Pith. sign in

REVIEW 3 major objections 6 minor 17 references

The paper argues that choosing which vision blocks to skip is a combinatorial search problem, and a source-balanced evolutionary search finds stronger fixed-budget routes than independent ranking or random selection.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 19:09 UTC pith:ULFM3HKJ

load-bearing objection A transparent and well-executed empirical study, but the central claim that evolutionary search beats naive pruning is evaluated on the same set used to freeze the searched routes, so the headline result is not yet sealed. the 3 major comments →

arxiv 2607.17052 v1 pith:ULFM3HKJ submitted 2026-07-19 cs.CV

Searching for Task-Specific Vision Paths: Evolutionary Block Pruning Across Vision-Language Models

Tarun Tomar This is my paper
classification cs.CV MSC 68T4568T20
keywords vision-language modelsvision block pruningevolutionary searchstructured pruningcapability-specific routingcombinatorial optimizationmodel efficiencyOCR transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether a vision-language model must run every vision block for every question, and whether the best blocks to skip can be found by looking at each block's importance alone. It shows that they cannot: blocks that look safe individually can become harmful when skipped together, so choosing a fixed set of K blocks to skip is a combinatorial search problem. The paper introduces a source-balanced evolutionary search that consistently constructs stronger same-budget routes than independent ranking, contiguous removal, or random selection on two architectures. It also shows that capability-specific routes — routes tuned for OCR, counting, spatial reasoning, attributes, or object existence — are not universally better or transferable. A shared route often beats a capability-specific route, especially on a sealed transfer test, so the paper concludes that combinatorial search reliably improves route construction but capability labels do not define stable vision pathways.

Core claim

On the paper's own terms, the central discovery is that route quality is non-additive: the damage from skipping a set of blocks is not predicted by the sum of individual single-block damages. A source-balanced evolutionary search treats a route as one complete set of exactly K blocks and optimizes a multi-objective loss over capability-source cells, including worst-source and variability terms. This search beats independent ranking by 1.03, 1.03, and 3.54 percentage points on Qwen at four, six, and eight skipped blocks, and by 4.91 points on SmolVLM2 at four skipped blocks. Capability specialization gives a borderline 2.17-point gain on Qwen at six blocks, driven by a 7.10-point OCR gain, bu

What carries the argument

A route is a set of exactly K vision-transformer blocks replaced by identity operations. The key mechanism is the source-balanced evolutionary search: a genetic algorithm with deterministic parent pairing, fixed-K crossover that keeps the intersection of two parental routes, and one-swap mutation, guided by Pareto survival on a multi-objective vector that balances mean damage, worst-source damage, and source variability (plus collateral damage for capability-specific routes). This machinery lets the search evaluate complete block combinations rather than composing individual importance scores, which is necessary because residual blocks interact.

Load-bearing premise

The headline comparisons use the 876-example selection set both to pick the frozen route and to measure the search advantage, so the results assume that selection over these 876 examples does not materially inflate the apparent gain of the searched routes.

What would settle it

Take the same two models and the same search and control methods, but freeze all routes on the 876-example set and then evaluate on a genuinely sealed benchmark from sources not used in any screening. If the searched routes' advantage over independent ranking shrinks or reverses there, the selection-set reuse rather than the search itself would be the likely explanation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Independent layer-importance scores should not be used to compose multi-block reductions in vision encoders; they miss block interactions.
  • Any fixed-budget depth pruning of a vision-language model is better done by combinatorial search than by ranking, especially at larger skip budgets.
  • Capability-specific routes should be validated on held-out sources before deployment; a route named for OCR may not transfer across OCR benchmarks.
  • The same skip budget has different costs across architectures, so pruning budgets need to be set per model, not globally.
  • Route search can reduce executed vision depth and give modest end-to-end speedups, but the checkpoint itself is not smaller until parameters are physically removed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A learned router that conditions on image statistics or question text rather than a known capability label is a natural next step; the paper's oracle-label assumption is a deployment bottleneck.
  • The interaction measure I(i,j) could be used to predict when independent ranking will fail, potentially giving a cheaper screening test before running full search.
  • The paper's search advantage is measured on the same selection set used to freeze routes; a strictly sealed multi-source evaluation would test whether the advantage persists without selection optimism.
  • Since the paper only varies skip budgets K=4,6,8 on one model and K=4 on another, extending to larger budgets and more architectures would show whether the search advantage grows monotonically.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper studies training-free, fixed-budget pruning of vision-encoder blocks in VLMs by replacing selected blocks with identity operations. It introduces a source-balanced evolutionary search for shared and capability-specific block routes, and compares the searched routes against independent-ranking, contiguous-removal, and random baselines at matched skip budgets on Qwen2.5-VL-3B and SmolVLM2-2.2B. The central positive claim is that combinatorial search reliably improves route construction; the central negative claim is that capability-specific routes are not stable or transferable. The paper also reports a sealed IIIT5K audit showing that the SmolVLM2 OCR-specific route underperforms the shared route by 13.6 pp.

