constructs reusable compressed visual representations and selects among cache views based on question type and answer-preservation criteria, with fallback to a safer base representation when compression is not certified. The method is evaluated on seven visual QA workloads using Qwen2.5-VL-7B-Instruct, showing small accuracy gains over the guarded base while reducing the average injected visual-token count.

Critical strengths
  1. The paper identifies a practically relevant problem: preserving answer quality when reusing compressed visual representations across repeated queries.

  2. The proposed framework is carefully designed around safe fallback and paired answer preservation, which makes the system more reliable than applying a fixed compression strategy uniformly.

Critical weaknesses

1. Limited methodological novelty

The methodological contribution appears limited relative to the existing literature on visual-token/cache compression and query-aware selection. The proposed APVC largely combines familiar ingredients—pooling/salient-token based initialization, attention-based feature aggregation, residual refinement, and query-conditioned slot selection—while AP-CVS adds a validation-based routing and fallback mechanism over a bank of existing or learned compressed views. Although this combination is reasonable and the answer-preservation criterion is useful, it is unclear that the paper introduces a sufficiently new compression or caching principle beyond composing existing techniques with an empirical routing rule. This substantially weakens the paper's technical contribution.

2. Insufficient experimental evidence for the claimed efficiency and generality

The experimental evaluation does not convincingly establish that VPCache is practically efficient or broadly applicable.

  • Limited baselines: The main comparisons use Fresh/Exact reuse, TokenPool, and KIVI-style/SnapKV-style proxies rather than strong implementations of several directly relevant recent visual-token and VLM cache-compression methods discussed in the paper itself. This makes it difficult to assess the proposed method against the actual state of the art.

  • Limited model generality: All main experiments use only Qwen2.5-VL-7B-Instruct. Since the proposed framework is presented as a general LVLM inference method, evaluation across different LVLM architectures is important to determine whether the observed behavior is model-specific.

  • Limited workload scale and diversity: The evaluation uses only 448 held-out requests (64 per workload), with a request-level rather than image-disjoint split. Moreover, the repeated-visual setting contains only about 2.95 questions per visual input on average. There is no video or large multi-frame benchmark, where visual context and cache size would be substantially larger and the claimed benefits of reusable compression would be more meaningful.

  • Modest effective compression: Although individual compressed paths can reduce the representation substantially, only 44 of 448 held-out requests actually use a shorter visual prefix. Consequently, the workload-level reduction is only 4.90% in injected visual tokens and 2.46% in selected memory slots. These reductions are quite small relative to the paper's efficiency motivation.

  • Missing end-to-end efficiency measurements: The paper does not provide controlled end-to-end latency, throughput, or serving-cost comparisons for the final method. Representation/token reduction alone does not establish actual system-level efficiency, particularly because routing, slot selection, and maintaining multiple candidate cache views can introduce additional overhead.

  • Incomplete memory analysis: The only peak-memory experiment is a small auxiliary diagnostic that predates the final APVC-hifi method. Moreover, its reported on-disk cache footprint provides no storage advantage and is even larger than Exact L1 caching in that diagnostic. Thus, the claimed memory benefit is not demonstrated for the final system.

  • No compression-cost/amortization analysis: APVC requires constructing a compressed memory for each visual input, but the cost of this construction is not quantified. For a repeated-query setting, the key quantity is the break-even point at which the one-time compression cost is amortized by subsequent queries. No FLOPs, construction latency, or efficiency-versus-number-of-queries analysis is provided.

Questions for authors
  1. What is the key methodological novelty of VPCache beyond combining existing compression, query-aware selection, and validation-based routing/fallback mechanisms?

  2. How does VPCache compare with stronger and more directly related visual-token and VLM cache-compression methods under the same backbone, compression budget, and evaluation setting, using faithful implementations rather than proxy variants?

  3. Does VPCache generalize to LVLM backbones other than Qwen2.5-VL-7B-Instruct?

  4. Would the conclusions remain consistent when evaluated on larger benchmark subsets, image-disjoint splits, and the official metrics of each benchmark?

  5. How does VPCache scale as the visual context becomes larger and the number of repeated queries per visual input increases, including multi-image or video settings?

  6. Given that compression is applied to only a limited fraction of requests, how does the quality–compression trade-off change when targeting substantially higher compression coverage or more aggressive compression ratios?

  7. What are the end-to-end latency and throughput improvements of the final VPCache system under controlled and identical serving conditions?

