acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Visual KV Cache Reuse for Efficient Multi-Turn Iterative Multimodal Reasoning

30 views

Abstract

Multi-turn visual reasoning acquires evidence through repeated cropping and local inspection, but overlapping observations lead to redundant visual encoding and prefill. We propose an adaptive visual KV cache reuse framework that reduces repeated processing while retaining access to additional image detail. The central difficulty is that skipping visual prefill also removes the interaction between a new observation and the current reasoning: historical keys remain bound to old positions, while historical values have not incorporated the intent formed between crops. Our framework first selects full re-encoding or regional reuse according to crop size and cache availability. On the reuse path, Positional Key Rebinding (\PKR) corrects attention matching at the new positions, and Intent-guided Value Enhancement (\IVE) uses pre-crop reasoning attention to strengthen content relevant to the current objective. Source states from fully re-encoded crops are retained for subsequent retrieval and reuse. The framework requires no additional training; on reuse hits, it bypasses the visual encoder and prefill for the selected visual tokens, then resumes normal text reasoning. Experiments show that our method reduces inference time while preserving the overall accuracy of multi-turn visual reasoning that relies on intermediate crop images.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-turn visual reasoning, KV cache reuse, efficient multimodal inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Q58jOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper reduces repeated visual processing in multi-turn reasoning by reusing cached visual KV states and processing fresh crops when greater magnification is requested.

It reuses visual KV states from earlier observations when the requested region is sufficiently large and the cache is available, while re-encoding smaller crops from the original pixels to obtain finer detail.

For reused states, it corrects the keys’ positional encodings and slightly strengthens relevant values using attention from the recent reasoning, without additional training.

On Mini-o3, the method reports up to 1.40× end-to-end speedups across five benchmark settings while maintaining similar overall accuracy.

Critical strengths

1, The paper retains full re-encoding for crops that request greater magnification, allowing the model to acquire detail absent from the cache, and requires no additional training.

2, The paper identifies both positional mismatch and missing interaction with recent reasoning, then addresses them through key-position correction and attention-guided value enhancement.

3, The paper avoids repeated visual encoding and prefill by reusing regional KV states, reporting 1.30–1.40× end-to-end speedups while maintaining similar overall accuracy.

Critical weaknesses

1, The novelty needs clearer evidence. Prior works, e.g., MPICand CONDUIT, already address multimodal cache reuse, positional relocation, or contextual adaptation. What does regional crop reuse and IVE add beyond these methods? Direct comparisons under matched conditions would help establish the contribution.

2, Figure 4 includes cache adaptation, but it is unclear where routing, regional mapping, source-bank maintenance, and attention extraction are counted. Figure 5 measures incremental per-crop VRAM rather than total memory. Could the authors clarify the accounting and report total peak memory as the number of turns increases?

3, The fixed area threshold of 0.2 does not directly measure whether the cache contains sufficient detail. Several successive crops can each pass this threshold while covering an increasingly small region of the cached source, potentially delaying fresh encoding.

Questions for authors

1, Could the authors reconcile the crop-reuse rate with the visual-processing savings? Figure 4 reports 44% crop reuse but a 64% reduction in visual-processing time, from 50 to 18 units. Under identical crop trajectories, this would require the reused crops to be about 2.3× more expensive on average than the remaining crops. Could the authors provide crop-level token counts and timings, and separate savings from reuse from changes in the reasoning trajectory?

2, What computation is included in the reuse FLOPs? Key rotation and value scaling alone appear much cheaper than the approximately (1010)–(1011)(10^{10})–(10^{11}) FLOPs plotted in Figure 5. Attention extraction or recomputation could explain the difference, but the algorithm describes attention as already saved during decoding. Could the authors give the FLOP breakdown and clarify which attention operations introduce additional work?

