Pith. sign in

REVIEW 2 major objections 5 minor 60 references

DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation

T0 review · 2 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Separate semantic and spatial cues, then learn which branch to trust pixel by pixel, yields stronger reasoning segmentation.

desk verdict Clean dual-branch + gate recipe that actually moves ReasonSeg numbers; soft spots are real but not load-bearing. read the letter →

arxiv 2607.04779 v1 pith:OZAB5BOV submitted 2026-07-06 cs.CV

Ruizhe Zeng , Siyu Cao , Lu Zhang , Zhiyong Liu This is my paper · ORCID
classification cs.CV
keywords reasoningsegmentationdynamicgatingsemantic-spatialcuespromptableMLLMlearnablefusiondual-branch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Reasoning segmentation asks a system to turn a complex language query into a precise pixel mask. Most current pipelines let a multimodal language model compress its reasoning into sparse cues such as points or boxes and then feed those cues into a single segmentation model. The compression step often injects ambiguity or noise, and when several cues are simply concatenated, one bad cue can spoil the whole mask. DGSeg instead asks the language model for two complementary descriptions of the same target—what it is and where it is—runs each cue through its own segmentation branch, and trains a lightweight gate to fuse the two predictions using relative branch-quality supervision. The gate learns to raise the weight of the more reliable branch in each region, so noisy or conflicting evidence is suppressed. On the hard ReasonSeg benchmark the method reaches 69.6 % and 67.3 % gIoU on validation and test with a 7B backbone, outperforming strong recent baselines while adding almost no compute.

What carries the argument

Dynamic gating module: a two-layer convolutional network that reads concatenated pixel-decoder features from the semantic and spatial branches and outputs a pixel-wise fusion weight map, trained with soft targets derived from each branch’s IoU against ground truth plus an entropy penalty.

What would settle it

Replace the learned gate with simple averaging or with the stronger single branch alone on the same dual-branch outputs; if the final gIoU on ReasonSeg no longer rises above the stronger branch or the average, the claim that relative-quality supervision is necessary collapses.

Watch

Extended reading notes

Core claim

Isolating complementary semantic and spatial cues into separate segmentation branches and training a dynamic gate with relative branch-quality supervision lets the system suppress ambiguous or erroneous regions that would otherwise contaminate a joint prediction, producing masks that stay closer to the original reasoning intent.

Load-bearing premise

The soft fusion target built from each branch’s IoU against the ground-truth mask is assumed to give an unbiased enough teaching signal for the gate, even though both branches share the same frozen backbone and their quality already depends on the cues produced in the first training stage.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. DGSeg addresses reasoning segmentation by having an MLLM produce complementary semantic (textual description) and spatial (bounding box) cues, routing them through separate SAM3 branches, and fusing the resulting mask logits with a lightweight dynamic gating module. The gate is trained with relative branch-quality supervision derived from per-branch IoU against ground truth (Eqs. 8–9), plus an entropy penalty, under a two-stage protocol (GRPO for cue generation, then supervised gate training). On ReasonSeg under a zero-shot setting the method reports 66.0/60.0 gIoU (3B) and 69.6/67.3 gIoU (7B), with competitive RefCOCO-family results and ablations that isolate dual-branch isolation, reward design, and learnable fusion.

Significance. If the reported gains hold, the work supplies a practical and efficient alternative to iterative MLLM refinement for mitigating noisy intermediate cues in MLLM-guided segmentation. The dual-branch isolation plus learnable fusion pipeline is cleanly motivated, the ablations (Tables 3–5) and correlation analysis (Fig. 3) give concrete evidence that the gate tracks relative branch quality, and the computational overhead is modest (Table 6). Code release further strengthens reproducibility. The contribution is incremental rather than paradigm-shifting, but it is a solid, well-executed advance for the reasoning-segmentation literature.

