- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes SRET, a training-free visual token reduction framework designed to preserve spatial reasoning capabilities in vision-language models (VLMs). The proposed method consists of three components: relation modeling based on visual similarity, shallow contextual features, and spatial coordinates; coverage-guided anchor adjustment with cross-group token budget reallocation; and structure-preserving evidence transport, which redistributes information from removed tokens to retained representations through norm-preserving updates. The authors evaluate SRET on two spatially specialized VLMs based on Qwen3-VL and InternVL3, across eight spatial reasoning benchmarks and five token retention ratios. The reported results show an average improvement of 2.060% over the strongest baseline when averaged across both models and all rentention ratios. Additional experiments investigate the contributions of selection and transport, alongside estimated computational costs and logical KV-cache usage.
Critical strengths
-
Comprehensive evaluation across multiple spatial reasoning tasks and compression budgets. The paper evaluates visual token reduction across eight benchmarks, two model variants, and five retention ratios. The evaluation covers single-image, multi-view, and video-based spatial reasoning. The consistent improvements reported under aggressive token compression suggest that preserving spatially distributed information is a worthwhile direction.
-
Attempts to isolate the contributions of representative selection and evidence transport. The module ablations in Table 2 and fixed-anchor transport interventions in Table 3 provide useful evidence that both representative selection and feature redistribution contribute to the reported improvements. In particular, the source-shuffling intervention is a meaningful attempt to distinguish appropriate source–receiver correspondence from the mere addition of removed-token information.
Critical weaknesses
1. The main performance claims are undermined by substantial evaluation sensitivity to answer parsing, while the corrected comparisons against baselines remain insufficiently reported. A major concern is the reliability of the reported performance comparisons, particularly given the findings disclosed in Appendix B.7. The authors acknowledge that the original BLINK answer parser incorrectly treats many valid InternVL3 responses as incorrect because it requires an answer-option letter at the beginning of the generated response. This causes a substantial discrepancy between the originally reported scores and the results obtained through supplementary answer parsing. Specifically, the authors report that:
- Across 250,932 prediction records, 29,289 outputs were initially unparsed.
- Supplementary parsing identifies 15,592 additional correct answers.
- The five-budget average BLINK accuracy of InternVL3 with SRET increases from 45.650% to 61.915%, a difference of 16.265%.
This is not a neglibigle evaluation artifact. It demonstrates that a substantial portion of the measured performance on this benchmark depends on response formatting rather than the underlying visual reasoning capability. More importantly, although the authors state that the parsing audit covers every method, Appendix B.7 reports the corrected results only for SRET. The principal comparisons in Table 1 continue to use the original strict parsing results, and the manuscript does not provide a complete baseline comparison or revised macro-average ranking under the supplementary parsing protocol. Consequently, it remains unclear whether SRET's reported improvements over competing approaches remain stable after correcting this evaluation artifact. Different token reduction methods may influence answer formatting differently, making it insufficient to assume that the same parser affects all methods equally.
This is particularly consequential because the headline improvement over the strongest baseline is only 2.060% when averaged across models and budgets. Without complete, consistently rescored comparisons, the extent to which this improvement represents genuine preservation of spatial reasoning rather than differences in output-format compliance remains uncertain. The issue is explicitly disclosed in the appendix, which is appreciated. However, acknowledging the evaluation problem without providing corrected baseline rankings and updated principal comparisons does not adequately establish the robustness of the central empirical claim.
2. The clamined inference efficiency is not substantiated by end-to-end measurements, despite the potentially substantial computational overhead of the proposed method. The paper positions SRET as an efficient token reduction framework. However, its efficiency analysis relies primarily on FLOPs and prefill KV-cache sizes rather than measured end-to-end inference performance. This limitation is particularly important because the proposed algorithm introduces potentially expensive operations beyond conventional token pruning. As explicitly described in Appendix B.1, constructing within-group token relations has complexity , where is the number of tokens in group . Furthermore, the method performs iterative anchor replacement, cross-group budget adjustment, endpoint matching, and evidence aggregation. The number of optimization iterations and their computational overhead may vary substantially with input resolution, the number of image tiles, and the length of multi-frame inputs. The reported resource analysis does not adequately establish whether these costs are offset by the reduced computation in subsequent decoder layers. In particular:
- No measured end-to-end latency, prefill latency, or throughput comparison is provided.
- The computational overhead of relation construction, iterative optimization, and evidence transport is not separately quantified.
- The reported KV-cache memory excludes temporary compression tensors and allocator overhead, as acknowledged in Appendix B.1.
- The primary accuracy-resource visualization in Figure 3 is restricted to InternVL3, leaving the practical efficiency of the other backbone insufficiently characterized.
Moreover, SRET preserves full visual encoding and the first two language-model decoder blocks, limiting its computational savings to the subsequent processing stages. Therefore, reducing the number of retained tokens does not necessarily imply an equivalent reduction in actual inference time or GPU memory consumption. Without direct latency measurements and peak memory profiling, the paper demonstrates an improvement in accuracy at fixed retained-token budgets, but does not convincingly establish that the proposed approach provides superior practical inference efficiency.
Questions for authors
1. Evaluation robustness and corrected baseline comparisons. Appendix B.7 reveals a substantial discrepancy between strict and supplementary BLINK parsing, including a 16.265% increase in InternVL3 SRET's five-budget average accuracy. Since the authors state that the parsing audit covers every method, could they provide the corrected BLINK results for all baselines and recompute the full Table 1 comparisons using a consistent supplementary parsing protocol? In particular, does SRET retain its reported 2.060% average advantage and its ranking across the ten model-budget configurations? Please provide the changes in relative performance, rather than only corrected absolute accuracy.
2. Actual inference latency and computational overhead. Could the authors provide measured end-to-end inference latency, prefill latency, throughput, and peak GPU memory consumption for SRET and the baselines under identical hardware and input conditions? Additionally, please clarify whether the estimated FLOPs in Figure 3 account for all computations introduced by SRET, including relation construction, iterative anchor optimization, cross-group exchanges, and evidence transport. A per-stage runtime breakdown, particularly for high-resolution and multi-frame inputs, would be important for establishing whether SRET actually accelerates inference compared with simpler pruning methods.
3. Baseline implementation details and comparison with recent spatially aware methods. The paper discusses spatially structured token compression approaches such as Nüwa, S2 Prune, and CoverPrune, but does not provide direct experimental comparisons with these methods. Given that spatial structure preservation is the central motivation of SRET, how does the method compare with the most directly related spatially aware token reduction approaches under comparable inference settings? Furthermore, could the authors clarify the exact compression layers, implementation adaptations, and hyperparameter configurations used for FastV, VisionZip, DivPrune, and GridPrune on the two backbones? These details are important for assessing the fairness and reproducibility of the main comparisons.
4. Generalization and performance degradation on video-based spatial reasoning. The experiments use SenseNova-SI-1.3 variants of Qwen3-VL and InternVL3, rather than demonstrating effectiveness on their standard publicly available base checkpoints. How much of SRET's improvement depends on the spatially specialized representations learned by these models? Moreover, Table A1a shows that SRET underperforms DivPrune by approximately 6.05% on ReVSI at 5% retention with Qwen3-VL. Can the authors explain this substantial degradation, particularly given that preserving spatial relations across observations is the primary motivation of the proposed method? An analysis of whether the degradation arises from per-view anchor allocation, local relation construction, or evidence transport would help determine the method's applicability to long video and multi-view reasoning.