3, What memory is included in the incremental VRAM measurement? With 28 layers, four KV heads, head dimension 128, and BF16 storage, a 2,304-token visual KV block alone occupies approximately 126 MiB. Could the authors report the actual number of reused tokens and clarify whether Figure 5 includes newly inserted KV states, retained source states, and attention buffers? Reporting total peak memory across increasing numbers of turns would make the memory trade-off much clearer.

8 comments
AuthorsOpenReview ↗

Response (1/6) — Weakness 1

Weakness 1. Technical contributions and comparisons with existing cache-reuse methods.

A1. Thank you for your comments on the technical contributions and the need for direct comparisons. Our work studies how to continually acquire new visual detail during dynamic cropping and local zooming while reducing redundant visual encoding and prefill. Existing work, including MPIC and CONDUIT, has investigated cache reuse and adaptation. However, when cropping and zooming produce new visual features, the existing cache lacks the corresponding KV states, and positional correction or contextual compensation of existing states cannot provide these new states. Applying cache reuse to this process therefore requires addressing when to re-encode, when to reuse, and how to incorporate newly encoded views into subsequent reuse.

To address these challenges, our framework integrates adaptive routing, Positional Key Rebinding (PKR), and Intent-guided Value Enhancement (IVE). Routing selects between re-encoding and regional reuse based on the crop-to-parent area ratio and cache availability, retaining newly encoded observations as cache sources for subsequent crops. During reuse, PKR aligns the positions of cached keys, while IVE uses pre-crop reasoning attention to apply bounded, token-wise gains to the original cached values, strengthening evidence relevant to the current inspection objective. Our core contribution is to unify new-evidence acquisition, cache-position alignment, and intent adaptation, enabling cache reuse to support multi-turn reasoning with changing observation regions, resolutions, and reasoning objectives.

To evaluate this design, we have added direct comparisons on VisualProbe using Mini-o3. Across reuse variants, we fix the routing threshold (τ=0.2\tau=0.2), regional mapping, and remaining inference settings. MPIC-32 recomputes the K/V states of the first min⁡(32,N)\min(32,N) visual tokens in the requested region. LegoLink-0 prepends four auxiliary tokens when constructing each source cache, removes their K/V entries after prefill, and uses its own positional handling during reuse. Neither adapted baseline additionally applies our PKR or IVE, and all associated overheads are included in end-to-end timing.

MethodAccuracyEnd-to-end speedup
Full re-encoding54.6%1.00×
Direct reuse40.3%0.92×
PKR only50.4%1.41×
MPIC-32 (adapted)51.9%1.28×
LegoLink-0 (adapted)52.6%1.20×
PKR + IVE54.8%1.39×

The complete method exceeds MPIC-32 and LegoLink-0 by 2.9 and 2.2 percentage points, respectively, while also achieving higher end-to-end speedup. With routing and PKR fixed, adding IVE increases accuracy from 50.4% to 54.8%, while speedup changes from 1.41× to 1.39×. The controls in Appendix A.5.2 further show that, even with the mean gain matched, IVE outperforms uniform scaling by 3.2 percentage points (54.8% vs. 51.6%). These results support the additional contribution of intent-aligned, token-wise gain allocation and the accuracy–efficiency advantage of the complete framework in dynamic cropping.

AuthorsOpenReview ↗

Response (2/6) — Weakness 2

Weakness 2. Timing accounting and total peak GPU memory across turns.

A2. Thank you for highlighting the need to clarify the accounting of computational overhead and memory usage. Figure 4 reports end-to-end latency for complete multi-turn inference, including routing decisions, regional mapping, source-cache construction and maintenance, attention acquisition, and cache adaptation. Attention acquisition includes storing reasoning Queries, reconstructing attention using existing Keys, and subsequent aggregation and normalization. The 1.39× speedup reported on VisualProbe therefore includes these costs.

Figure 5 is a controlled component-level benchmark, and its memory metric is component-level incremental VRAM rather than total memory usage during complete inference. In the revised manuscript, we explicitly clarify this scope in the figure caption and restrict the corresponding conclusion in the main text to the component-level measurement.

