betaFeedback

WorksComputer Science

SIEVE: Preserving Cross-Image Evidence via Depth-Aware Token Compression

Abstract

Multi-image vision-language models (VLMs) incur substantial computation when processing long visual sequences. Compressing these sequences must preserve complementary evidence across images and keep tokens from different images separate. Global selection can underrepresent question-relevant images, and pruning in early decoder layers risks removing relevant content before sufficient multimodal processing, yet existing methods spend most of the budget on these early layers. We propose SIEVE, a training-free, single-pass framework that jointly allocates visual tokens across images and decoder depth. SIEVE first pools each image into a compact spatial grid to reduce early processing cost. It then progressively selects representatives using prompt attention under equal per-image quotas and aggregates unselected tokens into similar representatives from the same image. These operations keep every image represented and its tokens separate from other images, while refining the retained representations. A closed-form schedule sets retention counts to target a prescribed average token budget across decoder layers. Across four VLM backbones, SIEVE achieves the highest average at a 6.25% budget on five multi-image benchmarks and four single-image benchmarks, and the highest mean on two video benchmarks. On LLaVA-OneVision-7B, it outperforms the strongest baseline by 9.96 and 9.03 points at 6.25% and 12.5% budgets, and at 12.5% it exceeds the uncompressed model by 2.01 points with 6.6x fewer prefill FLOPs.

Introduction

Multi-image VLMs integrate visual evidence across inputs [11, 20]. Their computational cost, however, grows with the number of images, as each image contributes a fixed token count. Token compression reduces this cost, but multi-image inputs raise a specific challenge. Task-relevant evidence can be complementary across images, so keeping more tokens from one image cannot compensate for removing the evidence needed from another. Effective compression must therefore preserve complementary visual evidence while controlling the cost of processing it.

Prior work reduces visual tokens before the decoder [1, 23], within the decoder [7, 22, 28], or at both stages [13, 25], and recent multi-image methods adapt retention to estimated image importance [27]. We instead treat aggressive multi-image compression as a joint token allocation problem across images and decoder depth. The retained tokens must support cross-image reasoning throughout inference while staying within a limited computational budget.

This requires three key properties. (1) Cross-image coverage. A global ranking can leave too few tokens for an image whose evidence the answer needs (Figure 1(a), bottom). (2) Image identity. Similar tokens from different images can describe different observations. If two images each contain a car, merging by similarity alone may fuse the two cars into one token, and the model can no longer tell which image each car came from or count them correctly. (3) Evidence throughout processing. Early reduction can remove useful tokens before their relevance is clear, whereas delaying reduction keeps costly tokens from every image in the early layers.

Figure 1. Overview of SIEVE. (a) Top: On COCO [15], attention to queried objects peaks in the latter half of decoder layers across backbones. Bottom: Fraction of global top-10% tokens assigned to the queried image (solid) and their overlap across questions (dashed), per backbone. Early rankings are largely query-agnostic, motivating per-image allocation and deferred attention-based selection. (b) Cumulative effect of each component at a 6.25% budget, averaged over five benchmarks. Pooling and budget scheduling provide most gains; progressive merging adds 2.27 points. (c) At a 12.5% budget, SIEVE leads on all five multi-image benchmarks with LLaVA-OneVision-7B, normalized to the uncompressed model (dashed ring: 100). (d) Three-image example: LLaVA-OneVision-7B answers correctly with only 6.25% of visual tokens. After per-image pooling, S(N)=3S(N) = 3 stages retain equal quotas per image and merge discarded tokens into their most similar retained tokens within the same image. Colors indicate the final token absorbing each region.

Figure 1(a) highlights two coupled challenges: when to select tokens and how to allocate the token budget across

images. In the latter half of the decoder, attention from the final prompt token to the queried object peaks, and 51–72% of globally top-ranked visual tokens shift toward the queried image as the question changes. In contrast, early-layer attention is unreliable for global token ranking. At layer 2, the global top-10% tokens overlap by 82–85% across questions targeting different images, allocate only the uniform 25% share to the queried image, and exhibit strong positional bias (e.g., 49% on the last image for LLaVA-OneVision and 63% on the first for Qwen3-VL). Yet compression cannot be deferred: retaining all visual tokens through layer 2 already consumes 117% of a 6.25% depth-averaged budget on LLaVA-OneVision-7B (see Eq. (2)). These observations motivate attention-independent early compression through per-image pooling and equal token quotas, followed by attention-guided selection at later layers.

Figure 1(b) progressively incorporates these components into one-step pruning. Pre-decoder pooling and budget-aware scheduling provide the largest gains by reducing early-layer costs. Equal per-image quotas prevent unreliable global rankings from suppressing individual images. Finally, similarity-based merging preserves information from unselected tokens, while repeated selection and merging progressively refine visual representations through image–text interaction.

We therefore propose SIEVE, a training-free, single-pass framework (Figure 1(d)). Spatial pooling first produces a compact grid for each image. Within the decoder, equal per-image quotas maintain coverage, prompt attention selects representatives, and each unselected token is merged into its most similar representative from the same image, which preserves image identity. An early selection stage limits processing cost, and later stages refine the representatives at greater depths. A closed-form schedule sets the number of retained tokens from the target budget, the stage locations, and the retention decay. SIEVE improves over pooling alone by 3.70 points, whereas applying FastV’s prune or PyramidDrop’s schedule to the pooled input gives no improvement (§ 6.1). Table 1 contrasts these mechanisms with those of prior methods.

MethodPre-decoder compressionText-guided reductionProgressive selectionPer-image coverageSame-image aggregation
VisionZip [23]✓✗✗✓†✓†
FastV [7]✗✓✗✗✗
PyramidDrop [22]✗✓✓✗✗
SparseVLM [28]✗✓✓✗✗
AVTP [27]✗✗✓✗✗
CLSE [5]✗✗✗✗✗
S²Prune [10]✗✗✗✓†✗
ToDRE [13]✓✓✗✗✗
SIEVE✓✓✓✓✓

Table 1. Comparison with prior methods. Unlike existing approaches, SIEVE unifies pre-decoding compression, text-guided progressive token selection, balanced per-image retention, and within-image token aggregation in a single framework.

We evaluate SIEVE on four VLM backbones with five multi-image, four single-image, and two video benchmarks. At a 6.25% budget, it achieves the highest multi-image and single-image averages on all four backbones and the highest video mean (Table 2, Table 5 and Table 6). On MuirBench prompts with at least five images, SIEVE leads on all four backbones at a 6.25% budget (Supp. Prompts with many images). At a 12.5% budget, it also reduces prefill FLOPs by 6.6× and prefill latency by 1.9× while exceeding the uncompressed model (Table 3). Our contributions are as follows:

MethodLLaVA-OneVision-7BQwen2.5-VL-7BQwen3-VL-8BInternVL3-8B
VenueMant.NLVR2MuirBLINKQB2AvgMant.NLVR2MuirBLINKQB2AvgMant.NLVR2MuirBLINKQB2AvgMant.NLVR2MuirBLINKQB2Avg
Full60.8389.2241.9649.8773.7063.1275.1287.9058.5857.4475.2070.8566.8287.1850.3158.8676.4067.9168.2085.7254.0456.3477.1068.28
Keep 6.25% of visual tokens
VisionZip [23]CVPR’2551.1566.4740.0841.4056.1051.0465.4470.2243.5845.7166.3058.2557.6066.8643.1938.0368.2054.7859.4572.7445.3846.2968.1058.39
FastV [7]ECCV’2436.8749.4526.0439.2445.8039.4842.8649.8626.6539.7246.2041.0630.8850.6243.8534.7747.1041.4441.4750.8420.9236.9345.2039.07
PyramidDrop [22]CVPR’2536.8749.4526.0439.2445.8039.4842.8649.8626.6539.7246.2041.0630.8850.6243.8534.7747.1041.4441.4750.8420.9236.9345.2039.07
AVTP [27]arXiv’2638.2550.7725.5838.6145.4039.7258.0656.6139.9639.8752.9049.4834.5652.5044.1234.3548.2042.7552.0754.5331.0436.3550.8044.96
CLSE [5]ECCV’26––––––––––––32.7250.5443.5832.4058.6043.57––––––
S²Prune [10]arXiv’2642.8667.8632.2340.2452.0047.0466.3666.1544.0342.9460.7056.0454.8461.7143.2339.5162.1052.2858.5363.6940.5842.0856.0052.18
ToDRE [13]ECCV’2652.0772.9241.8841.5652.7052.2370.5173.4748.6548.2468.9061.9559.9165.1547.2341.2967.8056.2866.3676.1751.8148.1973.8063.27
SIEVE63.5983.6147.7745.5070.5062.1971.4378.0550.9649.5068.4063.6760.8373.2047.2744.2968.5058.8267.7478.3050.9249.5071.4063.57
Keep 12.5% of visual tokens
VisionZip [23]CVPR’2555.3076.0442.8844.1962.1056.1067.7477.3849.1949.6170.3062.8459.9174.5143.7742.4072.8058.6863.5978.0248.2747.5071.8061.84
FastV [7]ECCV’2438.7152.0529.8540.0845.1041.1650.2353.2431.8542.1959.8047.4647.0055.7244.6545.5567.9052.1754.3861.1231.3841.6166.7051.04
PyramidDrop [22]CVPR’2539.1750.0132.3840.4546.4041.6840.0951.6030.4640.2955.2043.5341.9453.9845.6240.3568.0049.9848.3953.4126.5439.6160.9045.77
SparseVLM [28]ICML’2541.4753.9330.6940.4548.0043.1157.6060.1031.3445.0864.7051.6467.3645.6245.5572.0056.9668.6862.6832.6546.5570.5054.94
AVTP [27]arXiv’2641.9460.4029.9240.1448.7044.2263.1368.2447.3545.6662.1057.3052.5360.3344.8537.4561.6051.3559.9162.3539.5042.3560.5052.92
CLSE [5]ECCV’2641.9465.8735.9242.8247.9044.8970.9771.8447.2348.7168.1061.3759.9165.6143.3146.7373.3057.6361.2972.8340.1249.7671.3059.06
S²Prune [10]arXiv’2652.0777.0237.0844.4558.0053.7269.5978.1852.3147.8767.8063.1559.9172.5045.7743.2467.8057.8463.5973.6846.0046.8266.3059.28
ToDRE [13]ECCV’2653.0075.2742.5042.8256.4054.0069.5979.1254.0050.9273.6065.4563.5972.6147.8844.2471.9060.0471.8980.5154.0051.9775.9066.73
SIEVE65.9088.8349.3148.1973.4065.1372.8182.1652.3552.0371.7066.2165.4479.2548.0449.5071.9062.8369.1280.9052.1551.2972.6065.21
Keep 25% of visual tokens
VisionZip [23]CVPR’2562.2184.6642.8147.9769.3061.3970.5183.1352.2352.2972.7066.1763.5981.6146.3847.6175.9063.0265.4482.0350.8849.3475.4064.62
FastV [7]ECCV’2450.2370.2636.1943.4558.3051.6968.2072.2047.7352.4570.7062.2663.5973.1747.2753.0373.4062.0967.7482.4751.2353.7674.5065.94
PyramidDrop [22]CVPR’2552.5377.3938.1944.4565.7055.6569.1271.8544.3550.2475.2062.1567.7473.2547.1251.7674.4062.8567.7478.9648.8554.5575.1065.04
SparseVLM [28]ICML’2552.5377.3135.2744.3564.8054.8572.8182.6346.5853.9773.9065.9867.7482.1948.8556.2375.7066.1465.9083.0349.5455.0674.8065.77
AVTP [27]arXiv’2648.8569.2637.2341.3558.6051.0668.2076.5954.0850.1368.5060.5068.8372.6047.8143.1468.3058.5364.5270.5546.2347.1368.7059.43
CLSE [5]ECCV’2655.3078.9439.6943.8766.0056.7670.9782.1356.7753.1372.9066.7865.4479.5947.7351.0370.9062.3966.8282.3750.0453.2975.7065.64
S²Prune [10]arXiv’2656.2284.2042.4645.2466.0058.8273.7384.6856.0052.7672.4067.9162.2180.7248.2751.1870.9062.6666.3680.9549.5850.7171.6063.84
ToDRE [13]ECCV’2656.0881.4143.1946.3465.4058.8873.7383.7155.7754.1179.5068.1564.5279.8649.5051.3875.9066.0659.9185.5853.8756.7671.6067.11
SIEVE68.2090.4147.2751.2476.1066.6473.2784.4855.8554.8172.9068.2666.3683.3647.8854.0274.1065.1472.3582.9654.1254.9775.3067.94

Table 2. Performance comparison for multi-image benchmarks. The best average in each block is bold; Full is the model with no compression. A dash marks a method that cannot run at that budget, explained in Supp. Budgets a method cannot reach. At B=116B = \frac{1}{16}, Eq. (2) leaves FastV and PyramidDrop a single token after reduction, so their schedules coincide and their rows are identical.