major comments (2)
  1. Section 3.2 / Eq. (9): the soft target W* is computed from IoU of the same frozen SAM3 backbone against ground truth for both branches. While Table 3 shows that non-learnable and alternative learnable fusions already improve over joint encoding, the paper should more explicitly discuss whether this supervision can overfit to Stage-1 cue idiosyncrasies rather than generalizable reliability cues, and whether the gate transfers when Stage-1 cues are produced by a different MLLM or reward schedule.
  2. Table 1 and Section 4.1: training uses only 9k RefCOCOg instances with no ReasonSeg data, yet the strongest claims are zero-shot ReasonSeg numbers. A short analysis of failure modes on ReasonSeg (e.g., queries requiring multi-hop or functional reasoning where both cues fail) would better bound the claim that complementary cues plus gating preserve reasoning intent.
minor comments (5)
  1. Section 3.1: the precise architecture of F_dg (kernel sizes, channels of the two conv layers) is underspecified in the main text; a short description or pointer to the supplement would aid reimplementation.
  2. Eq. (2)–(4): notation for feature resolution (h,w) versus logit resolution is clear, but the upsampling operator for W↑ is not named; bilinear vs. nearest should be stated.
  3. Table 2: SAM3 Agent scores are markedly lower than other 7B baselines; a brief note on why the comparison is still informative would help readers interpret the gap.
  4. Fig. 3 caption and surrounding text: the Pearson r=0.637 is reported without sample size or p-value; adding these would strengthen the correlation claim.
  5. Minor typography: occasional missing spaces after commas/periods and inconsistent hyphenation of “semantic–spatial” vs. “semantic-spatial” appear throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical dual-branch + gating pipeline evaluated zero-shot on held-out benchmarks; no equation or claim reduces reported gIoU to a fitted input by construction.

full rationale

DGSeg is a standard two-stage empirical CV pipeline (GRPO cue generation on RefCOCOg followed by supervised dynamic-gate training). The soft target W* (Eq. 9) is ordinary relative-quality supervision for a fusion head; it is never presented as a first-principles derivation of the final masks, and the headline numbers (69.6/67.3 gIoU) are measured zero-shot on ReasonSeg, which is never used in training. Ablations (Tables 3–5) show gains from dual-branch isolation alone and from multiple alternative fusion schemes, so the result is not forced by the particular form of W*. Citations are to external baselines (LISA, Seg-Zero, SAM3, etc.) and do not supply a uniqueness theorem or ansatz that the present claims rest upon. No self-definitional loop, fitted-parameter-as-prediction, or load-bearing self-citation chain exists. Circularity score is therefore zero.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The paper is an empirical systems contribution. It inherits the capabilities of Qwen2.5-VL and SAM3, introduces a small set of training hyper-parameters and a soft-target temperature, and postulates that relative IoU supplies a usable supervisory signal for the gate. No new physical entities or free-form theoretical constants appear.

free parameters (3)
  • temperature τ in soft fusion target W*
    Controls sharpness of the soft target derived from branch IoUs (Eq. 9); value chosen by authors and annealed.
  • loss weights λ_w, λ_ent and their annealing schedule
    Balance segmentation loss against gate supervision and entropy penalty; annealed over iterations (Section 3.2).
  • Stage-2 learning rate and weight decay = 1e-4
    1e-4 / 1e-4 for the two-layer gate network; standard but free choices.
assumptions (3)
  • domain assumption Qwen2.5-VL can be RL-finetuned with GRPO to emit parseable semantic descriptions and boxes that are useful for SAM3.
    Stage-1 training and all subsequent results rest on this capability (Section 3.2).
  • domain assumption SAM3’s pixel-decoder features contain sufficient information to judge relative reliability of two prompt-conditioned branches.
    The dynamic gate is defined to operate solely on those features (Eq. 2).
  • ad hoc to paper Relative IoU of the two branch masks against ground truth is a valid soft target for learning fusion weights.
    Introduced in Stage 2 (Eqs. 8–9) without external theoretical justification.
invented entities (2)
  • dynamic gating module F_dg
    purpose: Produce pixel-wise fusion weights from concatenated dual-branch features so that noisy regions can be suppressed.
    Lightweight two-conv network introduced by the paper; no independent existence outside this architecture.
  • relative branch-quality supervision (soft target W*)
    purpose: Provide an auxiliary training signal that teaches the gate which branch is locally more accurate.
    Constructed from per-branch IoUs; exists only inside the training procedure of this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation." pith.science (2026). https://pith.science/paper/OZAB5BOV