To supplement this measurement with full-inference memory usage, we profile a six-crop example on the same GPU with a fixed crop and text trajectory and prefix caching enabled for both methods:

MethodSampled peak GPU process memory over the full run (GiB)
Full re-encoding24.689
PKR + IVE24.916

In this example, PKR + IVE cumulatively reuses 1,618 visual tokens, while sampled peak process memory increases by 0.227 GiB, approximately 0.92%. Retained source-cache storage grows from 141.75 MiB to 285.03 MiB and is already included in the memory measurements. In our response to Question 3, we additionally report the per-turn PyTorch allocated-memory peaks and retained source-cache sizes for all six crop steps, separately from the full-run sampled process-memory peaks, to show both cache growth and memory usage.

AuthorsOpenReview ↗

Response (3/6) — Weakness 3

Weakness 3. Preservation of visual detail during successive cropping and the routing threshold.

A3. Thank you for highlighting the potential loss of detail during successive cropping. We agree that the parent-relative area threshold is a routing heuristic and cannot, by itself, guarantee sufficient visual detail relative to the fully encoded ancestor after multiple crops. This motivates combining re-encoding with cache reuse and examining sensitivity to the routing threshold.

In our framework, crop requests fall into two categories: selecting a local region from the original image, and further cropping an existing local observation. The former always undergoes full re-encoding to acquire new local detail and establish a regional cache. The latter is routed to reuse or re-encoding according to the area threshold and cache availability, with newly encoded observations retained to support subsequent crops. For example, Figure 15 shows two re-encoding steps that acquire a fine-grained representation of the target region, followed by three consecutive nested reuse steps. Although the subsequent regions progressively shrink, the required details have already been acquired through the preceding re-encoding steps and preserved in the cache, allowing the model to complete the recognition task correctly.

The additional threshold experiments further characterize this trade-off. The following results are reported in Appendix A.5.1, Figure 8, with PKR, IVE (β=0.15\beta=0.15), and the remaining inference settings held fixed:

Routing threshold τ\tauVisualProbe accuracyEnd-to-end speedupKV reuse rate
010.3%0.55×70.6%
0.145.7%0.81×61.2%
0.254.8%1.39×44.0%
0.355.1%1.22×27.9%
0.554.4%1.02×2.7%
1.054.6%1.00×0.0%

At τ=0\tau=0 or 0.10.1, some samples exhibit repeated crop requests until the budget of at most 12 rounds is exhausted, accompanied by decreases in both accuracy and end-to-end efficiency. At τ=0.2\tau=0.2, we did not observe this unproductive repeated-cropping behavior in our evaluation; 44.0% of requests use reuse and the remaining 56.0% undergo re-encoding, achieving 54.8% accuracy and 1.39× speedup. Increasing the threshold to τ=0.3\tau=0.3 improves accuracy by only 0.3 percentage points, while reducing speedup to 1.22×. These results support an effective accuracy–efficiency trade-off on the evaluated tasks by using adaptive re-encoding to acquire new detail and regional reuse to reduce redundant computation.

Reviewer_2gcsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes training-free visual KV cache reuse for iterative multimodal reasoning. Crop requests are routed to either full re-encoding or regional cache reuse based on crop size and cache availability. Reused keys receive updated rotary positions through Positional Key Rebinding (PKR), while Intent-guided Value Enhancement (IVE) scales cached values using pre-crop reasoning attention. Experiments on Mini-o3 across five benchmark settings report comparable overall accuracy and 1.30–1.40× end-to-end speedups.

Critical strengths
  1. A practical reuse setting with meaningful visual dependence. The paper targets redundant processing across successive crop observations rather than compression within a single image. Removing intermediate images substantially reduces Mini-o3’s accuracy, providing a useful check that the evaluation actually requires intermediate visual evidence.

  2. Promising measured efficiency with preserved accuracy. The method reports 1.30–1.40× end-to-end speedups while maintaining overall accuracy across all five settings. The time breakdown includes cache adaptation, and the PKR-only versus PKR+IVE comparison demonstrates additional accuracy gains from value enhancement. Immutable source states also prevent cumulative value amplification during consecutive reuse.