Significance. If the positive claim were validated on properly sealed data, the paper would make a useful contribution: it demonstrates that block interactions matter in VLM vision encoders, that simple independent rankings do not compose, and that a source-balanced evolutionary search can find better fixed-cardinality routes. The negative result on capability-specific routing is also valuable and is supported by a genuinely fresh IIIT5K evaluation. The paper is commendably reproducible: frozen routes are listed in Table 2, search configurations are hashed, code is released, matched budgets are enforced, and the authors explicitly disclose that the 876-example set is method-selection evidence rather than a sealed test set. However, the headline search-vs-baseline comparison is evaluated on the same 876-example set used to freeze the searched route, so the positive claim currently rests on an in-sample comparison. The IIIT5K audit does not test the search-vs-baseline claim.

major comments (3)
  1. [§3.5, §4.2] The central comparison is partially in-sample. The frozen route is chosen as the lowest-loss finalist on the 876-example selection set (§3.5), and §4.2 then compares that route against independent, contiguous, and random baselines on the same 876 examples. The baselines are not selected on this set, so the searched route's advantage includes selection-induced optimism. At n=876, the standard error of an accuracy difference is roughly 1.3–1.5 pp; selecting the best of up to three finalists can inflate the apparent advantage by a comparable amount, which covers much of the reported Qwen gains (1.03, 1.03, 3.54 pp) and a nontrivial part of the SmolVLM2 gain (4.91 pp). The authors explicitly acknowledge in §4.1 and §5 that the 876 set is 'method-selection evidence rather than... a sealed test set,' but the IIIT5K audit in §4.5 is fresh only for the shared-vs-OCR comparison, not for search-vs
  2. [§4.2, §3.5] The Qwen search-vs-baseline results are reported as point differences without the paired bootstrap intervals promised in §3.5. For example, the searched shared route beats independent ranking by 1.03 pp at K=4 and 1.03 pp at K=6, and by 3.54 pp at K=8, but no intervals are given. With n=876, 1.03 pp is within plausible sampling noise, so the reader cannot judge whether these differences are meaningful. The SmolVLM2 replication reports intervals, but the Qwen numbers—which are the primary model's main evidence—do not. Report paired bootstrap intervals for all headline comparisons. If those intervals are computed on the 876 set after finalist selection, state explicitly that they do not correct for model selection.
  3. [§4.2, §4.4, Abstract] The cross-architecture generalization rests on a single SmolVLM2 budget (K=4). The SmolVLM2 replication runs only the four-block study, while the Qwen study covers K=4, 6, and 8. The abstract and conclusion say 'search transfers across architectures' and 'combinatorial search reliably improves route construction,' but the evidence for transfer is one architecture at one budget. This is a scope issue rather than an internal error, but it should be reflected in the claims: the manuscript should either add SmolVLM2 results at additional budgets or replace 'reliably' with a statement scoped to the tested budgets.
minor comments (6)
  1. [§2.2] The progressive ablation reports a 4.89% vision-encoder latency improvement for a route that is not listed in Table 2. Clarify which route this is and why this latency number is not used in §4.6.
  2. [Table 1] The text has a typo: 'T able 1' instead of 'Table 1'.
  3. [§4.6, Figure 8] The latency measurement is an unlocked same-VM comparison; the caption discloses this, but the main text should state prominently that the speedup is not a fixed-clock or edge-device measurement. It already does in the figure caption, but consider moving that caveat into the body text.
  4. [§4.2] The 'mean of three random routes' is reported without showing the spread of the three individual random routes. Reporting the individual values or a range would be more informative, since a single good random route could drive the mean.
  5. [Figure 5] The figure compares Qwen at K=6 with SmolVLM2 at K=4. The caption is transparent, but the visual pairing may be misread as a matched comparison. Consider labeling each panel with its budget more prominently.
  6. [§3.3] The scalar loss weights (0.50/0.30/0.20 and 0.45/0.30/0.15/0.10) are introduced without sensitivity analysis. A brief note that the conclusions are robust or not to these weights would strengthen the paper.