  8. What are the actual GPU-memory and cache-storage savings of the final APVC-hifi system, including all coexisting cache views and fallback representations?

  9. What is the computational cost of constructing the compressed cache, and after how many repeated queries is this one-time cost amortized relative to exact prefix reuse or fresh inference?

2 comments
AuthorsOpenReview ↗

Part 1 of 2: contribution, FastV/FoLDER comparison, InternVL and image-disjoint test

Thank you for the questions about technical contribution, cross-architecture behavior, and serving efficiency. VPCache connects a reusable visual bank to question-dependent view selection and QA-based qualification; each experiment below evaluates one part of that design.

Q1. A reusable bank with independently controlled views

VPCache separates reusable capacity, per-question extraction, and eligibility. APVC constructs an image-dependent, question-independent bank; the hifi point retains approximately 0.75N slots and injects approximately 0.5N tokens for an eligible question. Different questions select views from that shared bank. AP-CVS compares paired generated-answer scores on calibration data and freezes a dataset × question-class routing table. Online selection reads the question and table.

This decomposition makes the contribution testable: the fixed-budget control tests representation choice; three seeds test bank learning; independent-versus-transferred routing tests qualification; the serving trace charges construction and fallback. A full-budget interface control additionally reproduces all 147 Exact answers when native spatial grids are retained, confirming that the bank interface leaves uncompressed features unchanged.

Q2. Common-prompt comparison with FastV and FoLDER

Conditional view selection retains document evidence at this operating point: VPCache scores 80.42 DocVQA and 64.94 InfoVQA, compared with FastV's 69.34/44.46 and FoLDER's 68.63/45.87. The differences to FastV are +11.08 and +20.48 points.

Additional benchmarkFrozen VPCacheExactGuarded baseGuarded APVCFastV port + ExactFoLDER port
DocVQA80.4285.6280.4279.5969.3468.63
InfoVQA64.9464.9464.9464.9444.4645.87
ChartQA75.3977.3475.3975.0069.5360.94
TextVQA74.3080.3174.3074.5779.3476.80
AI2D78.2178.2178.2178.2173.5474.32
VQAv2, uniform79.2681.4879.2679.8879.6578.29

All methods use the same 1,434 questions, checkpoint, dataset templates, BF16, greedy decoding, a 262K-pixel input limit, and a 32-token output cap. VPCache compresses 241 requests (16.81% coverage, 8.73% total-token reduction); these FastV/FoLDER settings compress 100% of requests. The table specifies the resulting quality–coverage choices, including Exact's uncompressed reference. Under the conditional VQAv2 rule, the scores are VPCache 78.48, Exact 81.48, guarded base 78.48, guarded APVC 79.11, FastV 79.65, and FoLDER 75.56.

Q3. Qualification transfers across LVLM interfaces

The same qualification interface selects architecture-specific operators on InternVL2.5-2B. With frozen seed-0 weights and 448×448 single-tile preprocessing, our Exact path reproduces the native InternVL inference output on one calibration case from each benchmark. A matched 128-question VQAv2 prompt control yields 1.28 words per answer versus 5.59, with uniformly normalized scores of 68.67 versus 27.50.

The Exact-qualified table selects compressed base in TextVQA/binary and VQAv2/count, and Exact elsewhere. It compresses 42/1,434 requests and reduces total tokens by 1.46%.

Additional benchmarkQualified policyExactPaired difference, pp [95% CI]
TextVQA67.7768.16−0.39 [−1.17, 0.00]
VQAv2, uniform70.5470.78−0.23 [−1.79, 1.33]

On InternVL, qualification selects the base and Exact views that preserve this architecture's answers, so the QA-based qualification mechanism transfers across interfaces. The APVC learning comparison is reported separately at matched candidate budgets.

Q4. A further image-disjoint benchmark test

The additional test contributes 1,434 questions on 521 further images beyond the 3,080-question earlier test. Its images are checked at the pixel level against the fitting, qualification, and test partitions, and all methods are scored on the identical question set.

Whole image groups are selected independently of outcomes. Benchmark-specific scoring and paired image-cluster resampling preserve both the dataset's metric and dependence among questions about one image. A 485-question conservative ChartQA source-exclusion diagnostic is reported as a separate control.

AuthorsOpenReview ↗

Part 2 of 2: video contexts, coverage, latency, memory and amortization

Q5. Distinct questions and 32-frame contexts

Reuse reduces repeated visual work on distinct questions. On the archived 65-question / six-image natural-reuse trace, VPCache measures 206.93 ms versus Fresh 310.55 ms, a 33.37% reduction, with 8–16 distinct questions per image.