Critical weaknesses
  1. The comparisons do not sufficiently establish the contribution of the proposed adaptations. Mini-o3 with full re-encoding is the only direct efficiency baseline; the related cache-reuse and selective-refresh approaches discussed in Section 2.2 are not evaluated. Table 3 compares PKR-based reuse against removing intermediate images, which establishes the benefit of restoring visual evidence but does not isolate positional rebinding from reuse itself. Although adding IVE improves accuracy, no uniform-gain or shuffled-attention control tests whether the improvement specifically depends on current intent rather than generic amplification of visual values.

  2. The systems trade-off is incompletely characterized. Figure 5 reports incremental per-crop VRAM, not total peak memory including the persistent source bank and retained attention information. The paper acknowledges these additional storage costs but does not quantify them. It also does not explain sufficiently how full-context attention is collected across all layers within the vLLM implementation or separately measure its overhead. Consequently, the reported speedups are encouraging, but the complete compute–memory trade-off remains unclear.

Questions for authors
  1. Under the same routing rule, how do direct reuse, PKR-only reuse, and PKR+IVE compare in accuracy, repetition rate, and latency? Can you include a relevant position-corrected reuse or selective-refresh baseline from Section 2.2?

  2. Does IVE outperform uniform value scaling and shuffled token gains with matched average amplification? These controls would test whether the current reasoning-derived relevance pattern explains its benefit.

  3. What is the total peak GPU memory, including the source bank and attention buffers, as the number of crop turns increases? Please clarify the attention-collection implementation and whether its full overhead is included in end-to-end timing.

  4. How sensitive is the accuracy–latency trade-off to the routing threshold τ and enhancement strength β? In particular, can repeated crops exceed the parent-relative threshold while becoming too small relative to the fully encoded ancestor to retain adequate detail?

6 comments
AuthorsOpenReview ↗

Response (1/4)

Q1. Comparisons with direct reuse and existing cache-reuse methods.

A1. Thank you for suggesting more direct comparisons. We have conducted additional experiments on VisualProbe using Mini-o3. Across the reuse variants, we fix the routing rule (τ=0.2\tau=0.2), regional mapping, and remaining inference settings, and compare direct reuse, PKR-only reuse, PKR+IVE, and two adapted baselines.

The two existing methods are adapted to regional crop requests as follows:

  • MPIC-32 (adapted): For an eligible crop containing NN visual tokens, we recompute the K/V states of the first min⁡(32,N)\min(32,N) tokens in sequence order under the current context and reuse the cached states for the remaining tokens. This adapts MPIC’s selective recomputation mechanism to the requested region. [1]
  • LegoLink-0 (adapted): Following the zero-recomputation variant in EPIC, we prepend four auxiliary tokens when constructing each source observation’s cache and remove their K/V entries after prefill. For subsequent eligible crops, we retrieve the corresponding regional visual cache and insert it into the current context using the method’s own positional handling, without recomputing visual tokens during reuse. [2]

Both adapted baselines use the same crop-routing rule and regional mapping while retaining their own cache-processing mechanisms, without additionally applying our PKR or IVE. Selective recomputation and additional cache-construction costs are included in end-to-end timing. Speedups are measured relative to full re-encoding.

MethodAccuracyEnd-to-end speedup
Full re-encoding54.6%1.00×
Direct reuse40.3%0.92×
PKR only50.4%1.41×
MPIC-32 (adapted)51.9%1.28×
LegoLink-0 (adapted)52.6%1.20×
PKR + IVE54.8%1.39×

Positional correction contributes beyond reuse itself. Direct reuse retains the cached keys’ original positional bindings and leaves values unchanged. It achieves 40.3% accuracy and 0.92× speedup, meaning that it does not provide an end-to-end efficiency gain over full re-encoding. With the same routing rule, PKR-only reuse raises accuracy to 50.4%, an improvement of 10.1 percentage points, while achieving 1.41× speedup. Thus, retrieving cached visual states alone is insufficient in this setting; correcting their positions substantially improves the measured accuracy–efficiency trade-off.