@misc{pith2026260704779,
  author       = {Pith},
  title        = {Pith review of: DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZAB5BOV}},
  note         = {Machine review of arXiv:2607.04779}
}
read the original abstract

Reasoning segmentation aims to predict pixel-wise masks for targets given complex language queries. Existing approaches leverage Multimodal Large Language Models (MLLMs) for vision-language reasoning and generate intermediate target cues (e.g., points or boxes) to guide a segmentation model. However, compressing rich reasoning into sparse cues often introduces ambiguity and noise, preventing these cues from accurately preserving the reasoning intent. While multiple complementary cues can enrich target information, existing methods typically feed them jointly into a single segmentation process, allowing ambiguous or erroneous cues to affect the entire prediction. Therefore, we propose DGSeg, a reasoning segmentation framework that learns to fuse predictions guided by semantic and spatial cues. Specifically, the MLLM jointly reasons about both target identity and spatial location, producing complementary semantic and spatial cues that are fed into separate segmentation branches. Their predictions are adaptively integrated by a lightweight dynamic gating module trained with relative branch-quality supervision to suppress noisy or conflicting regions. Extensive experiments demonstrate that DGSeg consistently outperforms strong baselines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on the challenging ReasonSeg validation and test splits. Code is available at https://github.com/RZZeng/DGSeg.

Figures

Figures reproduced from arXiv: 2607.04779 by the authors.

Figure 1
Figure 1. Motivation of our method. Existing approaches (e.g., Seg-Zero [33]) use an MLLM to generate target cues for downstream segmentation. However, they feed cues directly without mitigating potential noise, causing results to deviate from reasoning due to ambiguous or noisy cues (e.g., oversized boxes or erroneous points). In contrast, we leverage complementary semantic–spatial reasoning and a learnable fusion process to… view at source ↗
Figure 2
Figure 2. Overall framework of DGSeg. Given an image and a language query, the MLLM first generates complementary semantic and spatial cues. These cues are pro￾cessed by separate segmentation branches to produce initial results. A learnable dy￾namic gating module evaluates the features from dual-branch segmentation and adap￾tively fuses these results. before producing final outputs, we guide Freason to first generate an expli… view at source ↗
Figure 3
Figure 3. Visualization of the analysis on the dynamic gating module. Correlation Analysis of Fusion Weights. We further examine whether the proposed learnable fusion module in DGSeg effectively assesses segmentation features and filters unreliable regions from both global and local perspectives. For global correlation, we compute the IoU of the semantic and spatial branch predictions with the ground truth mask, and use their… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of gating results in DGSeg. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on ReasonSeg. We compare strong baselines in￾cluding Seg-Zero [33] and LENS [58]. For each example, the left shows the query and the right visualizes the predictions and the generated target cues. DGSeg produces semantic-spatial cues and performs…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 14 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2309.16609 (2023) 16 Zeng et al

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023) 16 Zeng et al

    arXiv 2023

  2. [2]

    arXiv preprint arXiv:2308.12966 (2023)

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)

    arXiv 2023

  3. [3]

    arXiv preprint arXiv:2502.13923 (2025)

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  4. [4]

    In: ECCV (2024)

    Bao, X., Sun, S., Ma, S., Zheng, K., Guo, Y., Zhao, G., Zheng, Y., Wang, X.: Cores: Orchestrating the dance of reasoning and segmentation. In: ECCV (2024)