The expanded video diagnostic evaluates 64 additional clips and 256 questions, sampling 32 distinct frames per case. The image-trained checkpoint and policy are fixed.

32-frame methodCorrect / 256Mean latency, sPeak host feature tensors, MiB
Fresh2022.4480
Exact2021.8691047.59
Base2001.9161717.38
APVC1991.9261717.38

APVC reduces total visual tokens by 29.25% at 55.08% request coverage, with an observed 21.35% latency reduction relative to Fresh, although it was trained only on images. APVC−base accuracy is −0.39 points [−1.17, 0.00]; APVC−Exact is −1.17 [−3.52, 1.17], using paired clip-cluster intervals. Fresh and Exact predictions match verbatim.

Q6. Bank capacity and selective compression

Bank capacity and injected prefix are separate operating-point controls. On the archived 141-request stress fixture, hifi APVC scores 49.20, the smaller-bank checkpoint 46.39, and the shorter-injection candidate 37.34. The hifi point is 2.81 macro points above the smaller-bank configuration and 11.86 above the shorter injection. The bank configurations use their respective fixed checkpoints; the injection comparison reuses the hifi checkpoint. Guarded base on that fixture is 61.38 and includes its eligibility/fallback decisions.

Coverage is an explicit policy setting. The frozen Qwen selector compresses 547/3,080 archived requests and 241/1,434 additional requests, reducing total visual tokens by 9.05% and 8.73%, respectively. Qualifying against Exact changes the additional-test operating point to 30 compressed requests. The 0/0.2/0.5/1/2-point mean-calibration-loss margins all select 125 compressed requests, at 8.72% coverage and 4.43% token reduction, so the selection is stable across this range; the grid was fixed before test scores were inspected.

Q7. Complete online latency and throughput

The 6.82% latency and 7.32% throughput improvements over guarded base include the online selection path. The archived measurements use 389 distinct requests, two randomized-order repetitions, and batch-one sequential execution. Preprocessing, routing, construction, cache access, decoding, and fallback are charged. Mean generated lengths are 5.14 tokens for base and 5.08 for VPCache.

Mean latency is 192.99 ms for VPCache, 207.10 ms for guarded base, 187.15 ms for Exact, and 225.02 ms for Fresh; FoLDER's archived mean is 182.54 ms. The paired VPCache−base difference is −14.12 ms [−19.81, −8.41].

Q8. Physical cache, shared features, and deployment assets

Shared L1 removes duplicate visual-encoder work, while a common LRU policy enforces the declared feature-storage capacity. In the complete first repetition, shared VPCache reduces native vision-encoder calls from 173 to 157. The unrestricted cache remains 328.18 MiB because its representations coexist.

Physical feature-cache policyExact, MiBVPCache, MiB
Unrestricted273.26328.18
Common 160 MiB LRU, final stored cache158.81158.00
Common 160 MiB LRU, peak before eviction162.14162.11

Both methods remain within 160 MiB after each request, and capacity enforcement leaves all predictions unchanged. Under the same capacity, VPCache stores slightly less than Exact.

Archived allocated GPU peaks are 15.585 GiB for base/Exact and 15.678 GiB for VPCache; host memory is 3.208 GiB for base and 3.213 GiB for VPCache. The APVC checkpoint is a separate 98.15 MiB deployment file.

Q9. Construction cost under repeated-question reuse

Reuse distributes construction cost across subsequent questions about the same visual input. Cold APVC construction, including first-use loading, is 186.7–236.9 ms and is charged when the bank is first materialized. The outcome-independent request trajectories show the following measured costs:

Requests per imageFresh, msExact, msVPCache, ms
R=1281.37277.67274.72
R=8274.88194.52196.44
R=32269.87182.79184.11

At R=32, the VPCache and Exact points are 184.11 and 182.79 ms. For the independently calibrated selector, VPCache is already faster than Fresh from the first or second request: crossings against Fresh/Exact occur at R=2/32 on binary visual questions and R=1/1 on the other three standard-resolution categories. The separate 65-question natural-reuse result confirms the saving on distinct questions.

These measurements anchor the contribution in three linked choices: learning reusable visual evidence, qualifying the view for each question, and charging its complete online cost.