IVE further improves accuracy at comparable speed. Adding IVE increases accuracy from 50.4% to 54.8%, a gain of 4.4 percentage points, while speedup changes from 1.41× to 1.39×. This corresponds to approximately 1.4% longer end-to-end latency than PKR-only reuse. The complete method achieves accuracy comparable to full re-encoding (54.8% vs. 54.6%) while retaining 1.39× acceleration.

The adapted baselines improve accuracy over PKR-only reuse, but the complete method achieves a better accuracy–efficiency trade-off. MPIC-32 reaches 51.9% accuracy with 1.28× speedup, and LegoLink-0 reaches 52.6% with 1.20× speedup. PKR+IVE exceeds them by 2.9 and 2.2 percentage points, respectively, while also achieving higher end-to-end speedup. These comparisons support the effectiveness of combining positional correction with intent-guided value enhancement for dynamic regional reuse.

References

[1] Zhao, Shiju, et al. “MPIC: Position-Independent Multimodal Context Caching System for Efficient MLLM Serving.” arXiv preprint arXiv:2502.01960, 2025.

[2]Hu, Junhao, et al. “EPIC: Efficient Position-Independent Caching for Serving Large Language Models.” Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025.

AuthorsOpenReview ↗

Response (2/4)

Q2. Does IVE outperform uniform value scaling and shuffled token gains with matched average amplification?

A2. Thank you for suggesting these controls. We have added controlled comparisons in Appendix A.5.2 (p. 18, Figure 9) to examine whether IVE’s gains arise from intent-aligned weighting beyond simply increasing value magnitudes.

We fix the model, PKR, routing rule (τ=0.2\tau=0.2), and inference configuration, changing only value adaptation. We compare five variants.

  • PKR only uses the original cached values without enhancement.
  • Uniform scales every selected value by 1+β=1.151+\beta=1.15.
  • Uniform-Mean computes IVE’s gains and applies their arithmetic mean uniformly within each crop, layer, and KV head.
  • Shuffled randomly permutes the IVE gains within each crop, layer, and KV head, preserving their distribution and mean while disrupting their correspondence with visual tokens.
  • IVE applies the original intent-aligned, token-wise gains.
Value adaptationVisualProbeV* BenchHR-Bench 8K
PKR only50.4%83.6%71.9%
Uniform51.1%84.5%72.2%
Uniform-Mean51.6%84.1%72.2%
Shuffled48.7%82.6%71.1%
IVE54.8%87.4%73.1%

Uniform scaling provides modest improvements. With the mean gain matched, IVE still improves accuracy over Uniform-Mean by 3.2, 3.3, and 0.9 percentage points, respectively. Shuffled falls below PKR only on all three benchmarks, and IVE exceeds Shuffled by 6.1, 4.8, and 2.0 percentage points. These controls support an additional benefit from aligning reasoning-derived gains with the corresponding visual tokens.

AuthorsOpenReview ↗

Response (3/4)

Q3. Total peak GPU memory and attention-acquisition overhead.

A3. Thank you for this suggestion. Figure 5 reports component-level incremental memory usage. We additionally profile memory on a single GPU using a fixed crop and text trajectory. We report the two methods separately to distinguish their operations at each crop step.

Table 1. Baseline: full re-encoding at every crop.

Crop stepOperationPer-turn allocated peak (GiB)
1Full re-encoding23.248
2Full re-encoding22.973
3Full re-encoding22.914
4Full re-encoding22.972
5Full re-encoding23.211
6Full re-encoding22.952

Table 2. PKR + IVE: adaptive re-encoding or KV reuse.