MethodFLOPs (T) ↓Storage (MB) ↓Prefill (ms) ↓Acc ↑
Full26.46107.4100.463.12
Keep 12.5% of visual tokens
VisionZip [23]4.0616.947.056.10
FastV [7]4.1316.947.941.16
PyramidDrop [22]4.1316.952.041.68
SparseVLM [28]4.1316.953.843.11
AVTP [27]4.0716.946.644.22
CLSE [5]4.1016.954.446.89
S²Prune [10]4.1017.059.753.72
ToDRE [13]4.0616.971.354.00
SIEVE4.0216.753.565.13

Table 3. Efficiency analysis. Average across five benchmarks. Storage denotes key-value cache size. Prefill includes any pre-LLM token reduction. Measurements use a single NVIDIA H100.

MethodLLaVA-OneVision-7BQwen2.5-VL-7BQwen3-VL-8BInternVL3-8BAvg
Full58.7262.5966.3764.8163.12
Keep 6.25% of visual tokens
VisionZip [23]56.7454.8657.4955.4456.13
FastV [7]37.6435.3639.6738.7637.86
PyramidDrop [22]37.6435.3639.7538.7637.88
AVTP [27]39.0048.5246.8646.0845.12
CLSE [5]––40.32––
2Prune [10]46.7349.2653.7349.5749.82
ToDRE [13]57.4157.8159.3261.3658.97
SIEVE57.9658.5262.2160.3659.76
Keep 12.5% of visual tokens
VisionZip [23]58.8956.9761.4460.3259.41
FastV [7]46.8347.4853.3450.6049.56
PyramidDrop [22]43.6445.4248.9644.4845.63
SparseVLM [28]48.8355.2060.5156.6855.31
AVTP [27]45.2754.1954.0852.5551.52
CLSE [5]53.3556.3957.5860.0056.83
2Prune [10]56.2556.0860.1657.8857.59
ToDRE [13]58.4760.5061.1362.7060.70
SIEVE57.9360.6564.4861.7761.21

Table 5. Performance comparison for video benchmarks. Average across two benchmarks. Each clip enters as 16 frames sampled uniformly and handed to the decoder as images, so video runs the same code path as Table 2; the frame count is fixed for every method and budget.

MethodLLaVA-OneVision-7BQwen2.5-VL-7BQwen3-VL-8BInternVL3-8BAvg
Full80.6780.5981.7684.4481.87
Keep 6.25% of visual tokens
VisionZip [23]65.3355.4854.1761.0559.01
FastV [7]40.1338.8542.2840.1240.35
PyramidDrop [22]40.1338.8542.2840.1240.35
CLSE [5]——42.67——
2Prune [10]56.6354.9150.3059.1855.26
SIEVE69.6658.1959.5364.4462.96
Keep 12.5% of visual tokens
VisionZip [23]71.7661.2662.0866.1165.30
FastV [7]49.6245.9647.1458.4150.28
PyramidDrop [22]47.8845.9846.2349.3747.37
SparseVLM [28]52.8256.5258.8465.4758.41
CLSE [5]59.8467.0657.1571.7363.95
2Prune [10]68.8466.2758.8668.9865.74
SIEVE71.9765.2165.7075.4969.59

Table 6. Performance comparison for single-image benchmarks. Average across four benchmarks. SIEVE has the highest mean on all four backbones at 6.25% and on three at 12.5%. Per-benchmark results for all four backbones are in the supplementary material (see Supp. Single-image results).

  • We analyze image coverage, image identity, and compression timing, by measuring object attention across decoder depth. We also derive a feasibility condition for delayed reduction under a depth-averaged token budget.

  • We introduce SIEVE, a training-free framework that combines spatial pooling, per-image representative selection, and progressive same-image aggregation with a closed-form retention schedule.

  • We evaluate SIEVE across four VLM backbones on multi-image, single-image, and video tasks, showing strong performance under aggressive compression and measuring the contribution of each component.

Related Work

Visual token reduction. Several methods reduce visual tokens before the decoder. VisionZip [23] uses encoder attention to select tokens and aggregates the remaining tokens and CDPruner [26] conditions diversity-based selection on the instruction. Other methods apply token reduction within the decoder. FastV [7] prunes after an early layer, PyramidDrop [22] applies staged pruning, SparseVLM [28] uses text raters, and VScan [25] and ToDRE [13] combine reduction before and within the decoder. CLSE [5] scores tokens by the change in their high-frequency spatial information across adjacent decoder layers. S2^{2}Prune [10] splits each image into regions, keeps at least one token per region, then gives more tokens to regions with more structure. These in-decoder methods retain all tokens through the initial layers, consuming a substantial fraction of a restrictive depth-averaged budget. Most of them also score tokens by early attention, which our analysis identifies as a weak relevance signal. SIEVE instead keeps a coarse view of all images through pooling before the decoder, selects tokens at multiple decoder depths, and merges unselected tokens only into representatives from the same image.

Multi-image token reduction. Some methods are designed explicitly for multi-image inputs. AVAM [24] selects regions and decodes original and cropped inputs, requiring additional decoding. ERA [21] uses entropy-guided selection and token recycling. AVTP [27] selects the pruning layer and assigns image-dependent retention ratios using hidden-state importance scores. Allocations based on importance scores can be unreliable at early layers, and these methods do not explicitly control token counts across decoder depth under a strict depth-averaged budget. SIEVE instead assigns an equal quota to each image at every stage to maintain image coverage, restricts aggregation to tokens from the same image, and uses a schedule to match the depth-averaged budget.

Method

Notation

Inputs and visual tokens. Let I1,…,INI_1,\ldots,I_N be the NN input images and p the text prompt. A frozen vision encoder EE and projector gg map each image to a grid of visual tokens Xi=g(E(Ii))∈RPi×dX_i = g(E(I_i)) \in\mathbb{R}^{P_i \times d}. The total number of visual tokens is n=∑i=1NPin = \sum_{i=1}^{N} P_i.

Prompt sequence. Visual tokens are interleaved with the text embeddings to form the prefill sequence H(0)∈RT×dH^{(0)} \in\mathbb{R}^{T \times d}, and hth_t denotes row tt of the current H(ℓ)H^{(\ell)}. An owner map ω:{1,…,T}→{0,1,…,N}\omega: \{1,\ldots,T\} \to\{0,1,\ldots,N\} records the image of each position, with ω(t)=i\omega(t)=i for a visual token of image IiI_i and ω(t)=0\omega(t)=0 for a text token.

Decoder and surviving sets. The language model has LL decoder layers f0,…,fL−1f_0,\ldots,f_{L-1} with H(ℓ+1)=fℓ(H(ℓ))H^{(\ell+1)} = f_\ell(H^{(\ell)}), where LL depends on the backbone (§ 4.1). A stage after layer ℓ\ell acts on H(ℓ+1)H^{(\ell+1)}. We write V(ℓ)\mathcal{V}^{(\ell)} for the visual tokens that enter fℓf_\ell, so V(0)\mathcal{V}^{(0)} is the pooled set, and Vi(ℓ)={t∈V(ℓ)∣ω(t)=i}\mathcal{V}^{(\ell)}_i = \{t \in\mathcal{V}^{(\ell)} \mid\omega(t)=i\} for the tokens of image ii. With nℓ=∣V(ℓ)∣n_\ell= |\mathcal{V}^{(\ell)}|, we have n≥n0≥n1≥⋯≥nL−1n \ge n_0 \ge n_1 \ge\cdots\ge n_{L-1}. Text tokens are not counted.

Compression budget. The prefill cost of layer ℓ\ell grows with the number of tokens it processes, so we measure compression by the depth-averaged visual token count

nˉ=1L∑ℓ=0L−1nℓ=Bn,B∈(0,1],(1)\bar{n} = \frac{1}{L}\sum_{\ell=0}^{L-1} n_\ell= Bn, \qquad B \in(0,1], \tag*{(1)}

where BB is the token budget. Under a fixed depth-averaged budget, retaining more tokens in early layers requires more aggressive pruning in later layers.

Overview. Figure 1(d) summarizes the two-phase pipeline. The first phase pools each image before the decoder (§ 3.2). The second phase runs SS stages inside the decoder (§ 3.3). Each stage selects tokens with prompt attention under equal per-image quotas and merges every unselected token into its most similar kept token from the same image. A closed-form schedule sets how many tokens each stage keeps, to target Eq. (1). SIEVE is training-free, single-pass framework and uses no auxiliary models.

Pooling stage

Because Eq. (1) averages over all LL layers, tokens that survive until the first reduction layer ℓ1\ell_1 consume n0(ℓ1+1)n_0(\ell_1+1) of the total budget BnLBnL. Retaining all nn tokens up to layer ℓ1\ell_1 therefore consumes

n(ℓ1+1)BnL=ℓ1+1BL.(2)\frac{n(\ell_1+1)}{BnL}=\frac{\ell_1+1}{BL}. \tag*{(2)}

of the budget. This fraction grows as BB decreases and exceeds 1 when BL<ℓ1+1BL < \ell_1 + 1. In that case, no later pruning can meet the budget. We therefore spatially compress each image before the decoder. The pooled count must be small enough to keep this prefix cost low, yet larger than the budget BnBn, to retain candidates for subsequent selection. We set the pooled total to m=⌊min⁡(αB,β)n⌋m = \lfloor\min(\alpha B,\beta)n\rfloor, with α>1\alpha> 1 and B<β<1B < \beta< 1, where the headroom α\alpha keeps mm above the budget, leaving additional candidates for prompt-guided selection. The cap β\beta binds only when αB>β\alpha B > \beta, which no reported budget reaches (at B=14B = \frac{1}{4}, αB=β\alpha B = \beta), so it is not a tuned hyperparameter. Image ii receives mi=round⁡(mPi/n)m_i = \operatorname{round}(mP_i/n) tokens. Its grid is adaptively average-pooled to approximately mim_i tokens, each placed at the center of its pooling cell, which keeps the spatial layout of the image. Because grid sides are integers, the realized total m~\tilde{m} can differ slightly from mm.

Selection stage

The first stage is applied early, after layer ℓ1\ell_1, because every additional layer at full pooled width uses more of the budget and leaves fewer tokens for later stages. The other stages are placed in the second half of the remaining depth, where prompt attention best separates relevant tokens (Figure 1(a)), at layers

ℓj=ℓ1+⌊L−ℓ12⌋+⌊(j−1)(L−ℓ1)2S⌋,j=2,…,S.(3)\ell_j = \ell_1 + \left\lfloor\frac{L-\ell_1}{2}\right\rfloor+ \left\lfloor\frac{(j-1)(L-\ell_1)}{2S}\right\rfloor,\qquad j=2,\ldots,S. \tag*{(3)}

Merging happens after each fℓjf_{\ell_j}. The placement depends only on ℓ1\ell_1, SS and LL, with no per-benchmark tuning.