Reviewer_Fr3xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work tackles the inference overhead of large vision language models when processing the same image repeatedly by proposing a cache view selection framework called VPCache. To address the pain point that indiscriminate global compression of visual features can compromise question specific accuracy, VPCache aims to preserve answer correctness through two complementary components. First, it employs a compressor named APVC, which leverages gated fusion and evidence aware training to extract key visual representations. Second, it uses a selector named AP CVS to qualify candidate views based on actual QA correctness on development data, freezing the routing decisions prior to final testing. Evaluated on seven VQA benchmarks using Qwen2.5-VL-7B, VPCache matches the baseline compression rate, reducing visual tokens by 4.90 percent and memory slots by 2.46 percent, while correctly answering two additional queries on the held out set, improving both soft and exact matches, without introducing paired soft-match regressions.

Critical strengths

S1. A meaningful formulation of answer-preserving cache selection.

The paper addresses an important limitation of visual compression: a representation sufficient for one question may discard information needed for another. Evaluating candidate views through paired correctness rather than aggregate accuracy alone is a useful perspective, as it explicitly accounts for regressions on previously correct answers. The frozen routing policy and guarded fallback provide a practical mechanism for controlling such regressions.

S2. A unified evaluation framework with component-level analysis.

The paper compares exact reuse, deterministic memory, static compression proxies, and learned representations under a shared inference setup. The component ablation in Table 6 helps distinguish the contributions of cache-view selection and learned compression. The authors also explicitly acknowledge limitations in their efficiency measurements, which makes the scope of the empirical claims easier to assess.

Critical weaknesses

W1. Limited incremental improvements and insufficient evidence of practical efficiency benefits.

Empirical improvements are marginal: on 448 held-out requests, VPCache adds only two soft matches and two exact matches over FacetGuard-base. Table 6 shows identical average token and memory-slot reductions (4.90% and 2.46%). Thus, the selector does not demonstrate any incremental reduction in the reported representation-cost metrics over the guarded baseline. The learned APVC contributes just one extra exact match over the static-bank selector, with no soft-match gain. The main experiments lack controlled latency, throughput, and cache-memory measurements; the KIVI/SnapKV baselines are proxy implementations, not full systems. Whether the added complexity yields real-world advantages remains unclear.

W2. Concerns about evaluation independence and the reliability of answer-preservation claims.

The same development requests are used for both APVC training and AP CVS qualification, without an independent calibration set. Appendix A.1 reveals that diagnostics on the combined 896 request matrix, which includes held-out requests, informed memory-budget and objective choices, raising concerns about the independence of the final evaluation and potential test-set-informed model selection. View qualification relies on very small samples, for example only 12 development examples for infographic counting. Zero paired losses on such tiny sets do not guarantee preservation on unseen requests. While certification is defined empirically, stronger independent validation is needed, especially since avoiding answer regression is the paper's central goal.

Questions for authors

Q1 (Regarding W2). Independence of the held-out evaluation.

Appendix A.1 says diagnostics on the combined 896 request matrix guided memory budget and objective configuration choices. Could the authors specify which design decisions were affected? Can the main results be verified on a fully untouched test set, ideally with disjoint visual inputs and an independent calibration split?

Q2 (Regarding W1). Incremental efficiency benefits.

Table 6 shows VPCache and FacetGuard base have identical average token and memory slot reductions. What measurable end to end efficiency advantages does VPCache offer over the guarded baseline? Specifically, could the authors report controlled latency, throughput, and actual resident cache memory measurements, including costs from view construction and repeated query reuse?

Q3 (Regarding W1 & W2).. Robustness and contribution of APVC.

The learned APVC adds only one extra held-out exact match over the static bank selector. Does this improvement hold across independent test sets or different random seeds? Also, do the reported zero paired soft losses remain under stricter answer equivalence criteria or the official evaluation metrics of the respective benchmarks?

1 comment
AuthorsOpenReview ↗

Independent qualification, latency–storage operating point and repeatable APVC gains

Thank you for emphasizing paired answer outcomes and component attribution. The evidence connects learned representation quality, calibration-based eligibility, and complete online cost through separate controlled comparisons.

Q1. Fixed design and independent qualification

Independent qualification contributes +1.38 InfoVQA ANLS points over transferred routing, with a paired 95% interval of [0.19, 2.63]. This directly evaluates the routing decision with the representation checkpoint held fixed.

The APVC-hifi checkpoint, 0.75N reusable-bank capacity, approximately 0.5N injected budget, and evidence-weighted objective are fixed for the additional Qwen evaluation. The learning protocol separates 393 fitting, 772 qualification, and 3,080 test requests by image identity. The further 1,434-question test excludes the earlier images.