Crop stepOperationVisual tokens reused at this stepPer-turn allocated peak (GiB)Retained source cache (MiB)
1Full re-encoding023.385141.75
2KV reuse + IVE55223.324141.75
3Full re-encoding023.177157.50
4KV reuse + IVE56023.339157.50
5Full re-encoding023.489285.03
6KV reuse + IVE50623.453285.03

Shared table note. Both methods use the same GPU and fixed crop and text trajectory, with prefix caching enabled. Per-turn allocated peaks are measured by PyTorch and cover preparation, crop processing, and subsequent reasoning within each turn, including the model, context KV pool, retained source caches, and Query/attention buffers. Source-cache size is the retained source KV storage at the end of the turn and is already included in the memory measurements.

Overall peak memory. Across the complete inference run, the sampled peak GPU process memory is 24.689 GiB for the baseline and 24.916 GiB for PKR + IVE. PKR + IVE cumulatively reuses 1,618 visual tokens, with an increase in process peak memory of 0.227 GiB, approximately 0.92%. Process memory is sampled through NVML every 5 ms and is reported separately from the per-turn PyTorch allocated-memory peaks above.

The retained source cache grows from 141.75 MiB to 285.03 MiB, while per-turn peaks also depend on the temporary computation required by each crop. For example, the third crop encodes 288 visual tokens, compared with 2,592 in the first crop, so its allocated-memory peak is lower despite the growth in retained source cache.

For attention acquisition, we cache the pre-crop reasoning Queries during decoding and use existing contextual Keys to reconstruct the visual attention statistics defined in Equation (6). This reconstruction includes additional QK computations over the context and the image intervals. Within each layer and KV head, we aggregate attention over the relevant reasoning tokens and query heads sharing that KV head, then select and normalize within the crop region to obtain the token-wise weights πi\pi_i. The costs of storing Queries, reconstructing attention, aggregation, and normalization are all included in the reported end-to-end latency; the 1.39× speedup reported on VisualProbe therefore includes these costs.

Reviewer_Yh4cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an adaptive visual KV cache reuse framework for efficient multi-turn multimodal reasoning. The method combines area-based routing, Positional Key Rebinding (PKR), and Intent-guided Value Enhancement (IVE) to avoid redundant visual encoding and prefill during iterative cropping. Experiments on five benchmarks demonstrate 1.30–1.40× end-to-end speedups while largely preserving the original model's accuracy.

Critical strengths
  1. Practical motivation and efficiency gains. The paper addresses a meaningful computational bottleneck in iterative visual reasoning, where repeated cropping can introduce substantial redundant computation. The proposed training-free framework is practical and integrates with the existing inference pipeline without modifying model parameters.
  2. Promising empirical results. The method is evaluated across five visual reasoning benchmarks and consistently achieves 1.30–1.40× end-to-end speedups while maintaining comparable accuracy. The results demonstrate the practical potential of visual KV cache reuse for accelerating multi-turn reasoning.
Critical weaknesses
  1. Substantial overlap with existing work.
    The proposed method has substantial overlap with prior KV cache reuse methods, especially Position Rebinding Cache Reuse (PRCR) [1]. PRCR has already identified the decoding collapse caused by stale positional bindings and proposed rebinding cached keys through RoPE. Both the motivation and the PKR mechanism are very similar to this work. Kamera [2] has also explored position-independent multimodal KV reuse and contextual compensation.
    The main differences seem to be the adaptive routing strategy and IVE. However, the routing strategy is based on a simple area threshold, while IVE mainly applies attention-based scaling to cached Values. It is not clear what the main technical novelty is beyond combining these components with existing KV reuse techniques. I would like to see a clearer discussion of how this method differs from prior work, especially PRCR.
  2. Insufficient ablation studies.
    Table 3 shows that adding IVE improves accuracy, but it does not fully explain where the improvement comes from. In particular, it is unclear whether the gains are due to the proposed intent-guided weighting or simply to increasing the magnitude of visual Values. Simple baselines such as uniform Value scaling or shuffled attention weights are missing.
    Similarly, the adaptive routing strategy is not sufficiently studied. The authors use a fixed area threshold but do not show the sensitivity to this choice or whether it performs better than simpler routing policies.
    These ablations are important for understanding which parts of the method actually contribute to the improvement, particularly given the overlap with prior work.