Prompt-attention scores. At stage jj, each visual token is scored by the attention it receives from the final prompt token. Let Ah,t(j)A_{h,t}^{(j)} be this attention weight to position tt in head hh of the NhN_h heads at layer ℓj\ell_j, computed from the layer input H(ℓj)H^{(\ell_j)}. We set

st=1Nh∑h=1NhAh,t(j)∑t′∈V(ℓj)Ah,t′(j),t∈V(ℓj).(4)s_t=\frac{1}{N_h}\sum_{h=1}^{N_h}\frac{A_{h,t}^{(j)}}{\sum_{t'\in\mathcal{V}^{(\ell_j)}}A_{h,t'}^{(j)}},\qquad t\in\mathcal{V}^{(\ell_j)}. \tag*{(4)}

Renormalizing over visual tokens gives every head an equal weight.

Per-image quotas. Global top-kk selection can allocate few or no tokens for some images, which may remove evidence needed for cross-image comparison. We therefore split the stage budget kjk_j uniformly across images,

qi=clip⁡(⌊kjN⌋,qmin⁡,∣Vi(ℓj)∣),q_i=\operatorname{clip}\left(\left\lfloor\frac{k_j}{N}\right\rfloor,q_{\min},\left|\mathcal{V}_i^{(\ell_j)}\right|\right),

with a floor of qmin⁡q_{\min} tokens per image and any remainder distributed greedily, so that ∑iqi=kj\sum_i q_i=k_j whenever Nqmin⁡≤kj≤nℓjNq_{\min}\leq k_j\leq n_{\ell_j}. Each image then keeps its qiq_i highest scoring tokens, Ki⊆Vi(ℓj)\mathcal{K}_i\subseteq\mathcal{V}_i^{(\ell_j)}, and K=⋃i=1NKi\mathcal{K}=\bigcup_{i=1}^{N}\mathcal{K}_i. The quota sets how many tokens each image keeps, while attention scores determine their identities.

Similarity merging. Let D=V(ℓj)∖K\mathcal{D}=\mathcal{V}^{(\ell_j)}\setminus\mathcal{K} denote the unselected tokens. Using the layer outputs H(ℓj+1)H^{(\ell_j+1)}, each t∈Dt\in\mathcal{D} is assigned to its most similar kept token from the same image, π(t)=arg⁡max⁡u∈Ki(t)cos⁡(ht,hu)\pi(t)=\arg\max_{u\in\mathcal{K}_i(t)}\cos(\mathbf{h}_t,\mathbf{h}_u), and merged into it by score-weighted averaging,

hu←suhu+∑t: π(t)=usthtsu+∑t: π(t)=ust,u∈K.(5)\mathbf{h}_u\leftarrow\frac{s_u\mathbf{h}_u+\sum_{t:\,\pi(t)=u}s_t\mathbf{h}_t}{s_u+\sum_{t:\,\pi(t)=u}s_t},\qquad u\in\mathcal{K}. \tag*{(5)}

Similarity-based assignment follows the token-merging principle of aggregating related representations [4]. Restricting π\pi to the same image keeps image identity, so similar tokens from different images, such as two cars in two images, are never fused. Removing this restriction lowers the average score by 1.09 points at a 6.25% budget (§ 6.1).

Sequence update. The sequence entering layer ℓj+1\ell_j+1 keeps all text tokens and K\mathcal{K}, so V(ℓj+1)=K\mathcal{V}^{(\ell_j+1)}=\mathcal{K}. Unselected tokens add no key-value entries to later layers.

Budget schedule. The kept counts follow a geometric schedule, kj=k1γj−1k_j=k_1\gamma^{j-1} with γ∈(0,1)\gamma\in(0,1). The first ℓ1+1\ell_1+1 layers process the pooled tokens, and stage jj covers the δj=ℓj+1−ℓj\delta_j=\ell_{j+1}-\ell_j layers after it, with ℓS+1=L−1\ell_{S+1}=L-1. Substituting into Eq. (1) gives

k1=BnL−m(ℓ1+1)∑j=1Sγj−1δj.(6)k_1=\frac{BnL-m(\ell_1+1)}{\sum_{j=1}^{S}\gamma^{j-1}\delta_j}. \tag*{(6)}

Each kjk_j is then rounded and clipped to [1,m][1,m]. In all our settings the smallest kSk_S exceeds Nqmin⁡Nq_{\min} by at least 7 tokens, so the quota floor never conflicts with the schedule.

If the numerator of Eq. (6) is not positive, the first ℓ1+1\ell_1+1 layers already use the whole budget. Pooling keeps the numerator positive at every evaluated budget. Since the schedule uses mm rather than the realized m~\tilde{m}, the depth average departs slightly from BnBn; all reported FLOPs and token fractions use the measured average.

Position indices. Pooled tokens are numbered from zero in raster order of the pooled grid. Inside the decoder, a kept token keeps its index, gaps included, and merging changes only its content.

Adaptive stage count. More stages let each image recompute its selection and merging with deeper hidden states, but they also add cost and repeated merging. We therefore set the number of stages from the image count,

S(N)=clip⁡(N,Smin⁡,Smax⁡),Smin⁡=2, Smax⁡=4.(7)S(N)=\operatorname{clip}(N,S_{\min},S_{\max}),\qquad S_{\min}=2,\ S_{\max}=4. \tag*{(7)}

A prompt with one or two images uses two stages, while one with four or more uses four. The floor of two ensures that single-image prompts are also reduced in more than one stage; with a floor of one they would be reduced only once, after layer ℓ1\ell_1. The cap limits the extra cost of scoring and merging. Stage locations and token counts then follow from Eqs. (3) and (6) with S=S(N)S = S(N).

Experiments

Setup

Models and implementation. We evaluate SIEVE on LLaVA-OneVision-7B [12], Qwen2.5-VL-7B [3], Qwen3-VL-8B [2], and InternVL3-8B [30]. SIEVE modifies only prefilling; decoding is greedy throughout. Qwen3-VL has 36 decoder layers, and the other backbones have 28. LLaVA-OneVision represents each image in a multi-image prompt with a 384×384384 \times384 view and 729 visual tokens; Qwen2.5-VL uses dynamic resolution, so token counts depend on the input. We use area-mean pooling for LLaVA-OneVision and bilinear resampling for the other backbones, whose encoders already merge spatial patches.

Hyperparameters. We use one configuration for all backbones, budgets, and benchmarks, with no benchmark-specific tuning: first merging layer ℓ1=2\ell_1 = 2, maximum stage count Smax⁡=4S_{\max} = 4, retention decay γ=12\gamma= \frac{1}{2}, per-image quota floor qmin⁡=4q_{\min} = 4, pooling headroom α=3\alpha= 3, and pooling cap β=0.75\beta= 0.75 (neither the cap nor the quota floor is binding in the evaluated configurations).

Benchmarks and baselines. We test on five multi-image tasks (Mantis-Eval [11], NLVR2 [19], MuirBench [20], BLINK [9], and Q-Bench2 [29]), four single-image tasks (TextVQA [18], ScienceQA-IMG [16], MMStar [6], and POPE [14]), and two video tasks (Video-MME [8] and EgoSchema [17]). The baselines are VisionZip [23], FastV [7], PyramidDrop [22], SparseVLM [28], CLSE [5], S2^2Prune [10], ToDRE [13], and AVTP [27].

Token budget. Comparisons use the same depth-averaged visual-token retention B∈{116,18,14}B \in\left\{\frac{1}{16}, \frac{1}{8}, \frac{1}{4}\right\} (Eq. (1)). A method that keeps the full visual sequence through layer 2 of a 28-layer decoder already spends 171% of the 6.25% budget before pruning (Eq. (2)), so FastV and PyramidDrop exceed this target even when reduced to one token.

Multi-image results

Performance under tight budgets. Table 2 shows that SIEVE achieves the highest mean score on all four backbones at B=116B = \frac{1}{16}, improving over the strongest baseline by 9.96 points on LLaVA-OneVision and 1.72–2.54 on Qwen2.5-VL and Qwen3-VL. SIEVE retains 87–99% of the uncompressed mean at B=116B = \frac{1}{16} and 93–103% at B=18B = \frac{1}{8}. On LLaVA-OneVision it exceeds the uncompressed model at B=18B = \frac{1}{8} and 14\frac{1}{4} (65.13 and 66.64 vs. 63.12). Figure 2 shows that SIEVE is the most accurate at every compute level.

Figure 2. Accuracy vs. compute on LLaVA-OneVision at B∈{116,18,14}B \in\left\{\frac{1}{16}, \frac{1}{8}, \frac{1}{4}\right\} (five-benchmark average). SIEVE has the highest accuracy at every compute level and, at B≥18B \geq\frac{1}{8}, exceeds the uncompressed model (dashed). x-axis: estimated visual-token prefill TFLOPs (Supp. Efficiency measurement and compute estimate); measured totals in Table 3.

Why baselines fail: compression timing. On LLaVA-OneVision at B=18B = \frac{1}{8}, FastV, PyramidDrop, and SparseVLM score only 41.16–43.11, trailing SIEVE by 22.0–24.0 points. They keep all tokens through layer 2, which already uses 86% of the 18\frac{1}{8} budget (Eq. (2)), leaving little capacity after pruning. Figure 3 shows how SIEVE keeps the queried objects through successive reductions at B=116B = \frac{1}{16}.

Retained visual tokens at each merging stage

Figure 3. Retained visual tokens at each merging stage, LLaVA-OneVision-7B at 6.25%. Darkened cells are no longer separate tokens. Stages follow Eqs. (3) and (7): reduction after layers 2, 19, 23 for the three-image question and 2, 21 for the two-image one. Red rings mark the object each question names (discount and cereal box).

Position robustness and joint evidence. Two controlled probes test whether the equal per-image quota keeps evidence usable (Table 4). (1) Position. When the relevant MMStar [6] image moves among N−1N - 1 distractors, SIEVE has the highest mean and worst-position accuracy, 3.6 and 4.8 points above the best baseline in mean accuracy at N=4N = 4 and 8. (2) Joint use. On COCO image pairs that need a fact from each image, SIEVE reaches 86.6% and 77.8% joint accuracy at N=4N = 4 and 8, above both the best baselines and the uncompressed model. When either image is replaced so that the answer changes, it answers both versions correctly about twice as often as the uncompressed model at N=8N = 8. Replacing the equal quota with global top-kk or with a floor plus attention lowers every metric, so the equal split itself matters. Together, these probes show that the equal per-image quota keeps the relevant evidence accessible wherever it appears and usable jointly across images, which per-image allocation or token ranking alone does not provide. Protocols and more results are in Supp. Robustness to relevant-image position and Supp. Joint preservation of evidence from two images.

MethodPosition N=4N = 4 MeanPosition N=4N = 4 WorstPosition N=8N = 8 MeanPosition N=8N = 8 WorstJoint evidence N=4N = 4 JointJoint evidence N=8N = 8 JointJoint evidence N=8N = 8 CF-AJoint evidence N=8N = 8 CF-B
Full52.951.248.145.277.163.923.623.2
VisionZip [23]43.141.240.137.867.260.720.519.8
FastV [7]30.529.030.027.850.844.02.22.3
PyramidDrop [22]28.827.028.626.849.942.02.73.2
SparseVLM [28]30.428.829.827.451.943.74.24.9
AVTP [27]33.532.033.532.058.054.60.71.0
CLSE [5]38.333.034.830.051.743.96.26.5
S²Prune [10]44.043.040.438.264.354.113.612.8
ToDRE [13]42.040.439.737.668.259.320.919.4
SIEVE w/ global top-kk45.843.043.337.485.074.642.037.1
SIEVE w/ floor + attention45.141.443.236.684.074.641.537.4
SIEVE47.647.045.242.286.677.847.342.5

Table 4. Position robustness and joint evidence on LLaVA-OneVision-7B at B=18B = \frac{1}{8}. Position moves the relevant MMStar image among N−1N - 1 distractors and reports mean and worst-position accuracy. Joint uses COCO image pairs whose question needs a fact from each image. CF-A and CF-B require correct answers on both the original pair and the pair with image A or B replaced so that the answer changes.