Circularity Check

1 steps flagged

Headline search-vs-baseline comparison is evaluated on the same 876-example set used to freeze the searched route; the central claim therefore rests on an in-sample selection advantage rather than a sealed test.

specific steps
  1. fitted input called prediction [§3.5, §4.2, §5]
    "At most three routes with the lowest matching scalar loss advance to the 876-example image-disjoint selection set, and the route with the lowest selection loss is frozen. ... The 876-example set is used for method selection and is not described as sealed evidence because earlier discovery work touched its underlying benchmark sources. ... Only the 250-example IIIT5K audit is a genuinely fresh post-freeze source-transfer test."

    The frozen route is chosen as the lowest-loss candidate on the 876-example set, and §4.2 then reports that same route's accuracy on that same 876-example set against independent, contiguous, and random controls. The search route is therefore selected on the evaluation set, while the controls are not, so the reported gains (1.03–3.54 pp on Qwen; 4.91 pp on SmolVLM2) partly reflect best-of-at-most-three selection optimism on the exact examples used for the comparison. The paired bootstrap resamples example IDs but does not resample the selection procedure, so the intervals understate this in-sample advantage. The paper candidly labels the 876 set as method-selection evidence rather than a sealed test, but the central claim that combinatorial search reliably improves route construction is not

full rationale

The paper's mathematical machinery is internally consistent: the source-balanced objectives, Pareto survival, crossover, and mutation are not defined in terms of the headline results, and no self-citation chain is load-bearing. The circularity arises at the evaluation step. §3.5 freezes the route with the lowest loss on the 876-example selection set, and §4.2 uses that same set to compare the frozen route against non-selected baselines. This is a fitted-input-called-prediction pattern: the searched route is effectively a parameter selected on the evaluation data, and its reported superiority over independent/contiguous/random routes is at least partly an artifact of that selection. The paper explicitly acknowledges the limitation in §4.1 and §5, which is honest, but the acknowledgment does not remove the circularity from the central empirical claim. The IIIT5K audit is genuinely post-freeze and supports the secondary claim about non-transferable OCR pathways, but it does not validate the primary search-vs-baseline claim. The bootstrap intervals quantify sampling noise over examples, not selection-induced optimism, so the reported confidence intervals are too narrow for the claim they support. Overall, the derivation is partially circular: the central positive result is not backed by a sealed comparison, though the paper's own caveats prevent this from being a fully concealed circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claims rest on hand-chosen search weights and budgets, the oracle-label assumption, the identity-skip proxy, and the validity of the 876-set for both selection and comparison. No new physical or architectural entities are introduced.

free parameters (5)
  • Scalar loss weights (shared route) = 0.50, 0.30, 0.20
    Weights for mean damage, worst-source damage, and source variability in Lshared (§3.3). Hand-chosen; affect finalist ranking.
  • Scalar loss weights (capability route) = 0.45, 0.30, 0.15, 0.10
    Weights for target mean, target worst, collateral damage, and target variability (§3.3). Hand-chosen; affect which routes are preferred.
  • Evolutionary budgets = Qwen: pop 16, 3 generations, K=4,6,8; SmolVLM2: pop 12, 2 generations, K=4
    Search effort and skip budgets are chosen by hand and frozen before the SmolVLM2 replication (§3.5). They determine which routes can be found.
  • Search seeds = 20260715, 20260716, 20260717
    Three fixed seeds per search. Results may vary across seeds; the paper reports low Jaccard overlap between seed winners, indicating instability.
  • Finalist selection cap = at most 3 finalists per route family
    Limits how many seed finalists are evaluated on the 876-example selection set (§3.5). Affects the degree of selection bias in the comparison.
axioms (6)
  • domain assumption Identity operation is a valid substitute for a skipped block without fine-tuning or compensation
    Replacing a block with identity changes the input distribution to subsequent blocks; the paper treats this as measuring the effect of reduced depth. No low-rank compensation (unlike Short-LVLM) is applied.
  • domain assumption Capability labels are known (oracle) and each question maps to exactly one capability
    The capability-specific policy selects a route using a known label without routing errors or overhead (§3.2). This is an upper-bound evaluation of label-based routing.
  • ad hoc to paper Equal-weight source-balanced cell aggregation is the correct optimization objective
    Every capability-source cell receives equal weight regardless of example count (§3.3). This is a modeling choice that shapes which routes are considered best.
  • domain assumption The 876-example selection partition is representative of the capability distributions
    The partition has 155-193 examples per capability and is used for method selection. The paper acknowledges it is not a pristine held-out benchmark (§4.1).
  • domain assumption Greedy decoding and short-answer exact-match accuracy are valid evaluation metrics
    The paper evaluates deterministic short-answer correctness, not long-form generation, calibration, grounding, or safety (§5).
  • domain assumption The evolutionary search with small populations and few generations finds routes near the true optimum
    The authors state every selected route is 'the best found under the frozen search rather than a globally optimal subset' (§5). The search effort is finite.