[1] Wang, Mengzhao, et al. "Position Rebinding Cache Reuse: Replay-Free Visual Revisiting for Interleaved Multimodal Reasoning." arXiv preprint arXiv:2606.26631 (2026).

[2] Ma, Bole, et al. "Kamera: Unified Position-Invariant Multimodal KV Cache for Training-Free Reuse." arXiv preprint arXiv:2606.23581 (2026).

Questions for authors

I suggest authors to further clarify the difference between this work and the piror methods. Also, further exploring the effect of each proposed components is encouraged which could help others understand the paper. I would like to improve my rate if the questions are sufficiently answered.

4 comments
AuthorsOpenReview ↗

Response (1/2)

Q1. Substantial overlap with existing work.

A1. Thank you for your comments on the technical contributions. Our framework further extends and improves upon PRCR, and addresses a key challenge left unresolved by PRCR and Kamera—how to continually acquire new visual detail during multi-turn cropping and dynamic resolution changes while reducing redundant visual encoding and prefill.

PRCR focuses on reusing cached states through positional correction, while Kamera uses low-rank compensation to adapt them to changes in context. Both mechanisms rely on already encoded visual evidence. However, when cropping and zooming produce new visual features, the existing cache lacks the corresponding KV states, and positional correction or contextual compensation cannot generate these new states. This leaves the questions of when to re-encode, when to reuse, and how to incorporate newly encoded views into subsequent reuse.

To address these problems, our framework integrates adaptive routing, Positional Key Rebinding (PKR), and Intent-guided Value Enhancement (IVE). Routing selects between re-encoding and reuse based on the crop-to-parent area ratio and cache availability, retaining states from newly encoded observations for subsequent crops. During reuse, PKR aligns the positions of cached keys, while IVE uses pre-crop reasoning attention to apply bounded, token-wise gains to the original cached values, strengthening evidence relevant to the current objective. Our core contribution is to unify new-evidence acquisition, cache-position alignment, and intent adaptation, extending KV reuse from repeated access to already encoded views to multi-turn visual reasoning with continually acquired views, changing resolutions, and evolving inspection objectives.

AuthorsOpenReview ↗

Response (2/2)

Q2. Insufficient ablation studies.

A2. Thank you for pointing out the need for additional ablations. We have added three studies in Appendix A.5 (pp. 17–19, Figures 8–10) to examine the source of IVE’s gains, routing policies and threshold sensitivity, and enhancement-strength sensitivity.

1. Do IVE’s gains come solely from increasing value magnitudes? (A.5.2, Figure 9)

We fix the model, PKR, routing rule (τ=0.2\tau=0.2), and inference configuration, changing only value adaptation. We compare five variants.

  • PKR only uses the original cached values without enhancement.
  • Uniform scales every selected value by 1+β=1.151+\beta=1.15.
  • Uniform-Mean computes IVE’s gains and applies their arithmetic mean uniformly within each crop, layer, and KV head.
  • Shuffled randomly permutes the IVE gains within each crop, layer, and KV head, preserving their distribution and mean while disrupting their correspondence with visual tokens.
  • IVE applies the original intent-aligned, token-wise gains.
Value adaptationVisualProbeV* BenchHR-Bench 8K
PKR only50.4%83.6%71.9%
Uniform51.1%84.5%72.2%
Uniform-Mean51.6%84.1%72.2%
Shuffled48.7%82.6%71.1%
IVE54.8%87.4%73.1%

Uniform scaling provides modest improvements. With the mean gain matched, IVE still improves accuracy over Uniform-Mean by 3.2, 3.3, and 0.9 percentage points, respectively. Shuffled falls below PKR only on all three benchmarks, and IVE exceeds Shuffled by 6.1, 4.8, and 2.0 percentage points. These controls support an additional benefit from aligning reasoning-derived gains with the corresponding visual tokens.