Larger image sets and video. On our MuirBench prompts with N≥5N \geq5 (Supp. Prompts with many images), SIEVE leads on all four backbones at B=116B = \frac{1}{16}, by 3.84–8.58 points. For video (Video-MME [8] and EgoSchema [17]), 16 uniformly sampled frames as images, SIEVE averages 59.76 and 61.21 at B=116B = \frac{1}{16} and 18\frac{1}{8}, above the strongest evaluated baseline, ToDRE, by 0.79 and 0.51 points (Table 5).

Single-image results

For a single image, the per-image quota reduces to global top-kk and the remaining operations are unchanged. Across four backbones and four benchmarks, SIEVE averages 62.96 and 69.59 at B=116B = \frac{1}{16} and 18\frac{1}{8} (77% and 85% of the uncompressed), and is the best method on all four backbones at B=116B = \frac{1}{16} and on three at B=18B = \frac{1}{8} (Table 6), so the pipeline remains competitive without cross-image allocation.

Efficiency

Table 3 reports realized cost on LLaVA-OneVision at B=18B = \frac{1}{8} (five-benchmark average on one H100). SIEVE cuts prefill computation from 26.46 to 4.02 TFLOPs (6.6×) and KV-cache storage from 107.4 to 16.7 MB (6.4×), the lowest of all methods, while exceeding the uncompressed mean by 2.01 points. At the same budget, all methods stay within 3% of these FLOPs and storage, yet SIEVE is 9.03 points more accurate than the best baseline. Prefill latency falls from 100.4 to 53.5 ms (1.9×), within 6.9 ms of the fastest method, which is 20.91 points less accurate.

Ablations

Table 7 replaces one component of SIEVE at a time on LLaVA-OneVision at B=116B = \frac{1}{16}, and groups the variants by the three properties in § 1 including cross-image coverage, image identity, and evidence throughout processing. Δ\Delta is the change from SIEVE (62.19), each group header states SIEVE’s setting, and all variants use the same depth-averaged budget.

Cross-image coverage

Budget allocation. A global ranking can leave an image with too few tokens. Replacing the equal per-image quota with one global top-kk costs 2.08 points, and a per-image minimum with the remainder allocated by attention costs 1.71, so the equal split itself matters, not only the guarantee that no image is starved.

Selection score. Quotas decide how many tokens each image keeps, and prompt attention decides which. Randomizing only the first stage costs 1.64, versus 0.66 for the later stages only, since later stages choose only among the tokens the first stage keeps.

Image identity

Merge target. Merging alone can fuse tokens from different images, so SIEVE merges each unselected token into the most similar kept token of the same image. Allowing merges across images costs 1.09 points, dropping unselected tokens costs 1.45, and spatial-neighbor, attention-based, and random assignment cost 1.46, 1.64, and 2.37.

MantisNLVR2MuirBLINKQ-Bench2AvgΔ\Delta
SIEVE63.5983.6147.7745.5070.5062.19
(1) Cross-image coverage
Budget allocation (SIEVE: equal quota per image)
one global top-kk, no per-image quota59.9182.1747.9640.9369.6060.11-2.08
minimum per image, remainder by attention59.4582.9946.8544.2468.9060.48-1.71
Selection score (SIEVE: prompt attention)
random score at all stages60.3783.6749.2740.9367.4060.33-1.86
random score at the first stage only61.2983.7148.8141.5667.4060.55-1.64
random score at later stages only63.5983.6448.9241.2970.2061.53-0.66
(2) Image identity
Merge target (SIEVE: most similar kept token of the same image)
drop unselected tokens, no merging60.8383.2848.3142.0869.2060.74-1.45
merge into spatially nearest kept token63.1382.6848.0841.0868.7060.73-1.46
most similar kept token from any image63.1383.6248.0840.9869.7061.10-1.09
assign by attention weight62.2182.6048.0841.1468.7060.55-1.64
merge into a random kept token60.3782.4548.0440.1468.1059.82-2.37
(3) Evidence throughout processing
Initial pooling (SIEVE: area-mean pooling before layer 0)
no pooling (first selection moves to layer 0)45.6262.0133.2741.3550.3046.51-15.68
bilinear resampling instead of area-mean61.7582.2446.3543.6670.9060.98-1.21
strided subsampling instead of area-mean57.1474.0343.6542.5055.7054.60-7.59
Pooling headroom (SIEVE: α=3\alpha= 3)
less headroom, α=1.5\alpha= 1.559.9182.7648.6539.8764.0059.04-3.15
Position indices (SIEVE: raster-order indices, a kept token keeps its index)
average the positions of merged tokens61.2983.5448.0841.0370.2060.83-1.36
keep original pre-pooling indices59.4582.0643.8543.3567.5059.24-2.95
Reduction after pooling (SIEVE: progressive select-and-merge)
pooling only, no later reduction61.2980.8747.1240.4562.7058.49-3.70
pooling + FastV’s pruning57.6078.2142.6243.9268.1058.09-4.10
pooling + PyramidDrop’s schedule58.0676.6541.8144.4071.2058.42-3.77
First stage (SIEVE: after layer ℓ1=2\ell_1 = 2)
first stage at ℓ1=0\ell_1 = 062.2182.5248.6242.6668.6060.92-1.27
first stage at ℓ1=4\ell_1 = 459.9181.3747.6539.8269.5059.65-2.54
first stage at ℓ1=8\ell_1 = 858.0677.1645.1940.8770.8058.42-3.77
Later-stage placement (SIEVE: second half of the remaining depth)
later stages evenly spaced over depth60.3784.5845.9641.7170.0060.52-1.67
later stages in the first half of the remaining depth58.5380.0240.3840.4570.6058.00-4.19
Number of stages (SIEVE: S(N)=clip⁡(N,2,4)S(N) = \operatorname{clip}(N, 2, 4))
fixed S=2S = 263.1383.6947.5840.6670.0061.01-1.18
fixed S=3S = 364.0684.6847.8841.5670.4061.72-0.47
Retention decay (SIEVE: γ=12\gamma= \frac{1}{2})
faster decay, γ=14\gamma= \frac{1}{4}62.2184.3149.2741.8270.7061.66-0.53
slower decay, γ=34\gamma= \frac{3}{4}63.1383.3147.1940.9869.2060.76-1.43
Budget schedule (SIEVE: geometric decay)
same retention ratio at every stage63.1383.7248.0041.2469.7061.16-1.03

Table 7. Ablation study for each property. Each row ablates one component of SIEVE. Δ\Delta denotes the change relative to SIEVE.

Keep 6.25%Keep 12.5%
OneVisionQwen2.5OneVisionQwen2.5
SIEVE62.1963.6765.1366.21
Recursive, ℓ=1\ell= 158.0560.7563.7463.76
Recursive, ℓ=2\ell= 257.8960.5964.0363.09
Recursive, ℓ=3\ell= 355.8158.0862.0161.90
No pooling, ℓ1=0\ell_1 = 046.5156.1451.7563.77
No pooling, ℓ1=2\ell_1 = 2——46.6455.67

Table 8. Reduction schedule and pooling. Mean accuracy over the five multi-image benchmarks. Recursive repeats reduction at a single scoring layer ℓ\ell, its scoring passes counted toward the budget. No pooling removes only the initial pooling and selects first at layer ℓ1\ell_1; —: infeasible (Eq. (2)).

Evidence throughout processing

Initial pooling. Keeping all tokens through the early layers spends the budget before any selection (Eq. (2)). Without pooling, the mean falls to 46.51 (−15.68); pooling with the same ℓ1=0\ell_1 = 0 gives 60.92, so the drop comes from removing pooling. Bilinear resampling or strided subsampling instead

of area-mean pooling costs 1.21 and 7.59 points.

Pooling headroom. Reducing the headroom from α=3\alpha= 3 to 1.5 costs 3.15 points, because fewer pooled tokens remain as candidates for prompt-guided selection.

Position indices. SIEVE numbers pooled tokens in raster order, and a kept token keeps its index while merging changes only its content. Averaging the positions of merged tokens costs 1.36 points, and keeping the original pre-pooling indices instead of renumbering costs 2.95.

Reduction after pooling. Pooling alone trails SIEVE by 3.70 points. Applying FastV’s pruning or PyramidDrop’s schedule to the pooled sequence gives 58.09 and 58.42, no better than pooling alone, so the gain comes from progressive select-and-merge, not from later reduction in general.

First stage. The first stage runs after layer ℓ1\ell_1, and every extra layer at full pooled width uses more of the budget (§ 3.3). Delaying it from layer 2 to 4 or 8 costs 2.54 and 3.77 points, and starting it at layer 0 costs 1.27.

Later-stage placement. Attention to the queried object peaks in the second half of the decoder (Figure 1(a)), where SIEVE places its later stages. Moving them into the first half costs 4.19 points, and spacing them evenly over depth costs 1.67.

Number of stages. SIEVE sets S(N)=clip⁡(N,2,4)S(N) = \operatorname{clip}(N, 2, 4) from the image count. Fixing S=2S = 2 or 3 lowers the mean by 1.18 and 0.47 points.

Retention decay. SIEVE keeps tokens with a geometric schedule, γ=12\gamma= \frac{1}{2}. A faster decay (γ=14\gamma= \frac{1}{4}) lowers the mean by 0.53 points and a slower one (γ=34\gamma= \frac{3}{4}) by 1.43.

Budget schedule. A constant retention ratio at every stage costs 1.03 points relative to the geometric schedule.

Progressive versus recursive reduction. Under matched budgets including repeated scoring, SIEVE outperforms the best recursive schedule by 1.10–4.14 points on LLaVA-OneVision and 2.45–2.92 on Qwen2.5-VL (Table 8).

Pooling across budgets. On LLaVA-OneVision, removing pooling with layer-0 selection costs 15.68 and 13.38 points at 6.25% and 12.5%; layer-2 selection also underperforms (46.64 on LLaVA-OneVision and 55.67 on Qwen2.5-VL at 12.5%) and is infeasible at 6.25% (Table 8).

Conclusion

We present SIEVE, a training-free, single-pass framework for multi-image token compression: each image is pooled before decoding, then tokens are selected and merged at chosen decoder layers under equal per-image quotas and a depth-averaged budget. Across four VLM backbones, it achieves the highest mean scores on five multi-image benchmarks, four single-image benchmarks, and two video benchmarks at a 6.25% depth-averaged token budget. These results highlight the importance of jointly optimizing compression timing, prompt-guided selection under equal per-image quotas, and same-image token merging.

Motivation

Global attention ranking

We examine whether global attention ranking across images reliably guides early token reduction. Using 500 four-image sets from the COCO joint-evidence probe (relevant images in slots 1 and 3), we evaluate up to seven questions per set: three probe questions and four image-specific queries of the form “Is there a ⟨object⟩ in image pp?”, totaling 3,485 questions per backbone (an image without an annotated object gets no image-specific query). Without pruning, we rank all pooled visual tokens using attention from the final prompt token at five relative decoder depths and the final layer, retaining the global top-10%. Table S1 reports pairwise Jaccard overlap across questions (random: 5%) and the fraction assigned to the queried image (uniform: 25%). Bootstrap 95% confidence intervals (10,000 image-set resamples) remain within ±0.7 points. In the first third of the decoder, rankings are largely query-agnostic, with 81–86% overlap and near-uniform allocation to the queried image. Only in the second half does the queried image receive most of the selected tokens, with per-backbone peaks of 51–72%. At the final layer, LLaVA-OneVision and InternVL3 retain this preference (63% and 45%), whereas Qwen3-VL reverts toward uniform allocation (26%, 73% overlap).

