- 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 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
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:
| Method | Sampled peak GPU process memory over the full run (GiB) |
|---|---|
| Full re-encoding | 24.689 |
| PKR + IVE | 24.916 |
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.
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 (), and the remaining inference settings held fixed:
| Routing threshold | VisualProbe accuracy | End-to-end speedup | KV reuse rate |
|---|---|---|---|
| 0 | 10.3% | 0.55× | 70.6% |
| 0.1 | 45.7% | 0.81× | 61.2% |
| 0.2 | 54.8% | 1.39× | 44.0% |
| 0.3 | 55.1% | 1.22× | 27.9% |
| 0.5 | 54.4% | 1.02× | 2.7% |
| 1.0 | 54.6% | 1.00× | 0.0% |
At or , 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 , 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 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.
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), regional mapping, and remaining inference settings. MPIC-32 recomputes the K/V states of the first 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.
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.