2. Routing-threshold sensitivity and simpler routing policies (A.5.1, Figure 8)

We vary τ\tau on VisualProbe while keeping PKR and IVE (β=0.15\beta=0.15) fixed. Setting τ=0\tau=0 removes the area restriction for otherwise eligible requests, while τ=1\tau=1 always re-encodes.

Routing threshold τ\tauAccuracyEnd-to-end speedupKV reuse rate
010.3%0.55×70.6%
0.145.7%0.81×61.2%
0.254.8%1.39×44.0%
0.355.1%1.22×27.9%
0.554.4%1.02×2.7%
1.054.6%1.00×0.0%

The τ=0.2\tau=0.2 point retains the reported main-experiment result. At τ=0\tau=0, high reuse is accompanied by substantially lower accuracy and slower end-to-end inference. At τ=0.2\tau=0.2, accuracy is 54.8% with 1.39× speedup, compared with 54.6% accuracy for full re-encoding. Increasing τ\tau to 0.30.3 yields 55.1% accuracy and 1.22× speedup. Across the tested thresholds from 0.20.2 to 1.01.0, accuracy remains within 54.4%–55.1%, while speedup decreases as reuse becomes more conservative. These results characterize the accuracy–efficiency tradeoff and show that maximizing reuse does not necessarily improve end-to-end efficiency.

3. Enhancement-strength sensitivity (A.5.3, Figure 10)

We fix τ=0.2\tau=0.2, PKR, and the remaining inference configuration, and vary the enhancement strength β\beta on two benchmarks. Setting β=0\beta=0 corresponds to PKR-only reuse.

Enhancement strength β\betaVisualProbeMME-RealWorld-Lite
050.4%59.2%
0.0553.5%61.7%
0.1053.0%62.8%
0.1554.8%64.5%
0.3054.1%63.2%

The results at β=0.15\beta=0.15 and the VisualProbe PKR-only reference point retain the previously reported values. All tested positive enhancement strengths outperform PKR only, by 2.6–4.4 percentage points on VisualProbe and 2.5–5.3 points on MME-RealWorld-Lite. Increasing β\beta from 0.150.15 to 0.300.30 reduces accuracy by 0.7 and 1.3 points, respectively. These results show that enhancement benefits hold across multiple tested settings, while accuracy does not increase monotonically with enhancement strength.

Thanks for the author's reply! My second concern has been solved but there is still a quesiton remaining. I agree that iterative cropping introduces an additional challenge of acquiring new visual evidence. However, my main concern is about the technical novelty of the proposed components. In particular, PKR appears very similar to the position-rebinding mechanism in PRCR. The response mainly highlights differences in the application setting rather than explaining how the underlying mechanism differs. While the added routing strategy and IVE provide practical benefits, I am still not fully convinced that the proposed framework offers sufficient technical novelty beyond combining existing KV reuse techniques with relatively simple additional components.

Thank you for your follow-up comments and for confirming that the additional ablations have addressed your second concern. Regarding the technical overlap, we would like to clarify the following distinctions.

Both PRCR and PKR use RoPE rebinding, but their positional assignment differs. PRCR normalizes selected coordinates and maps them into a controlled interval; PKR applies a shared offset that preserves the original pairwise coordinate differences among the selected tokens.

PRCR uses attention to select tokens while leaving cached values unchanged; IVE uses pre-crop reasoning attention to apply bounded, token-wise gains within the requested region. Each enhancement operates on the original cached values, avoiding cumulative amplification.

At the framework level, PRCR focuses on reinserting previously encoded states. Our framework adaptively routes between re-encoding and regional reuse, retaining newly encoded observations as cache sources for subsequent crops. This unifies the acquisition of new visual detail, positional alignment through PKR, and intent adaptation through IVE.

An updated version of PRCR is now available on arXiv, and we would be grateful if you could consult this version when considering the comparison. Thank you again for your careful review and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.