An Exact-qualified variant uses the 772 calibration questions with corrected prompts, at least eight examples per dataset × question-class cell, and zero individual calibration-score losses. Its table is frozen before the additional test.

Q2. The measured latency–storage operating point

The online selection pipeline converts a measured view-storage budget into lower latency than guarded base. On the archived 389-request trace, mean latency is 192.99 versus 207.10 ms, sequential throughput is 5.18 versus 4.83 requests/s, and TTFT is 109.18 versus 117.77 ms. The paired latency difference is −14.12 ms [−19.81, −8.41].

The alternative views add 9.64 MiB of feature-cache storage and 95.24 MiB of peak allocated GPU memory relative to base, corresponding to less than 1% of the base GPU allocation. Timing charges preprocessing, routing, online construction, disk access, generation, and fallback; model startup is excluded. For reference, Exact runs at 187.15 ms with 273.26 MiB of feature cache, and Fresh at 225.02 ms. This accounting states both the system benefit and the resource assigned to achieve it.

Q3. Repeatable APVC gains and empirical preservation

All three APVC initializations improve VQAv2 under the archived fixed guarded protocol.

InitializationVQAv2 gain, ppPaired 95% CI, ppUniform-normalization gain, pp
Seed 0+2.14[0.83, 3.52]+2.73
Seed 1+1.68[0.26, 3.24]+2.26
Seed 2+2.07[0.70, 3.56]+2.65

The intervals use 5,000 paired image-cluster resamples. The archived complete selector changes 215 routes, with five score improvements, 208 ties, and two decreases relative to base.

Answer preservation holds on the independent test. On the additional 1,434-question test with corrected prompts, the frozen policy compresses 241 requests with zero per-question score losses relative to guarded base. This is the behavior the preservation criterion is designed to produce on unseen images. For the separate Exact-qualified variant, 30 requests receive compression; VQAv2 is 79.69 versus Exact 81.48. For full-coverage raw candidates on the additional test, APVC−base macro difference is +0.91 [−0.22, 2.08] points on Qwen and +0.22 [−0.74, 1.18] on InternVL. These equal-weight six-benchmark candidate summaries are reported separately from the guarded pipeline and from official individual benchmark metrics.

Reviewer_SLHYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies repeated-visual LVLM inference, where multiple questions share the same visual input. VPCache combines calibrated routing with a learned visual compressor to select reusable representations for repeated questions about the same image. Experiments show modest answer-quality improvements over a guarded baseline at the same representation budget.

Critical strengths
  1. Practical problem. The paper identifies a practical challenge in reusable visual compression: a representation sufficient for one question may discard evidence needed by another. Selecting representations according to question categories is a reasonable response to this issue.
  2. Useful component analysis and cost accounting. The paper separates the contributions of routing and learned compression, reports both soft and exact answer changes, and averages representation savings over all requests.
Critical weaknesses
  1. Limited benefit. The improvement over FacetGuard-base is only 2/448 for both soft and exact matches, approximately 0.45 percentage points.
  2. Evaluation independence. Appendix A.1 states that diagnostics over the combined development and held-out data informed memory-budget and objective configurations. Although the held-out requests were excluded from parameter fitting and route selection, they were therefore not untouched during method design. The single-backbone evaluation and small development cells further limit evidence for generalization.
  3. Unclear practical efficiency. Proxy baselines and token/slot reductions do not establish end-to-end speedups or total resident-memory savings.
  4. Incomplete discussion of related work. The paper lacks a sufficiently comprehensive review and analysis of closely related studies, such as [1, 2, 3].

[1] Chen L, Zhao H, Liu T, et al. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 19-35.

[2] Wang H, Yu Z, Spadaro G, et al. Folder: Accelerating multi-modal large language models with enhanced performance[C]//2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025: 23614-23625.

[3] Ju C, Wang H, Cheng H, et al. Turbo: Informativity-driven acceleration plug-in for vision-language large models[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 436-455.

Questions for authors
  1. Can the method be evaluated on a fully untouched test set, with uncertainty estimates and separate results for unseen images?
  2. Do the quality gains persist under official benchmark metrics, and how many exact-match changes reflect formatting rather than semantic correctness?
  3. What end-to-end latency and total resident-memory savings remain after accounting for view construction, coexisting representations, and exact fallback?
1 comment
AuthorsOpenReview ↗

Representation choice, image-disjoint test, official metrics and serving cost

Thank you for focusing on the value of representation choice, independent evaluation, and complete serving costs. Our central claim is that reusable visual representations should be learned and qualified for the questions they will serve.