What prompt attention tells apart, by depth

Figure S1. What prompt attention tells apart, by depth. (i) Attention to the queried object in one image (solid) and to a same-area centred box (dotted). (ii) Global top-10% over four images: share in the asked-about image (solid) and overlap between questions (dashed).

Method details

Algorithm

Algorithm 1 summarizes the SIEVE prefill procedure.

Backbone-specific pooling

Image ii receives mi=round⁡(mPi/n)m_i = \operatorname{round}(mP_i/n) tokens from the pooling target (§ 3.2). Coordinates ct∈R2c_t \in\mathbb{R}^{2} denote rows

Backbonerelative depth
0.070.300.520.700.851.00
LLaVA-OneVision-7BOverlap82.581.044.825.023.025.2
LLaVA-OneVision-7BAsked24.925.444.369.970.062.7
Qwen3-VL-8BOverlap85.385.781.022.753.173.2
Qwen3-VL-8BAsked25.125.028.671.938.826.1
InternVL3-8BOverlap84.482.667.031.735.735.7
InternVL3-8BAsked24.925.229.851.548.845.0

Table S1. Global top-10% by relative decoder depth. Overlap: mean Jaccard between questions on the same images (random: 5%). Asked: share inside the image the question asks about (uniform: 25%). 500 image sets per backbone; depths 2, 8, 14, 19, 23 of 28 layers (LLaVA-OneVision, InternVL3) and 2, 10, 18, 24, 30 of 36 (Qwen3-VL), and the last layer.

and columns on the unpooled grid; pooled coordinates c~\tilde{c} denote cell centers in the same frame.

LLaVA-OneVision, multi-image. Each image has a 27×2727 \times27 grid (G=27G = 27), with mi=round⁡(m/N)m_i = \operatorname{round}(m/N). Adaptive average pooling gives X~i=AvgPool⁡G~×G~(Xi)\tilde{X}_i = \operatorname{AvgPool}_{\tilde{G} \times\tilde{G}}(X_i), where G~=⌈mi⌉\tilde{G} = \lceil\sqrt{m_i}\rceil, where ⌈⋅⌉\lceil\cdot\rceil rounds to the nearest integer. Cell (a,b)(a,b) has center c~=((a+12)G/G~−12,(b+12)G/G~−12)\tilde{c} = ((a+\frac{1}{2})G/\tilde{G}-\frac{1}{2}, (b+\frac{1}{2})G/\tilde{G}-\frac{1}{2}). Grid rounding yields m~=∑iG~2\tilde{m} = \sum_i \tilde{G}^{2}, which may differ from mm (Sec. B.4).

LLaVA-OneVision, single image. AnyRes tokens are binned by coordinate into the smallest G~×G~\tilde{G} \times\tilde{G} grid with G~≥⌈mi⌉\tilde{G} \ge\lceil\sqrt{m_i}\rceil and at least mim_i occupied cells. Features are averaged within each occupied cell; row separators are retained.

Qwen2.5-VL. The Gh×GwG_h \times G_w grid is pooled to ⌈miGh/Gw⌉×⌈miGw/Gh⌉\lceil\sqrt{m_iG_h/G_w}\rceil\times\lceil\sqrt{m_iG_w/G_h}\rceil cells, preserving its aspect ratio up to rounding and using the cell centers defined above.

InternVL3. Each image has up to max⁡(1,min⁡(6,⌈64/N⌉))\max(1,\min(6,\lceil64/N\rceil)) crops of 4482448^{2} pixels and a thumbnail. The tile mosaic and 16×1616 \times16 thumbnail are pooled separately, in proportion to their token counts, then concatenated in that order.

Position indices after reduction

Pooled tokens follow raster order, with text, row separators, and image markers preserving their original order (§ 3.3). Cell centers are stored separately from position indices for spatial bookkeeping, including AnyRes binning, per-image quotas, neighbor search, and merge-target selection.

LLaVA-OneVision and InternVL3. Both use one-dimensional rotary embeddings with consecutive positions

Algorithm 1 SIEVE: multi-image token compression
Require: Visual grids $\{X_i\}_{i=1}^{N}$, prompt p, decoder $\{f_\ell\}_{\ell=0}^{L-1}$, target budget $B$
Ensure: Compressed prefill states and layer-wise KV caches
1: Set $\alpha = 3$, $\beta = 0.75$, $\gamma = \frac{1}{2}$, $\ell_1 = 2$, $q_{\min} = 4$
1. Spatial pooling before the decoder
2: $P_i \leftarrow |X_i|$; $n \leftarrow \sum_i P_i$
3: $m \leftarrow \lfloor \min(\alpha B,\beta)n\rfloor$
4: for $i=1,\ldots,N$ do
5: $m_i \leftarrow \operatorname{round}(mP_i/n)$
6: Resample $X_i$ to approximately $m_i$ tokens using the backbone-specific operator
7: end for
8: Interleave pooled grids with p to form $H$; initialize owner map $\omega$ and position indices
2. Image-adaptive, budget-aware scheduling
9: $S \leftarrow \min(4,\max(2,N))$
10: Compute $\ell_2,\ldots,\ell_S$ using {xref}`Eq. (3)`
11: Set $\ell_{S+1} \leftarrow L-1$ and $\delta_j \leftarrow \ell_{j+1}-\ell_j$
12: $c \leftarrow \frac{BnL-m(\ell_1+1)}{\sum_{j=1}^{S}\gamma^{j-1}\delta_j}$
13: $k_j \leftarrow \operatorname{clip}(\operatorname{round}(c\gamma^{j-1}),1,m)$ for each stage $j$
14: for $\ell=0,\ldots,L-1$ do
15: Run $f_\ell$ on $H$, retaining its outputs and KV cache
16: if $\ell=\ell_j$ for some stage $j$ then
3. Prompt-attention scoring
17: Compute $s_t$ from the final prompt token’s attention in $f_\ell$, normalized over visual tokens per head ({xref}`Eq. (4)`)
4. Per-image coverage
18: Set $V_i \leftarrow \{t:\omega(t)=i\}$
19: Allocate $k_j$ slots uniformly across images with floor $q_{\min}$ and capacity $|V_i|$; distribute the remainder ({xref}`Sec. 3.3`)
20: $K_i \leftarrow \operatorname{TopK}(\{s_t:t\in V_i\},q_i)$; $K\leftarrow \bigcup_i K_i$
5. Within-image evidence merging
21: Using layer outputs, assign each unselected $t$ to $\pi(t)\leftarrow\arg\max_{u\in K_{\omega(t)}}\cos(h_t,h_u)$
22: Update all representatives simultaneously by score-weighted averaging ({xref}`Eq. (5)`)
6. Sequence and position update
23: Keep all text tokens and merged representatives in sequence order; retain their owner and position indices
24: Pass the reduced sequence to the next layer; retain the caches already computed
25: else
26: Pass the layer outputs to the next layer
27: end if
28: end for
29: return Final prefill states and layer-wise KV caches

0, …, T−1T-1 on the pooled sequence. InternVL3 thumbnail coordinates are rescaled to the mosaic frame for spatial bookkeeping only.

Qwen2.5-VL and Qwen3-VL. After pooling, we reconstruct the three M-RoPE components: text runs use arange⁡(n)\operatorname{arange}(n) in each component, and image blocks use (0,r,c)(0,r,c), with the running offset advancing by max⁡(r,c)+1\max(r,c)+1. Here, rr and cc are dense ranks of the distinct row and column coordinates within each image. Rank compression supports non-rectangular token sets, whereas the official index builders assume a dense Gh×GwG_h \times G_w grid.

Decoder stages. Surviving tokens retain their indices, including gaps from removed tokens (e.g., 1, 2, 3, 4 → 2, 4). Each M-RoPE stage indexes the table constructed before the first layer. Merging tt into uu preserves uu’s index and coordinates.

Merged position ablation. In the row average the positions of merged tokens of Table 7, each surviving token’s position index is the average of the indices it absorbed, weighted by the content weights in Eq. (5). Fractional rotary positions are maintained separately from integer cache indices.

Stage count and schedule shape

Worked example. For LLaVA-OneVision with N=4N=4, n=2916n=2916, and B=18B=\frac{1}{8}, Eq. (6) gives

stagetokensper image
pooling (before f0f_0)2916 → 1156289
after f2f_21156 → 37293
after f18f_{18}372 → 18646–47
after f21f_{21}186 → 9323–24
after f24f_{24}93 → 4611–12

Table 7 (PDF p. 12).

The depth-averaged count is 371.25, compared with the target Bn=364.5Bn=364.5. The pooling target m=1093m=1093 gives mi=273m_i=273, rounded to a 17×17 grid per image (m~=1156\tilde m=1156). The 63 additional tokens across three prefix layers contribute 6.75 depth-averaged tokens, or 1.9% of the budget. At B=116B=\frac{1}{16}, m=546m=546, mi=136m_i=136, and G~=12\tilde G=12 yield m~=576\tilde m=576, adding 1.8% before subsequent retained counts are rounded.

Stage count. At a fixed budget, more stages increase the initial retained count and decrease the final count. For the same prompt at B=116B=\frac{1}{16} (m=546m=546), with at most four stages (Eq. (7)):

stagesretained countsafter last stage
S=2S=2158, 7979
S=3S=3173, 87, 4343
S(N)S(N) (SIEVE)186, 93, 47, 2323

Table 8 (PDF p. 12).

Table 7 compares fixed stage counts. At γ=1\gamma=1, later stages remove no tokens, yielding the same predictions as a single-stage configuration.

Schedule variants. Table 7 varies decay γ\gamma, pooling headroom α\alpha, stage placement, and schedule shape individually. Alternative placements distribute stages uniformly after $\ell_1 or within the first half of the remaining depth. The constant-ratio variant uses nj=mrjn_j = mr^j, as in PyramidDrop, instead of decay from the solved k1k_1. The cap β\beta is fixed because it is active only when αB>0.75\alpha B > 0.75.

Evaluation details

Backbones and baselines

All backbones share the hyperparameters in § 4.1; decoder depth LL determines the schedule and stage placement. On Qwen3-VL-8B (L=36L = 36), two, three, and four stages follow layers {2,27}\{2, 27\}, {2,24,30}\{2, 24, 30\}, and {2,23,27,31}\{2, 23, 27, 31\}, respectively. Following their published depth-based rules, PyramidDrop and SparseVLM use {2,9,18,27}\{2, 9, 18, 27\} and {2,7,17}\{2, 7, 17\} on Qwen3-VL, respectively; their InternVL3 placements match those on LLaVA-OneVision and Qwen2.5-VL. Our VisionZip adaptation ranks tokens by feature norm on InternVL3 and Qwen2.5-VL, whose visual representations lack a class token.

Budgets a method cannot reach

A dash in Table 2 to Table 6 denotes an infeasible budget; SparseVLM is omitted from all 6.25% rows. The full-width prefix consumes (ℓ1+1)/(BL)(\ell_1 + 1)/(BL) of the target budget, exceeding it when BL<ℓ1+1BL < \ell_1 + 1 (2). Thus, SparseVLM (ℓ1=2\ell_1 = 2) cannot meet 6.25% on any backbone. CLSE (ℓ1=1\ell_1 = 1) requires at least 7.14% on the three 28-layer backbones, but is feasible on Qwen3-VL (5.56%). Seven Qwen2.5-VL prompts are infeasible for S²Prune at 6.25%; excluding them changes each of the two affected cells by less than 0.1 points.

Benchmarks and table conventions

Selected splits are evaluated in full, subject to the infeasible-prompt exclusions in Sec. C.2 and Table S9. Q-Bench2 uses the publicly labeled development split. Full denotes the uncompressed reference; bold marks the highest average in Table 2.