    2024

  5. [5]

    arXiv preprint arXiv:2511.16719 (2025)

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala,K.V.,Khedr,H.,Huang,A.,etal.:Sam3:Segmentanythingwithconcepts. arXiv preprint arXiv:2511.16719 (2025)

    arXiv 2025

  6. [6]

    In: ECCV (2020)

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: ECCV (2020)

    2020

  7. [7]

    In: CVPR (2024)

    Chen, B., Xu, Z., Kirmani, S., Ichter, B., Sadigh, D., Guibas, L., Xia, F.: Spa- tialvlm: Endowing vision-language models with spatial reasoning capabilities. In: CVPR (2024)

    2024

  8. [8]

    In: ECCV (2024)

    Chen, Y.C., Li, W.H., Sun, C., Wang, Y.C.F., Chen, C.S.: Sam4mllm: Enhance multi-modal large language model for referring expression segmentation. In: ECCV (2024)

    2024

Show all 60 references
  1. [9]

    arXiv preprint arXiv:2412.05271 (2024)

    Chen, Z., Wang, W., Cao, Y., Liu, Y., Gao, Z., Cui, E., Zhu, J., Ye, S., Tian, H., Liu, Z., et al.: Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271 (2024)

    2024 arXiv

  2. [10]

    In: CVPR (2024)

    Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In: CVPR (2024)

    2024

  3. [11]

    In: NeurIPS (2017)

    Christiano, P.F., Leike, J., Brown, T.B., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. In: NeurIPS (2017)

    2017

  4. [12]

    In: ICRA (2024)

    Collins, J.A., Houff, C., Tan, Y.L., Kemp, C.C.: Forcesight: Text-guided mobile manipulation with visual-force goals. In: ICRA (2024)

    2024

  5. [13]

    In: NeurIPS (2023)

    Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)

    2023

  6. [14]

    In: CVPR (2009)

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)

    2009

  7. [15]

    In: ICLR (2026)

    Du, T., Li, H., Fan, Z., Zhang, J., Pan, P., Zhang, Y.: Sam-veteran: An mllm-based human-like sam agent for reasoning segmentation. In: ICLR (2026)

    2026

  8. [16]

    In: ICLR (2025)

    Fan, Q., Tao, X., Ke, L., Ye, M., Zhang, D., Wan, P., Tai, Y.W., Tang, C.K.: Stable segment anything model. In: ICLR (2025)

    2025

  9. [17]

    In: ICLR (2025)

    He, H., Li, G., Geng, Z., Xu, J., Peng, Y.: Analyzing and boosting the power of fine-grained visual recognition for multi-modal large language models. In: ICLR (2025)

    2025

  10. [18]

    arXiv preprint arXiv:2512.24023 (2025)

    He, X., Zhang, Y., Gao, S., Li, W., Hong, L., Chen, M., Jiang, K., Fu, J., Zhang, W.: Rsagent: Learning to reason and act for text-guided segmentation via multi- turn tool invocations. arXiv preprint arXiv:2512.24023 (2025)

    2025

  11. [19]

    In: ICLR (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)

    2022

  12. [20]

    In: CVPR (2024) DGSeg 17

    Hu, Y., Stretcu, O., Lu, C.T., Viswanathan, K., Hata, K., Luo, E., Krishna, R., Fuxman, A.: Visual program distillation: Distilling tools and programmatic rea- soning into vision-language models. In: CVPR (2024) DGSeg 17

    2024

  13. [21]

    In: NeurIPS (2025)

    Huang, J., Xu, Z., Zhou, J., Liu, T., Xiao, Y., Ou, M., Ji, B., Li, X., Yuan, K.: Sam-r1: Leveraging sam for reward feedback in multimodal segmentation via rein- forcement learning. In: NeurIPS (2025)

    2025

  14. [22]

    In: ICML (2021)

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., V.Le, Q., Sung, Y., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)

    2021

  15. [23]

    In: CVPR (2025)

    Kang, S., Kim, J., Kim, J., Hwang, S.J.: Your large vision-language model only needs a few attention heads for visual grounding. In: CVPR (2025)

    2025

  16. [24]

    In: EMNLP (2014)

    Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.L.: Referitgame: Referring to objects in photographs of natural scenes. In: EMNLP (2014)

    2014

  17. [25]

    In: ICCV (2023)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)

    2023

  18. [26]

    In: CVPR (2024)

    Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. In: CVPR (2024)

    2024

  19. [27]

    In: NeurIPS (2021)

    Li,J.,Selvaraju,R.R.,Gotmare,A.D.,Joty,S.,Xiong,C.,Hoi,S.C.H.:Alignbefore fuse: Vision and language representation learning with momentum distillation. In: NeurIPS (2021)

    2021

  20. [28]

    In: CVPR (2023)

    Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023)

    2023

  21. [29]

    In: ICLR (2025)

    Lin, Y., Li, H., Shao, W., Yang, Z., Zhao, J., He, X., Luo, P., Zhang, K.: Samrefiner: Taming segment anything model for universal mask refinement. In: ICLR (2025)

    2025

  22. [30]

    In: CVPR (2023)

    Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: CVPR (2023)

    2023

  23. [31]

    In: NeurIPS (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)

    2023

  24. [32]

    arXiv preprint arXiv:2510.16785 (2025)

    Liu, J., Chen, L.: Segmentation as a plug-and-play capability for frozen multimodal llms. arXiv preprint arXiv:2510.16785 (2025)

    2025

  25. [33]

    arXiv preprint arXiv:2503.06520 (2025)

    Liu, Y., Peng, B., Zhong, Z., Yue, Z., Lu, F., Yu, B., Jia, J.: Seg-zero: Reasoning-chain guided segmentation via cognitive reinforcement. arXiv preprint arXiv:2503.06520 (2025)

    2025 arXiv

  26. [34]

    arXiv preprint arXiv:2505.12081 (2025)

    Liu, Y., Qu, T., Zhong, Z., Peng, B., Liu, S., Yu, B., Jia, J.: Visionreasoner: Unified reasoning-integrated visual perception via reinforcement learning. arXiv preprint arXiv:2505.12081 (2025)

    2025

  27. [35]

    In: ACL (2025)

    Lu, Y., Cao, J., Wu, Y., Li, B., Tang, L., Ji, Y., Wu, C., Wu, J., Zhu, W.: Rsvp: Reasoning segmentation via visual prompting and multi-modal chain-of-thought. In: ACL (2025)

    2025

  28. [36]

    arXiv preprint arXiv:2510.11173 (2025)

    Lu, Z., Li, L., Wang, J., Feng, Y., Chen, B., Chen, K., Wang, Y.: Coprs: Learning positional prior from chain-of-thought for reasoning segmentation. arXiv preprint arXiv:2510.11173 (2025)

    2025

  29. [37]

    MIT press (2010)

    Marr, D.: Vision: A computational investigation into the human representation and processing of visual information. MIT press (2010)

    2010

  30. [38]

    In: CVPR (2025)

    Qian, R., Yin, X., Dou, D.: Reasoning to attend: Try to understand how< seg> token works. In: CVPR (2025)

    2025

  31. [39]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  32. [40]

    In: ICLR (2025) 18 Zeng et al

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. In: ICLR (2025) 18 Zeng et al

    2025

  33. [41]

    arXiv preprint arXiv:2401.14159 (2024)

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)

    2024 arXiv

  34. [42]

    In: CVPR (2024)

    Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixellm: Pixel reasoning with large multimodal model. In: CVPR (2024)

    2024

  35. [43]

    arXiv preprint arXiv:2402.03300 (2024)

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    2024 arXiv

  36. [44]

    arXiv preprint arXiv:2505.18816 (2025)

    Shen, Y., Li, C., Xiong, F., Jeong, J.O., Wang, T., Latman, M., Unberath, M.: Reasoning segmentation for images and videos: A survey. arXiv preprint arXiv:2505.18816 (2025)

    2025 arXiv

  37. [45]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  38. [46]

    In: CVPR (2024)

    Vuong, A.D., Vu, M.N., Huang, B., Nguyen, N., Le, H., Vo, T., Nguyen, A.: Language-driven grasp detection. In: CVPR (2024)

    2024

  39. [47]

    arXiv preprint arXiv:2409.12191 (2024)

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

    2024 arXiv

  40. [48]

    Wang,S.,Fang,G.,Kong,L.,Li,X.,Xu,J.,Yang,S.,Li,Q.,Zhu,J.,Wang,X.:Pix- elthink:Towardsefficientchain-of-pixelreasoning.arXivpreprintarXiv:2505.23727 (2025)

    2025 arXiv

  41. [49]

    In: ICASSP (2025)

    Wang, Y., Gao, D., Li, B., Long, R., Yi, L., Cai, X., Yang, L., Zhang, J., Yu, S., Xuan, Q.: Cof: Coarse to fine-grained image understanding for multi-modal large language models. In: ICASSP (2025)

    2025

  42. [50]

    In: NeurIPS (2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E.H., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. In: NeurIPS (2022)

    2022

  43. [51]

    In: ECCV (2024)

    Yan, C., Wang, H., Yan, S., Jiang, X., Hu, Y., Kang, G., Xie, W., Gavves, E.: Visa: Reasoning video object segmentation via large language models. In: ECCV (2024)

    2024

  44. [52]

    In: CVPR (2022)

    Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language-aware vision transformer for referring image segmentation. In: CVPR (2022)

    2022

  45. [53]

    Yao, Y., Li, L., Song, J., Chen, C., He, Z., Wang, Y., Wang, X., Gu, T., Li, J., Teng, Y., et al.: Argus inspection: do multimodal large language models possess the eye of panoptes? In: ACMMM (2025)

    2025

  46. [54]

    arXiv preprint arXiv:2506.22624 (2025)

    You, Z., Wu, Z.: Seg-r1: Segmentation can be surprisingly simple with reinforce- ment learning. arXiv preprint arXiv:2506.22624 (2025)

    2025 arXiv

  47. [55]

    In: ECCV (2016)

    Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: ECCV (2016)

    2016

  48. [56]

    arXiv preprint arXiv:2501.04001 (2025)

    Yuan, H., Li, X., Zhang, T., Sun, Y., Huang, Z., Xu, S., Ji, S., Tong, Y., Qi, L., Feng, J., Ming-Hsuan, Y.: Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos. arXiv preprint arXiv:2501.04001 (2025)

    2025 arXiv

  49. [57]

    In: AAAI (2026)

    Zhou, Q., Yang, L., Jia, Y., Gao, J., Ni, W., Wu, J., Wang, Q.: Reasoning via implicit self-supervised emergence for instruction segmentation. In: AAAI (2026)

    2026

  50. [58]

    In: AAAI (2026)

    Zhu, L., Ouyang, B., Zhang, Y., Cheng, T., Hu, R., Shen, H., Ran, L., Chen, X., Yu, L., Liu, W., Xinggang, W.: Lens: Learning to segment anything with unified reinforced reasoning. In: AAAI (2026)

    2026

  51. [59]

    In: CVPR (2025) DGSeg 19

    Zhu, M., Tian, Y., Chen, H., Zhou, C., Guo, Q., Liu, Y., Yang, M., Shen, C.: Segagent: Exploring pixel understanding capabilities in mllms by imitating human annotator trajectories. In: CVPR (2025) DGSeg 19

    2025

  52. [60]

    In: NeurIPS (2023)

    Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: NeurIPS (2023)

    2023

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.