W1. Representation choice at a fixed budget

The three-initialization APVC experiment shows that the learned bank improves VQAv2 over guarded base by +2.14, +1.68, and +2.07 points, with positive lower bounds for all three paired 95% intervals. The fixed-budget control isolates representation choice on 448 questions, adding two correct answers with both injected-token and selected-memory-slot budgets unchanged.

The complete selection pipeline is tested online: on the archived 389-request trace, it changes latency from 207.10 to 192.99 ms relative to guarded base, with a paired difference of −14.12 ms [−19.81, −8.41].

Q1. Evaluation on independent visual inputs

Fitting, qualification, and testing are separated by image identity. The additional test extends this separation to another 521 images.

PartitionQuestionsImages
Fitting393138
Qualification772274
Earlier test3,0801,078
Additional test1,434521

Pixel-level checks verify zero overlap among the three earlier partitions and between the additional images and those partitions. Additional image groups are selected in a fixed order independent of outcomes, with 2–4 distinct questions per image and no filtering by correctness. InfoVQA contributes 151 eligible questions; the other benchmarks contribute 256–257 each. Paired confidence intervals use 5,000 image-cluster bootstrap resamples.

Q2. Absolute model scores and benchmark protocols

Exact is the uncompressed model reference; guarded base is a conditional compression policy. We report the same 1,434 additional questions under two explicit Exact configurations, using dataset VQA/MCQ templates, BF16, and greedy decoding.

BenchmarkExact: 262K pixels / 32 tokensExact: native pixel limits / 2048 tokens
DocVQA85.6294.84
InfoVQA64.9478.91
ChartQA77.3480.47
TextVQA80.3187.58
AI2D78.2175.88
VQAv2, uniform normalization81.4880.43

The native configuration uses the native Qwen2.5-VL pixel limits. Its paired changes are +9.22 [5.47, 13.34] DocVQA, +13.97 [4.88, 22.27] InfoVQA, and +7.27 [3.32, 11.37] TextVQA points.

A matched 147-question calibration control isolates prompting at 262K pixels: ChartQA changes from 55.56 to 88.89 on 18 questions, and conditionally normalized VQAv2 from 61.03 to 79.66 on 29 questions. We retain benchmark-specific metrics and both VQAv2 conventions: the additional main table uses uniform normalization; native-reference VQAv2 under the archived conditional rule is 76.93.

Q3. Online efficiency and the Exact operating point

At the declared compressed operating point, VPCache delivers 6.82% lower latency and 7.32% higher sequential throughput than guarded base. The archived trace contains 389 distinct requests on 136 images, with two randomized-order repetitions.

Archived serving methodMean latency, msRequests/sSerialized feature cache, MiB
Fresh225.024.440
Exact L1187.155.34273.26
Guarded base207.104.83318.55
VPCache, independent qualification192.995.18328.18

The paired VPCache−base latency difference is −14.12 ms [−19.81, −8.41]; VPCache−Exact is +5.83 ms [−0.46, 12.15], an interval that includes zero. FoLDER's archived mean is 182.54 ms; its QA operating point is given under W4.

VPCache allocates storage to alternative views and Exact fallback, while controlling the prefix injected for an eligible request. In the first repetition, the cache holds 258.67 MiB of fallback features, 49.68 MiB of base views, and 19.83 MiB of APVC views. Under a common 160 MiB LRU policy, the complete first-repetition cache audit gives VPCache 158.00 MiB and Exact 158.81 MiB, with unchanged predictions.

W4. Positioning relative to token-reduction methods

VPCache makes the reusable-view eligibility decision explicit. FastV uses decoder attention for token pruning; FoLDER supplies feature merging; Turbo concerns informativity-driven token reduction. VPCache learns a persistent bank and selects an admitted view for each question class using paired calibration outcomes.

Our executable Qwen comparisons use an attention-based FastV port at K=2 and the FoLDER authors' merging function on projected features. The comparison reports both QA and compression coverage: the frozen VPCache policy compresses 16.81% of additional requests, while these FastV and FoLDER settings compress every request. At this setting VPCache scores 80.42 DocVQA and 64.94 InfoVQA, against 69.34/44.46 for FastV and 68.63/45.87 for FoLDER.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.