pith-pipeline@v1.3.0-alltime-deepseek · 12347 in / 12346 out tokens · 118895 ms · 2026-08-01T19:09:13.290216+00:00 · methodology

0 comments
read the original abstract

Vision-language models normally execute the same complete vision encoder for every question, even when OCR, counting, object, attribute, and spatial queries may not require identical computation. We study whether fixed-budget combinations of vision blocks can be skipped without fine-tuning. A shared K-block route skips one searched set of exactly K blocks for every question, while a capability-specific K-block policy selects one same-size route using a known capability label. We introduce a source-balanced evolutionary search and compare it with independent ranking, contiguous removal, and random routes at matched budgets. Experiments use Qwen2.5-VL-3B-Instruct, SmolVLM2-2.2B-Instruct, and an 876-example image-disjoint selection split. Search transfers across architectures: on SmolVLM2, the searched shared four-block route beats independent construction by 4.91 percentage points. Capability specialization is less stable. On Qwen, the six-block capability policy beats the shared route by 2.17 points, driven by a 7.10-point OCR gain. On sealed IIIT5K, however, the SmolVLM2 OCR-specific route trails its shared route by 13.6 points. Combinatorial search reliably improves route construction, but capability labels do not define universally transferable vision pathways.

Figures

Figures reproduced from arXiv: 2607.17052 by Tarun Tomar.

Figure 1
Figure 1. Figure 1: Accuracy drop after replacing one Qwen2.5-VL-3B vision block at a time with identity. Most blocks cause limited average damage when skipped alone, but sensitivity varies by capability and depth. This is screening evidence: a low single-block drop does not imply that several such blocks can be safely combined or that a capability is stored in a specific block. We use Qwen2.5-VL-3B-Instruct as the primary mo… view at source ↗
Figure 2
Figure 2. Figure 2: Fixed-budget evolutionary route search. Candidate routes are evaluated with paired source-aware objectives, filtered by Pareto survival, evolved through crossover and mutation, and frozen only after three-seed finalist selection. Held-out transfer is an audit after the route is frozen. The removal and addition are again selected by seeded SHA-256 ordering. Duplicate children are rejected, and a determinist… view at source ↗
Figure 3
Figure 3. Figure 3: Qwen selection accuracy as the number of skipped blocks increases. Every condition at a given x-axis position skips exactly the same number of vision blocks. SmolVLM2 is omitted because only its four-block study was completed [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy for matched four-block controls. Dashed lines indicate the corresponding full-model accuracy. Search improves route construction on both architectures, while SmolVLM2 remains more damaged by four skipped blocks overall. Attribute Counting Object OCR Spatial Qwen: skip 6 SmolVLM2: skip 4 +1.2 +1.1 +0.0 +7.1* +2.2 -0.6 +7.2* -0.5 -13.5* +9.4* * paired 95% interval excludes zero Capability-specific g… view at source ↗
Figure 5
Figure 5. Figure 5: Capability-policy accuracy minus shared-route accuracy. Stars mark paired 95% intervals that exclude zero. Qwen is shown at its strongest specialization budget, while SmolVLM2 is shown at its completed four-block budget. MME and OCRBench examples used during selection, exploited architecture-specific interactions that do not support IIIT5K text, or won because several low-stability candidates had similar d… view at source ↗
Figure 6
Figure 6. Figure 6: Post-freeze source transfer on IIIT5K. Both compressed conditions skip four SmolVLM2 vision blocks. Routes and all evaluation choices were frozen before any IIIT5K prediction was produced. Shared Attribute Counting Object OCR Spatial 0.0 0.2 0.4 0.6 0.8 1.0 Mean pairwise Jaccard across seed winners Low route stability warns against a fixed capability map Qwen: skip 4 SmolVLM2: skip 4 [PITH_FULL_IMAGE:figu… view at source ↗
Figure 7
Figure 7. Figure 7: Mean pairwise Jaccard overlap among three seed-level four-block development winners. Low overlap indicates that the frozen route is one strong solution among several distinct candidates, not a uniquely identified block map [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Four-block parameter fractions for both models and measured SmolVLM2 latency. The latency measurement is an unlocked same-VM comparison on one RTX 4090, not fixed-clock or edge-device evidence. For the searched SmolVLM2 shared four-block route, repeated batch-size-one measurements show an 8.60% vision-encoder speedup and a 4.19% end-to-end speedup on the same RTX 4090 VM. The smaller end-to-end gain is exp… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 1 canonical work pages

  1. [6]

    Searching for Task-Specific Vision Paths 14 Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh

    URL https://arxiv.org/abs/2306.13394. Searching for Task-Specific Vision Paths 14 Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. InIEEE Conference on Computer Vision and Pattern Recognition,

    Pith/arXiv arXiv

  2. [12]

    Fangyu Liu, Guy Emerson, and Nigel Collier

    URL https://arxiv.or g/abs/2305.10355. Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. InTransactions of the Association for Computational Linguistics, 2023a. URL https://arxiv.org/abs/2205.00363. Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench:...

    Pith/arXiv arXiv

  3. [13]

    URL https://arxiv.org/abs/2507.23362

    doi: 10.1145/3746027.3755122. URL https://arxiv.org/abs/2507.23362. Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. Shortgpt: Layers in large language models are more redundant than you expect.arXiv preprint arXiv:2403.03853,

    arXiv

  4. [14]

    Anand Mishra, Karteek Alahari, and C

    URL https://arxiv.org/abs/2403.03853. Anand Mishra, Karteek Alahari, and C. V. Jawahar. Scene text recognition using higher order language priors. InBritish Machine Vision Conference,

    Pith/arXiv arXiv

  5. [17]

    Ce Zhang, Kaixin Ma, Tianqing Fang, Wenhao Yu, Hongming Zhang, Zhisong Zhang, Haitao Mi, and Dong Yu

    URL https://proceedings.neurips.cc/paper_files/paper/2025/hash/8 8441fee0c0dacceb95e4ad77ece5d0a-Abstract-Conference.html. Ce Zhang, Kaixin Ma, Tianqing Fang, Wenhao Yu, Hongming Zhang, Zhisong Zhang, Haitao Mi, and Dong Yu. Vscan: Rethinking visual token reduction for efficient large vision-language models.Transactions on Machine Learning Research,

    2025

  6. [18]

    Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al

    URL https://arxiv.org/abs/2505.22654. Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevlm: Visual token sparsification for efficient vision-language model inference. InInternational Conference on Machine Learning,

    arXiv

  7. [19]

    URL https: //arxiv.org/abs/2410.04417

    Pith/arXiv arXiv

  8. [2002]

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji

    doi: 10.1109/4235.996017. Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394,

    arXiv

  9. [2012]

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh

    doi: 10.5244/C.26.127. Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. InAdvances in Neural Information Processing Systems,

    work page doi:10.5244/c.26.127

  10. [2017]

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al

    URL https://arxiv.org/abs/1612.00837. Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models.arXiv preprint arXiv:2203.15556,

    Pith/arXiv arXiv

  11. [2019]

    Saleh Ashkboos, Maximilian L

    URL https://arxiv.org/abs/1810.12440. Saleh Ashkboos, Maximilian L. Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, and James Hensman. Slicegpt: Compress large language models by deleting rows and columns. InInternational Conference on Learning Representations,

    Pith/arXiv arXiv

  12. [2020]

    Saeed Khaki, Nima Safaei, and Kamal Ginotra

    URL https://arxiv.org/abs/2001.08361. Saeed Khaki, Nima Safaei, and Kamal Ginotra. Understanding pruning regimes in vision-language models through domain-aware layer selection.arXiv preprint arXiv:2603.20275,

    Pith/arXiv arXiv 2001

  13. [2021]

    Sreetama Sarkar, Saket Gurukar, Yuexi Zhang, Monami Banerjee, and Ashwin Chandra

    URL https://arxiv.org/abs/2106.02034. Sreetama Sarkar, Saket Gurukar, Yuexi Zhang, Monami Banerjee, and Ashwin Chandra. Flashvlm: Exploiting layer redundancy via visual attention for efficient vision-language inference.OpenReview preprint,

    Pith/arXiv arXiv

  14. [2022]

    Hugging Face

    URL https://arxiv.org/abs/2203.15556. Hugging Face. Smolvlm2-2.2b-instruct model card. Hugging Face model repository,

    Pith/arXiv arXiv

  15. [2023]

    Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T

    URL https://arxiv.org/abs/2210.09461. Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii.IEEE Transactions on Evolutionary Computation, 6(2):182–197,

    Pith/arXiv arXiv

  16. [2024]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al

    URL https://arxiv.org/abs/2401.15024. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923,

    Pith/arXiv arXiv

  17. [2025]

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman

    URL https://arxiv.org/abs/2502.13923. Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. InInternational Conference on Learning Representations,

    Pith/arXiv arXiv