Multi-imageSingle-image
MantisNLVR2MuirBLINKQ-Bench2AvgTextVQASQA-IMMStarPOPEAvg
LLaVA-OneVision-7B, keep 6.25% of visual tokens
SIEVE (progressive)63.5983.6147.7745.5070.5062.1957.9084.5349.7386.4669.66
recursive, ℓ=1\ell= 161.2979.0946.1540.5063.2058.0552.3283.4948.3384.0667.05
recursive, ℓ=2\ell= 258.0678.7746.4239.7266.5057.8956.5582.5546.6784.8367.65
recursive, ℓ=3\ell= 354.8475.5343.9240.3564.4055.8153.5281.7145.2784.0966.15
LLaVA-OneVision-7B, keep 12.5% of visual tokens
SIEVE (progressive)65.9088.8349.3148.1973.4065.1362.9586.7151.2087.0271.97
recursive, ℓ=1\ell= 165.4487.0047.5846.4072.3063.7455.7085.8249.0786.1669.19
recursive, ℓ=2\ell= 266.8286.8248.2746.1372.1064.0361.3686.2749.4086.0170.76
recursive, ℓ=3\ell= 361.7583.9147.6946.1970.5062.0159.3384.9347.2785.0869.15
Qwen2.5-VL-7B, keep 6.25% of visual tokens
SIEVE (progressive)71.4378.0550.9649.5068.4063.6731.6780.0740.1380.8858.19
recursive, ℓ=1\ell= 171.4376.6043.9647.3464.4060.7531.6280.1838.6381.0457.87
recursive, ℓ=2\ell= 270.9775.2443.8147.7165.2060.5930.1078.8338.2079.2356.59
recursive, ℓ=3\ell= 365.4472.0141.9646.7964.2058.0828.0478.5237.3477.5755.37
Qwen2.5-VL-7B, keep 12.5% of visual tokens
SIEVE (progressive)72.8182.1652.3552.0371.7066.2149.0381.8545.6084.3465.21
recursive, ℓ=1\ell= 171.8980.0949.6950.4566.7063.7649.3581.0144.4183.6864.61
recursive, ℓ=2\ell= 270.0578.6349.8150.4566.5063.0947.5880.9144.0582.6963.81
recursive, ℓ=3\ell= 369.1276.5047.6950.2965.9061.9045.7380.1241.6681.5462.26

Table S9. Progressive versus recursive reduction, per benchmark. The recursive variant repeatedly reduces the sequence at scoring layer ℓ\ell; its scoring passes count toward the budget. The highest average in each block is bold. On Qwen2.5-VL, infeasible recursive configurations exclude up to 70 of 1,500 MMStar, 15 of 2,017 SQA-I, and 1 of 1,901 BLINK prompts per cell.

Efficiency measurement and compute estimate

Latency is measured on a single NVIDIA H100 GPU without concurrent workloads (Table 3). At B=18B = \frac{1}{8}, methods realize similar budgets; AVTP, VisionZip, and FastV are faster (46.6–47.9 ms vs. 53.5 ms) because SIEVE performs scoring and merging at S(N)S(N) stages. At B=116B = \frac{1}{16}, FastV and PyramidDrop exceed the budget despite retaining only one token after reduction. On NLVR2, their three-layer full-width prefix alone contributes 3n/L=156.23n/L = 156.2 depth-averaged tokens against a target of 91.1; realized retention is 157.1. Across five benchmarks, their computation exceeds that of SIEVE by 46% at B=116B = \frac{1}{16} (Figure S2). Visual-token prefill computation (Figure 2) is estimated as L(clinnˉ+cattnnˉ2/2)L(c_{\mathrm{lin}}\bar{n} + c_{\mathrm{attn}}\bar{n}^2/2), where clinc_{\mathrm{lin}} and cattnc_{\mathrm{attn}} are the 7B decoder’s per-token projection/MLP and per-pair attention costs. Text-token costs are excluded. Measured nˉ\bar{n} (1) gives the exact linear term; the approximate quadratic term accounts for less than 4% of the total. Figure S2 provides per-benchmark results, including the concurrent baseline AVTP [27].

Figure S2. Accuracy versus estimated prefill computation by benchmark on LLaVA-OneVision at B=116B = \frac{1}{16} and 18\frac{1}{8}. SIEVE leads at both evaluated budgets and exceeds the uncompressed model (dashed line) on Mantis-Eval and MuirBench. At B=116B = \frac{1}{16}, FastV and PyramidDrop exceed their targets despite retaining one token after pruning (2.56 TFLOPs on Mantis-Eval versus 1.51 for SIEVE). SparseVLM and CLSE are feasible only at B=18B = \frac{1}{8} (Sec. C.2).

Additional quantitative results

Single-image results

Table S2 reports the per-benchmark single-image results on all four backbones; Table 6 reports their means.

Prompts with many images

On the 1,096 MuirBench prompts with N≥5N \ge5 (Table S3), SIEVE ranks first on all four backbones at Bˉ=116\bar{B} = \frac{1}{16} and three at Bˉ=18\bar{B} = \frac{1}{8}. At Bˉ=14\bar{B} = \frac{1}{4}, it leads on LLaVA-OneVision and InternVL3; S²Prune and AVTP lead on Qwen2.5-VL and Qwen3-VL, respectively. It exceeds uncompressed LLaVA-OneVision at every evaluated budget.

Table S3. Prompts with at least five images. Accuracy on the 1096 MuirBench prompts with N≥5N \ge5, computed from the same predictions as Table 2. SIEVE attains the highest scores for all backbones at 6.25% token budget.

MethodLLaVA-OneVisionQwen2.5-VLQwen3-VLInternVL3
Full47.9063.1452.6556.11
Keep 6.25% of visual tokens
VisionZip [23]40.3347.4547.5447.35
FastV [7]26.0029.4746.4418.98
PyramidDrop [22]26.0029.4746.4418.98
AVTP [27]28.3843.7047.5433.12
CLSE [5]——48.08—
S²Prune [10]37.2348.7247.2640.15
SIEVE48.9154.7451.9251.92
Keep 12.5% of visual tokens
VisionZip [23]45.7153.2848.4548.45
FastV [7]32.9437.2350.2733.12
PyramidDrop [22]38.0536.8653.2829.20
SparseVLM [28]34.3141.0649.7335.58
AVTP [27]32.7648.8149.3643.16
CLSE [5]37.1449.8248.5440.51
S²Prune [10]41.2455.3851.0947.17
SIEVE52.0157.2151.9254.56
Keep 25% of visual tokens
VisionZip [23]46.6256.9350.1851.73
FastV [7]37.9651.4650.6450.46
PyramidDrop [22]40.3351.8251.8248.72
SparseVLM [28]37.2351.1950.9150.09
AVTP [27]41.8856.0252.7449.64
CLSE [5]45.4457.5750.7351.09
S²Prune [10]46.7260.5851.8252.19
SIEVE50.0060.2251.0056.66

Table S3.

Video results by benchmark

Table S4 separates the two benchmarks averaged in Table 5, Video-MME (V) and EgoSchema (E), for each backbone. SIEVE has the highest average at both budgets. At both budgets it leads on Video-MME for LLaVA-OneVision and on EgoSchema for Qwen2.5-VL and Qwen3-VL, while ToDRE leads on Video-MME for the other three backbones.

Table S4. Performance comparison for video benchmarks. Each clip uses 16 uniformly sampled frames as images, with the same processing as Table 2. V: Video-MME, all 2,700 questions; E: EgoSchema’s 500-question labeled subset (0.2 points per question; the remaining 4,531 questions require leaderboard evaluation). Avg is the average of Table 5. The best score in each column is bold and the second best _underlined_; a dash marks a method that cannot meet the budget.

MethodLLaVA-OneVision-7B VLLaVA-OneVision-7B EQwen2.5-VL-7B VQwen2.5-VL-7B EQwen3-VL-8B VQwen3-VL-8B EInternVL3-8B VInternVL3-8B EAvg
Full53.6363.8056.7868.4060.9371.8059.8169.8063.12
Keep 6.25% of visual tokens
VisionZip [23]50.6762.8049.9359.8053.7861.2052.4858.4056.13
FastV [7]39.0736.2040.3330.4041.7437.6041.5236.0037.86
PyramidDrop [22]39.0736.2040.3330.4041.8937.6041.5236.0037.88
AVTP [27]40.1937.8048.6348.4043.3350.4048.5643.6045.12
CLSE [5]––––42.6338.00–––
S²Prune [10]45.8547.6050.5248.0049.8557.6050.3348.8049.82
ToDRE [13]51.2263.6054.2261.4056.4462.2058.5264.2058.97
SIEVE54.1161.8053.4463.6055.8168.6056.9363.8059.76
Keep 12.5% of visual tokens
VisionZip [23]52.3765.4051.1562.8056.6766.2055.6365.0059.41
FastV [7]44.2649.4046.5648.4051.4855.2050.0051.2049.56
PyramidDrop [22]42.4844.8045.6345.2049.5248.4048.9640.0045.63
SparseVLM [28]44.2653.4051.1959.2056.2264.8053.5659.8055.31
AVTP [27]43.7446.8051.3757.0049.5658.6051.3053.8051.52
CLSE [5]47.8958.8051.7861.0053.1562.0055.5964.4056.83
S²Prune [10]51.3061.2053.5658.6056.1164.2055.1560.6057.59
ToDRE [13]52.7464.2056.0065.0058.2664.0059.0066.4060.70
SIEVE53.8562.0053.8967.4057.9671.0057.9365.6061.21

Table S4.

Ablations on additional backbones

Table S5 to Table S7 extend the ablations to the other three backbones. All variants underperform SIEVE, with removing pooling causing the largest drops (−7.53, −5.02, and −6.00 points). Strided pooling causes smaller drops (−0.97, −0.48, and −0.72) than on LLaVA-OneVision (−7.59; Table 7), likely because Qwen and InternVL already combine neighboring patches, preserving local information that striding discards in LLaVA-OneVision. Size-proportional quotas allocate tokens based on each image’s remaining token count.

Table S5. Ablation study for each property on Qwen2.5-VL-7B. Each row ablates one component of SIEVE. △\triangle denotes the change relative to SIEVE.

MantisNLVR2MuirBLINKQ-Bench2Avg△\triangle
SIEVE71.4378.0550.9649.5068.4063.67
(1) Cross-image coverage
Budget allocation (SIEVE: equal quota per image)
one global top-kk, no per-image quota67.2877.8445.6548.9265.6061.06-2.61
size-proportional share per image70.0577.3145.7749.3465.4061.57-2.10
Selection score (SIEVE: prompt attention)
random score at all stages67.7477.3945.1547.4565.8060.71-2.96
random score at the first stage only67.7477.4145.1547.4065.7060.68-2.99
random score at later stages only70.0577.7545.7749.5165.1061.64-2.03
(2) Image identity
Image target (SIEVE: most similar kept token of the same image)
drop unselected tokens, no merging70.9777.5745.5048.5566.2061.76-1.91
merge into spatially nearest kept token59.5977.7245.3848.7664.3061.15-2.52
most similar kept token from any image69.5977.8845.8849.4065.3061.61-2.06
assign by attention weight71.4377.2945.2348.7665.6061.66-2.01
merge into a random kept token65.4476.8044.5448.9764.2059.99-3.68
(3) Evidence throughout processing
Initial pooling
no pooling60.3766.2842.0046.2465.8056.14-7.53
strided pooling72.3575.1847.5849.2969.1062.70-0.97
Reduction after pooling (SIEVE: progressive select-and-merge)
pooling only, no later reduction68.2076.4242.8147.1364.3059.77-3.90
pooling + FastV’s pruning65.4473.2743.2348.1964.7058.97-4.70
pooling + PyramidDrop’s schedule65.4471.8142.3147.9766.3058.77-4.90
First stage (SIEVE: after layer ℓ1=2\ell_1 = 2)
first stage at ℓ1=0\ell_1 = 071.4378.9446.0848.5065.9062.17-1.50
first stage at ℓ1=4\ell_1 = 469.5976.8545.0848.8265.6061.19-2.48
first stage at ℓ1=8\ell_1 = 864.0671.9042.9646.8265.8058.31-5.36
Later-stage placement (SIEVE: second half of the remaining depth)
later stages evenly spaced over depth71.4377.8744.3549.0366.3061.80-1.87
later stages in the first half of depth66.3674.2240.9246.9865.6058.82-4.85
Number of stages (SIEVE: S(N)=clip⁡(N,2,4)S(N) = \operatorname{clip}(N, 2, 4))
fixed S=2S = 270.0577.7245.5849.0865.2061.53-2.14
fixed S=3S = 370.9778.2745.9249.1865.9062.05-1.62
Budget schedule (SIEVE: geometric decay)
same retention ratio at every stage70.5177.7746.0449.3465.4061.81-1.86

Table S5.

Table S7. Ablation study for each property on InternVL3-8B. Each row ablates one component of SIEVE. Δ\Delta denotes the change relative to SIEVE.

MantisNLVR2MuirBLINKQ-Bench2AvgΔ\Delta
SIEVE67.7478.3050.9249.5071.4063.57
(1) Cross-image coverage
Budget allocation (SIEVE: equal quota per image)
one global top-kk, no per-image quota64.5275.6947.1947.3465.4060.03-3.54
size-proportional share per image59.9075.2146.6547.2966.1060.23-3.34
Selection score (SIEVE: prompt attention)
random score at all stages59.9173.2645.8847.2463.0057.86-5.71
random score at the first stage only59.9173.2745.7347.4563.0057.87-5.70
random score at later stages only66.3675.9047.1247.3464.6060.26-3.31
(2) Image identity
Merge target (SIEVE: most similar kept token of the same image)
drop unselected tokens, no merging65.9076.0347.0847.3465.3060.33-3.24
merge into spatially nearest kept token66.3674.7846.4246.8663.9060.02-3.55
most similar kept token from any image65.9075.9046.7347.2464.7060.09-3.48
assign by attention weight64.9875.9246.5048.3464.6060.07-3.50
merge into a random kept token64.5274.9446.0846.9864.4059.38-4.19
(3) Evidence throughout processing
Initial pooling
no pooling60.3769.7942.2745.8269.6057.57-6.00
strided pooling68.2075.5150.3847.6672.5062.85-0.72
Position indices (SIEVE: raster-order indices, a kept token keeps its index)
keep original pre-pooling indices67.7474.6241.9647.7662.3058.88-4.69
Reduction after pooling (SIEVE: progressive select-and-merge)
pooling only, no later reduction61.7572.6344.3847.9264.1058.16-5.41
pooling + FastV’s pruning64.9876.1945.4648.7664.7060.02-3.55
pooling + PyramidDrop’s schedule67.2874.4945.1950.0867.0060.81-2.76
First stage (SIEVE: after layer ℓ1=2\ell_1 = 2)
first stage at ℓ1=0\ell_1 = 064.0676.0447.0047.9264.0059.80-3.77
first stage at ℓ1=4\ell_1 = 465.4474.9745.9647.4064.8059.71-3.86
first stage at ℓ1=8\ell_1 = 865.9072.7944.1246.9263.2058.59-4.98
Later-stage placement (SIEVE: second half of the remaining depth)
later stages evenly spaced over depth66.3676.1044.2747.0865.5059.86-3.71
later stages in the first half of depth66.8274.5941.2346.9864.1058.74-4.83
Number of stages (SIEVE: S(N)=clip⁡(N,2,4)S(N) = \operatorname{clip}(N, 2, 4))
fixed S=2S = 264.9875.9446.3546.9264.7059.78-3.79
fixed S=3S = 366.3676.1746.9247.4065.5060.47-3.10
Budget schedule (SIEVE: geometric decay)
same retention ratio at every stage64.5275.9447.1547.0364.7059.87-3.70

Table S7.

Progressive versus recursive reduction

Table S9 reports the per-benchmark scores of the progressive-versus-recursive comparison in § 6.1.

Additional qualitative results

Figure S3 and Figure S4 show more examples using SIEVE.

Figure S3. Examples from MuirBench, NLVR2, and Mantis. Final-stage tokens cover the queried objects: a red towel, a bird, a bike, and hats and bags.

Figure S4. Additional examples involving a dog’s open mouth, a person in graduation attire, two formulae, and two shops.

Additional analyses

Compression timing

One-step pruning. On LLaVA-OneVision-7B, FastV reaches BLINK [9] accuracies of 39.24 and 40.08 at $B =

{#eq:}
MethodTextVQA OVTextVQA Qwen2.5TextVQA Qwen3TextVQA IVL3SQA-I OVSQA-I Qwen2.5SQA-I Qwen3SQA-I IVL3MMStar OVMMStar Qwen2.5MMStar Qwen3MMStar IVL3POPE OVPOPE Qwen2.5POPE Qwen3POPE IVL3Avg OVAvg Qwen2.5Avg Qwen3Avg IVL3
Full76.2184.5083.5381.5395.9388.3094.5598.3661.3361.7360.1367.2089.2287.8288.8290.6680.6780.5981.7684.44
Keep 6.25% of visual tokens
VisionZip [23]49.9321.9221.6524.0384.4879.9278.5886.1244.2740.0037.4747.2082.6380.1078.9886.8665.3355.4854.1761.05
FastV [7]9.159.039.109.9374.3773.0373.5773.9729.5327.9328.4730.4047.4645.4257.9646.1740.1338.8542.2840.12
PyramidDrop [22]9.159.039.109.9374.3773.0373.5773.9729.5327.9328.4730.4047.4645.4257.9646.1740.1338.8542.2840.12
CLSE [5]––9.66–––73.13–––29.07–––58.81–––42.67–
S²Prune [10]35.1442.0215.9835.7681.2675.5877.7980.8139.5334.9633.9341.1370.5867.0873.4879.0356.6354.9150.3059.18
SIEVE57.9031.6730.8234.1084.5380.0781.6686.6649.7340.1341.2749.0086.4680.8884.3987.9969.6658.1959.5364.44
Keep 12.5% of visual tokens
VisionZip [23]62.3633.4136.0533.4587.9081.5184.0989.5449.8045.1343.8752.3386.9785.0084.3389.1271.7661.2662.0866.11
FastV [7]19.9826.2221.5732.6577.0974.8175.4181.4636.0730.8732.6041.9365.3651.9458.9777.6249.6245.9647.1458.41
PyramidDrop [22]22.6725.6316.1117.8775.6174.7175.0676.1533.8730.4731.9334.9359.3853.1161.8068.5347.8845.9846.2349.37
SparseVLM [28]23.6342.8741.2743.1378.9878.7879.7786.1739.4735.2738.0047.8069.1969.1476.3084.7752.8256.5258.8464.47
CLSE [5]34.9467.7737.6653.2581.5180.6677.3492.2744.5344.2038.5354.2778.3875.6075.0687.1459.8467.0657.1571.73
S²Prune [10]53.7562.2029.9351.9887.8080.6181.3687.8550.0047.3842.2049.2083.8278.5081.9386.8868.8466.2758.8668.98
SIEVE62.9549.0346.8861.1586.7181.8583.6493.2151.2045.6046.2757.6087.0284.3486.0190.0071.9765.2165.7075.49

Table S2. Performance comparison for single-image benchmarks on LLaVA-OneVision-7B (OV), Qwen2.5-VL-7B (Qwen2.5), Qwen3-VL-8B (Qwen3) and InternVL3-8B (IVL3). Per-benchmark scores and their average; the best average for each backbone and budget is bold and the second best _underlined_. A dash marks a method that cannot meet the budget.

116\frac{1}{16} and 18\frac{1}{8}, respectively, near the 37.7 chance level (mean reciprocal of answer-option counts). Its full-width prefix leaves little or no budget after pruning (Eq. (2)).

Aggregation. After removing half the visual tokens at layer 2, an object token has cosine similarity 0.77 to its most similar retained representative, versus 0.44 to its nearest spatial neighbor. Score-weighted merging transfers discarded-token features to retained representatives (Eq. (5)).

Repeated merging. A single merge improves accuracy by 0.04 points, compared with 2.27 points when selection and merge assignments are recomputed across depths (Figure 1b). This supports progressive aggregation with updated representatives.

Quota effects by image count

Per-image quota gains over global top-kk vary with image count on LLaVA-OneVision-7B at B=116B = \frac{1}{16} (percentage points):

N=2N = 2N∈{3,4}N \in\{3,4\}N≥5N \ge5
Mantis-Eval+2.8+5.6n/a
MuirBench−1.6+0.5+0.7

Table S2.

Quotas improve all reported subsets with N≥3N \ge3, but have mixed effects at N=2N = 2. The aggregate in Table 7 obscures this variation; NLVR2 and Q-Bench2 contain only two-image questions.

Robustness to relevant-image position

Protocol. We sample 500 MMStar questions approximately uniformly across six categories, independently of model predictions. Each relevant image is combined with N−1N - 1 distinct MMStar distractors for N∈{4,8}N \in\{4, 8\}, using nested distractor sets and fixed relative order. We test every relevant-image position, updating only its question index; the image, options, and answer remain fixed across positions and methods. Evaluation uses LLaVA-OneVision-7B at B=18B = \frac{1}{8} (Table S8).

Table S8. Relevant-image position robustness on LLaVA-OneVision-7B at B=18B = \frac{1}{8}. Mean/Worst: average/minimum accuracy over relevant-image positions (%); Gap: best–worst difference (points).

MethodN=4N = 4N=8N = 8
MeanWorstGap↓\downarrowMeanWorstGap↓\downarrow
Full52.951.22.648.145.27.2
VisionZip [23]43.141.24.640.137.87.2
FastV [7]30.529.05.030.027.89.2
PyramidDrop [22]28.827.03.228.626.83.2
SparseVLM [28]30.428.83.229.827.47.0
AVTP [27]33.532.03.033.532.03.6
CLSE [5]38.333.013.834.830.018.0
S2^{2}Prune [10]44.043.02.640.438.25.2
ToDRE [13]42.040.44.639.737.66.0
SIEVE w/ global top-kk45.843.06.243.337.413.4
SIEVE w/ floor + attention45.141.49.043.236.614.6
SIEVE47.647.01.445.242.25.4

Table S8.

Metrics. Mean and Worst denote average and minimum accuracy across positions; Gap is the maximum minus minimum. A small gap alone does not imply high accuracy. We estimate 95% confidence intervals from 10,000 paired bootstrap resamples over questions, retaining all positions per sampled question.

Joint preservation of evidence from two images

Protocol. We construct 500 COCO val2017 image pairs [15]: 167 sum, 167 comparison, and 166 category-matching questions. Each joint question combines one fact per image, each also queried separately. Pairs are selected independently of predictions; comparison and matching answers are balanced between yes and no. Counting images contain one to four annotated instances of the queried category, excluding crowd labels, with instance area ≥0.4%\ge0.4\% of the image and pairwise box IoU ≤0.3\le0.3. Animal images contain exactly one animal category occupying ≥5%\ge5\% of the image.

Each pair is tested at all N(N−1)N(N - 1) ordered placements for N∈{2,4,8}N \in\{2, 4, 8\}. The N−2N - 2 distractors contain no annotated queried category (no animal for matching); their sets are nested across NN with fixed relative order. Joint and constituent questions use separate calls on the same sequence without shared history, changing only image indices across placements. Dependency controls replace either relevant image with a screened image that changes the annotated joint answer, keeping the question, placement, and distractors fixed. Replacements cover 394 pairs for A, 373 for B, and 315 for both.

Each method processes 158,690 prompts using LLaVA-OneVision-7B with greedy decoding. Compressed methods target B=18B = \frac{1}{8}, realizing 12.4–12.7% depth-averaged retention. Table S10 reports N=4N = 4 and N=8N = 8.

MethodN=4N = 4N=8N = 8
MethodJointBothAll 3WorstCF-ACF-BJointBothAll 3WorstCF-ACF-B
Full77.177.662.470.447.242.063.962.140.549.023.623.2
VisionZip [23]67.257.740.957.629.426.160.756.734.947.620.519.8
FastV [7]50.824.820.941.22.31.744.019.012.836.82.22.3
PyramidDrop [22]49.924.016.744.41.31.042.017.18.836.02.73.2
SparseVLM [28]51.929.824.445.22.12.243.721.913.637.44.24.9
AVTP [27]58.039.435.256.41.91.654.635.930.551.20.71.0
CLSE [5]51.736.627.137.27.86.343.930.314.232.86.26.5
S²Prune [10]64.352.438.958.223.222.454.144.523.942.613.612.8
ToRDRE [13]68.254.440.858.432.629.959.354.333.146.420.919.4
SIEVE w/ global top-kk85.083.673.181.060.251.974.674.957.864.442.037.1
SIEVE w/ floor + attention84.083.171.979.657.451.174.674.958.063.241.537.4
SIEVE86.683.674.781.862.955.077.877.862.769.647.342.5

Table S10. Joint evidence preservation on LLaVA-OneVision-7B at B=18B = \frac{1}{8}: 500 COCO pairs over all ordered placements (12 for N=4N = 4, 56 for N=8N = 8). Joint: joint-question accuracy; Both: both constituents correct; All 3: all answers correct; Worst: minimum positional joint accuracy. CF-A/CF-B: correct joint answers before and after replacing A/BA/B (394/373 eligible pairs). Bold marks the best compressed method.

Metrics. Joint averages joint-question accuracy over pairs and placements. Both requires both constituent answers to be correct; All 3 also requires the joint answer. Worst is minimum joint accuracy across placements. CF-A/CF-B require correct joint answers on both original and replacement pairs within the eligible subset. We use 10,000 paired bootstrap resamples over image pairs, retaining all placements, constituent calls, and replacements per sampled pair.

Table S6. Ablation study for each property on Qwen3-VL-8B. Each row ablates one component of SIEVE. Δ\Delta denotes the change relative to SIEVE.

MantisNLVR2MuirBLINKQ-Bench2AvgΔ\Delta
SIEVE60.8373.2047.2744.2968.5058.82
(1) Cross-image coverage
Budget allocation (SIEVE: equal quota per image)
one global top-kk, no per-image quota59.4572.0543.8144.2967.3057.38-1.44
size-proportional share per image59.4572.8443.9644.7167.2057.63-1.19
Selection score (SIEVE: prompt attention)
random score at all stages58.9972.4843.9641.4564.4056.26-2.56
random score at the first stage only58.9972.4643.8541.6164.4056.26-2.56
random score at later stages only59.4573.6244.0844.7168.4058.05-0.77
(2) Image identity
Merge target (SIEVE: most similar kept token of the same image)
drop unselected tokens, no merging60.3773.1044.1244.7167.8058.02-0.80
merge into spatially nearest kept token58.5372.9743.4644.2967.6057.37-1.45
most similar kept token from any image59.4573.4243.6944.4568.3057.86-0.96
assign by attention weight58.0673.3943.5044.0867.1057.23-1.59
merge into a random kept token60.8372.4644.3844.0867.2057.79-1.03
(3) Evidence throughout processing
Initial pooling
no pooling53.4660.0442.5443.0869.9053.80-5.02
strided pooling64.0670.8947.2340.7268.8058.34-0.48
Reduction after pooling (SIEVE: progressive select-and-merge)
pooling only, no later reduction58.9970.8142.1242.8265.0055.95-2.87
pooling + FastV’s pruning57.6067.1745.1544.1967.9056.40-2.42
pooling + PyramidDrop’s schedule62.2167.4944.9645.8267.6057.62-1.20
First stage (SIEVE: after layer ℓ1=2\ell_1 = 2)
first stage at ℓ1=0\ell_1 = 059.4572.7044.6943.7167.4057.59-1.23
first stage at ℓ1=4\ell_1 = 460.8372.2744.9644.5067.3057.97-0.85
first stage at ℓ1=8\ell_1 = 858.5369.1844.8143.6166.9056.61-2.21
Later-stage placement (SIEVE: second half of the remaining depth)
later stages evenly spaced over depth59.4574.3245.0445.1368.0058.39-0.43
later stages in the first half of depth57.6070.3945.5043.7767.7056.99-1.83
Number of stages (SIEVE: S(N)=clip⁡(N,2,4)S(N) = \operatorname{clip}(N, 2, 4))
fixed S=2S = 258.9973.6344.1544.8768.4058.01-0.81
fixed S=3S = 362.2174.3244.0844.6168.1058.66-0.16
Budget schedule (SIEVE: geometric decay)
same retention ratio at every stage59.4573.4744.0444.6668.5058.02-0.80

Table S6.

References

  1. [1]Saeed Ranjbar Alvar, Gursimran Singh, Mohammad Akbari, and Yong Zhang. Divprune: Diversity-based visual token pruning for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9392–9401, 2025. 2arxiv.org/abs/2503.02175
  2. [2]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shitong Jiang, Zhao-hai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Jun-yang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yucheng Sun, Jun Tang, Jianhong Tu, Jian-qiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yux-uan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu. Qwen3-vl technical report, 2025. 5arxiv.org/abs/2511.21631
  3. [3]Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao-hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2025. 5arxiv.org/abs/2502.13923
  4. [4]Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster, 2023. 4arxiv.org/abs/2210.09461
  5. [5]Bin Chen, Yuxiang Cai, Yadan Luo, Yi Zhang, Jianwei Yin, and Zhi Chen. Spectral evolution-guided token pruning in multimodal large language models, 2026. 2, 3, 5, 6, 7, 14, 17, 18, 19arxiv.org/abs/2606.24165
  6. [6]Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, pages 27056–27087. Curran Associates, Inc., 2024. 5arxiv.org/abs/2403.20330
  7. [7]Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models, pages 19–35. Springer Nature Switzerland, 2024. 2, 3, 5, 6, 7, 14, 17, 18, 19arxiv.org/abs/2403.06764
  8. [8]Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Caifeng Shan, Ran He, and Xingqi Sun. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24108–24118, 2025. 5arxiv.org/abs/2405.21075
  9. [9]Xingyu Fu, Yushu Hu, Bangzhen Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. BLINK: Multimodal Large Language Models Can See but Not Perceive, page 148–166. Springer Nature Switzerland, 2024. 5, 13arxiv.org/abs/2404.12390
  10. [10]Yuanyuan Jia, Shunpu Tang, and Qianqian Yang. S²prune: Spatially structured visual token pruning for multimodal large language models, 2026. 2, 3, 5, 6, 7, 14, 17, 18, 19arxiv.org/abs/2609.01224
  11. [11]Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning, 2024. 2, 5arxiv.org/abs/2405.01483
  12. [12]Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 5arxiv.org/abs/2408.03326
  13. [13]Duo Li, Zuhua Yang, Xiaoqin Zhang, Ling Shao, and Shijian Lu. Todre: Effective visual token pruning via token diversity and task relevance. In European Conference on Computer Vision, pages 635–653. Springer, 2026. 2, 3, 5, 6, 7, 17, 18, 19arxiv.org/abs/2505.18757
  14. [14]Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 292–305, Singapore, 2023. Association for Computational Linguistics. 5arxiv.org/abs/2305.10355
  15. [15]Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context, page 740–755. Springer International Publishing, 2014. 1, 15arxiv.org/abs/1405.0312
  16. [16]Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, pages 2507–2521. Curran Associates, Inc., 2022. 5arxiv.org/abs/2209.09513
  17. [17]Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. In Advances in Neural Information Processing Systems, pages 46212–46244, 2023. 5arxiv.org/abs/2308.09126
  18. [18]Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5arxiv.org/abs/1904.08920
  19. [19]Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6418–6428, Florence, Italy, 2019. Association for Computational Linguistics. 5arxiv.org/abs/1811.00491
  20. [20]Fei Wang, Xingyu Fu, James Y. Huang, Zekun Li, Qin Liu, Xiaogen Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, Tianyi Lorena Yan, Wenjie Jacky Mo, Hsiang-Hui Liu, Pan Lu, Chunyuan Li, Chaowei Xiao, Kai-Wei Chang, Dan Roth, Sheng Zhang, Hoifung Poon, and Muhao Chen. Muirbench: A comprehensive benchmark for robust multi-image understanding, 2024. 2, 5arxiv.org/abs/2406.09411
  21. [21]Yuhao Wang, Mu Qiao, Haiwen Diao, Yunzhi Zhuge, Pingping Zhang, Xindong Zhang, Lei Zhang, and Huchuan Lu. Era: Entropy-guided visual token pruning with rectified attention for efficient mllms, 2026. 3arxiv.org/abs/2606.31982
  22. [22]Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zhang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, and Dahua Lin. Conical visual concentration for efficient large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14593–14603, 2025. 2, 3, 5, 6, 7, 14, 17, 18, 19DOI
  23. [23]Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19792–19802, 2025. 2, 3, 5, 6, 7, 14, 17, 18, 19arxiv.org/abs/2412.04467
  24. [24]Kang Zeng, Guojin Zhong, Jintao Cheng, Jin Yuan, and Zhiyong Li. Avam: a universal training-free adaptive visual anchoring embedded into multimodal large language model for multi-image question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2292–2302, 2025. 3
  25. [25]Ce Zhang, Kaixin Ma, Tianqing Fang, Wenhao Yu, Hongming Zhang, Zhiyong Zhang, Haitao Mi, and Dong Yu. Vscan: Rethinking visual token reduction for efficient large vision-language models, 2026. 2, 3arxiv.org/abs/2505.22654
  26. [26]Qizhe Zhang, Mengzhen Liu, Lichen Li, Ming Lu, Yuan Zhang, Junwen Pan, Qi She, and Shanghang Zhang. Beyond attention or similarity: Maximizing conditional diversity for token pruning in mllms. In Advances in Neural Information Processing Systems, pages 25438–25468. Curran Associates, Inc., 2025. 3arxiv.org/abs/2506.10967
  27. [27]Rongyang Zhang, Chengqiang Lu, Cong Li, Hongchao Gu, Tingjia Shen, Xuyang Zhi, Qiming Wang, Yan Gao, Yi Wu, Yao Hu, Hao Wang, and Enhong Chen. Multi-image visual token pruning in large visual language models, 2026. 2, 3, 5, 6, 7, 13, 17, 18, 19arxiv.org/abs/2608.26806
  28. [28]Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis A Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang. SparseVLM: Visual token sparsification for efficient vision-language model inference. In Proceedings of the 42nd International Conference on Machine Learning, pages 74840–74857. PMLR, 2025. 2, 3, 5, 6, 7, 14, 17, 18, 19arxiv.org/abs/2410.04417
  29. [29]Zicheng Zhang, Haoning Wu, Erli Zhang, Guangtao Zhai, and Weisi Lin. Q-bench+: A benchmark for multi-modal foundation models on low-level vision from single images to pairs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10404–10418, 2024. 5arxiv.org/abs/2402.07116
  30. [30]Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, Hao Li, Jiahao Wang, Nianchen Deng, Songze Li, Yinan He, Tan Jiang, Jiapeng Luo, Yi Wang, Conghui He, Botian Shi, Xingcheng Zhang, Wenqi Shao, Junjun He, Yingtong Xiong, Wenwen Qu, Peng Sun, Penglong Jiao, Han Lv, Lijun Wu, Kaipeng Zhang, Huipeng Deng, Jiaye Ge, Kai Chen, Limin Wang, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 5

Paper details

vlnc:W_3qjyn85s
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents