pith. sign in

arxiv: 2606.04046 · v1 · pith:OJOFRNHNnew · submitted 2026-06-02 · 💻 cs.CV · cs.AI· cs.CL· cs.LG· cs.RO

Dive into the Scene: Breaking the Perceptual Bottleneck in Vision-Language Decision Making via Focus Plan Generation

Boyuan Xiao , Bohong Chen , Yumeng Li , Ji Feng , Yao-Xiang Ding , Kun Zhou This is my paper

Pith reviewed 2026-06-28 10:26 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.CLcs.LGcs.RO
keywords embodied AIvision-language modelsvisual hallucinationsscene graphfocus plan generationrobotic manipulationperceptual bottleneckvision-language-action models
0
0 comments X

The pith

SceneDiver generates focus plans through scene graphs and iterative decomposition to reduce visual hallucinations in vision-language models for embodied tasks.

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

Vision-language models and vision-language-action models face a perceptual bottleneck in robotic manipulation and navigation because they cannot reliably distinguish task-relevant objects from distractors, leading to visual hallucinations. The paper proposes SceneDiver, a coarse-to-fine focus plan generation approach that first builds a holistic scene graph for initial scene comprehension and then applies iterative cycles of recognition, understanding, and analysis to decompose tasks into simpler sub-problems. A lightweight adapter distills this deliberate focus ability into reactive VLAs. If the method works, it would improve performance on standard embodied AI benchmarks while maintaining computational efficiency for fast execution tasks.

Core claim

SceneDiver constructs a holistic scene graph to establish initial comprehension and then progressively decomposes the task into simpler sub-problems through an iterative cycle of recognition, understanding, and analysis, enabling accurate identification and focus on critical objects while filtering out irrelevant ones for both VLMs and VLAs.

What carries the argument

The coarse-to-fine focus plan generation process that begins with a holistic scene graph and proceeds via iterative recognition-understanding-analysis cycles.

If this is right

  • Substantially reduces visual hallucinations for both VLMs and VLAs on standard embodied AI benchmarks.
  • Preserves computational efficiency in tasks requiring fast execution.
  • Leverages VLMs' long-term planning strengths for focus while enabling reactive control in VLAs via the adapter.

Where Pith is reading between the lines

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

  • The scene graph representation could serve as a reusable intermediate structure for multiple sequential tasks in the same environment.
  • The distillation adapter might allow similar focus improvements in other model families that lack built-in planning capacity.
  • Iterative decomposition cycles could be tested in non-robotic vision-language settings such as image captioning with distractors.

Load-bearing premise

The assumption that a holistic scene graph plus iterative recognition-understanding-analysis cycles will produce effective deep scene understanding and focus without introducing new errors or excessive overhead.

What would settle it

A direct comparison on an embodied AI benchmark where SceneDiver produces no reduction in visual hallucination rates relative to baseline VLMs or VLAs, or where the added steps increase latency beyond the level needed for reactive control.

Figures

Figures reproduced from arXiv: 2606.04046 by Bohong Chen, Boyuan Xiao, Ji Feng, Kun Zhou, Yao-Xiang Ding, Yumeng Li.

Figure 1
Figure 1. Figure 1: Failures in VLM decision making during direct text￾image queries. Left: When queried about the number of green objects, the VLM’s attention fails to capture the target objects, instead allocating excessive focus to irrelevant background ele￾ments. Right: When asked for the color of the object grasped by the robotic arm, the model exhibits attention mismatching; the focus incorrectly shifts to the yellow bl… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SceneDiver: From the input image we build a scene graph, perform graph reasoning to decompose the complex global scene into a series of simpler local sub-scenes corresponding to individual nodes in the graph (Stage 1), autonomously explore each local sub-scene using naturally designed exploration strategies to identify task-relevant objects (Stage 2), and use the resulting focus to modify the i… view at source ↗
Figure 3
Figure 3. Figure 3: Distilling SceneDiver into a lightweight adapter to transfer the deliberate focus ability to a VLA for reactive control. Slot Attention learns the structured representations of scene graphs, while the mask prediction module learns the two-stage reasoning process. tion objectives: a Structure Loss supervising Slot Attention to learn structured reasoning-related representations, and a Mask Loss supervising t… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization of the SceneDiver execution process for robot manipulation. The focus of each step is highlighted with a red bounding box (for visualization purposes only; the box is invisible to the VLM). The workflow of SceneDiver is illustrated using Step 4 as a representative example. minimal computational overhead. Finally, we perform com￾prehensive ablation studies to validate the necessity… view at source ↗
Figure 5
Figure 5. Figure 5: Fine-grained Verification and Exploration results of three different agentic behaviors In the fine-grained exploration stage, we perform a zoom-in operation on each candidate node to construct a local scene graph, recovering details that may be invisible in the global view. Based on the local observations, the VLM adaptively executes one of three exploration primitives, as illustrated in [PITH_FULL_IMAGE:… view at source ↗
Figure 6
Figure 6. Figure 6: Instruction: navigate to the Pot in the room and be as close as possible to it [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Instruction: navigate to the DeskLamp in the room and be as close as possible to it. Common Sense. Compared to the base task, this task involves a more complex instruction, requiring stronger reasoning capabilities from the MLLMs [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Instruction: I need a soft cushion to support my head while sleeping. Can you navigate to that object and stay close? [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Instruction: I’d like to view a decorative sculpture representing a figure or person. Can you navigate to that object and stay close? 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Instruction: The sound of someone walking upstairs adds a subtle rhythm to the quiet morning. There’s a folded towel on the counter, and the air smells faintly of butter. Could you navigate to the toaster for me? It’s a peaceful start to the day [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Instruction: The rhythmic ticking of the kitchen clock blends with the occasional drip from the faucet. There’s a small pile of onions on the table, freshly chopped. Please move towards the stove burner for me. The kitchen has a comforting hum to it. Visual Appearance Task. The instruction in this task describes the visual appearance of the target object, requiring the MLLMs to possess strong visual compr… view at source ↗
Figure 12
Figure 12. Figure 12: Instruction: Approach the tall green container with a smooth texture [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Instruction: Move closer to the small round object with a green surface and a cylindrical shape. Libero-plus. In this section, we present the qualitative results of the SceneDiver Adapter on the LIBERO benchmark, illustrating the masking outcomes across various scenarios. Our method does not merely memorize objects from the training data; instead, it inherently learns to identify task-relevant entities an… view at source ↗
Figure 14
Figure 14. Figure 14: Instruction: Put both the alphabet soup and the tomato sauce in the basket. Primary View Wrist View [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Instruction: Put the yellow and white mug in the microwave and close it. Primary View Wrist View [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Instruction: Pick up the book and place it in the back compartment of the caddy. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Instruction: Put the white mug on the left plate and put the yellow and white mug on the right plate. Primary View Wrist View [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Instruction: Put the white mug on the plate and put the chocolate pudding to the right of the plate. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_18.png] view at source ↗
read the original abstract

In embodied vision-language decision making tasks such as robotic manipulation and navigation, Vision-Language and Vision-Language-Action Models (VLMs & VLAs) are powerful tools with different benefits: VLMs are better at long-term planning, while VLAs are better at reactive control. However, their performance is limited by the same perceptual bottleneck: visual hallucinations arise due to the models' inability to distinguish task-relevant objects from distractors. In principle, accurate identification and focus on critical objects while filtering out irrelevant ones is the key to break this limitation. A straightforward solution is one-step focus: directly attending to essential objects. However, this approach proves ineffective because effective focus inherently requires deep scene understanding. To this end, we propose SceneDiver, a coarse-to-fine focus plan generation method for VLMs leveraging their long-term planning abilities, that first constructs a holistic scene graph to establish initial comprehension, then progressively decomposes the task into simpler sub-problems through an iterative cycle of recognition, understanding, and analysis. To enable reactive control, we also design a lightweight adapter for distilling the deliberate focus ability into VLAs. Evaluations on standard embodied AI benchmarks confirm that our method substantially reduces visual hallucinations for both VLMs and VLAs, while preserving computational efficiency in tasks requiring fast execution. Our code and data are released at: https://future-item.github.io/SceneDiver.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper introduces SceneDiver, a coarse-to-fine focus plan generation method for VLMs in embodied vision-language decision making. It first builds a holistic scene graph for initial scene comprehension, then applies iterative cycles of recognition-understanding-analysis to decompose tasks and focus on relevant objects, thereby reducing visual hallucinations. A lightweight adapter distills the focus capability into VLAs for reactive control. The central claim is that this approach substantially reduces hallucinations on standard embodied AI benchmarks while preserving computational efficiency for fast-execution tasks; code and data are released.

Significance. If the empirical claims hold with adequate controls, the work could meaningfully improve reliability of VLMs and VLAs in robotic manipulation and navigation by addressing the perceptual bottleneck. The dual treatment of long-horizon planning (VLMs) and reactive control (VLAs), combined with the public release of code and data, strengthens reproducibility and potential adoption.

major comments (2)
  1. [§4] §4 Experiments: the abstract and method overview assert 'substantial reductions' in visual hallucinations, yet the provided text supplies no concrete baseline comparisons, metrics (e.g., hallucination rate, success rate), error bars, or statistical tests; without these details the central empirical claim cannot be evaluated.
  2. [§3.2] §3.2 Iterative Cycle: the claim that the recognition-understanding-analysis loop produces effective deep scene understanding without introducing new errors or excessive overhead is load-bearing for both the hallucination-reduction and efficiency assertions, but no ablation or error-propagation analysis is described to bound this risk.
minor comments (3)
  1. [Abstract] Abstract: name the specific embodied AI benchmarks used rather than referring generically to 'standard' ones.
  2. [§3.1] Notation: the distinction between 'holistic scene graph' and subsequent 'focus plan' is introduced without a formal definition or diagram; a small illustrative figure would improve clarity.
  3. [§2] Related work: add explicit comparison to prior scene-graph and attention-focusing methods in VLMs to clarify incremental contribution.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point-by-point below, indicating where revisions will be made to strengthen the paper.

read point-by-point responses

  1. Referee: [§4] §4 Experiments: the abstract and method overview assert 'substantial reductions' in visual hallucinations, yet the provided text supplies no concrete baseline comparisons, metrics (e.g., hallucination rate, success rate), error bars, or statistical tests; without these details the central empirical claim cannot be evaluated.

    Authors: We agree that the abstract and method overview would benefit from more immediate quantitative grounding. The experiments section (§4) reports results on standard embodied AI benchmarks with baseline comparisons, including task success rates and hallucination metrics, along with error bars from repeated trials. To address the concern directly, we will revise the manuscript to add a summary of the key numerical improvements (e.g., relative reductions in hallucination rates and gains in success rate) to the abstract and include explicit statistical test results in §4. revision: yes

  2. Referee: [§3.2] §3.2 Iterative Cycle: the claim that the recognition-understanding-analysis loop produces effective deep scene understanding without introducing new errors or excessive overhead is load-bearing for both the hallucination-reduction and efficiency assertions, but no ablation or error-propagation analysis is described to bound this risk.

    Authors: This is a fair observation regarding the load-bearing nature of the iterative cycle. We will add a dedicated ablation study in the revised §4 that compares the full recognition-understanding-analysis loop against ablated variants (e.g., single-pass or non-iterative focus). The analysis will report effects on hallucination rates, task success, runtime overhead, and any observed error propagation, thereby bounding the risks and supporting the efficiency claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes a proposed method (SceneDiver) that constructs a holistic scene graph then applies iterative recognition-understanding-analysis cycles plus an adapter for distillation; the central claims rest on asserted empirical reductions in hallucinations on standard embodied benchmarks. No equations, fitted parameters, self-citations used as load-bearing uniqueness theorems, or any derivation steps that reduce by construction to the inputs appear in the abstract or method sketch. The approach is therefore self-contained against external benchmarks with no circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, no explicit free parameters, axioms, or invented entities are described; the method relies on standard scene graph and planning concepts from prior VLM literature.

pith-pipeline@v0.9.1-grok · 5806 in / 1138 out tokens · 21820 ms · 2026-06-28T10:26:25.770249+00:00 · methodology

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

70 extracted references · 18 canonical work pages · 5 internal anchors

  1. [1]

    European conference on computer vision , pages=

    Spice: Semantic propositional image caption evaluation , author=. European conference on computer vision , pages=. 2016 , organization=

    2016

  2. [2]

    Terra: Hierarchical Terrain-Aware 3D Scene Graph for Task-Agnostic Outdoor Mapping

    Terra: Hierarchical Terrain-Aware 3D Scene Graph for Task-Agnostic Outdoor Mapping , author=. arXiv preprint arXiv:2509.19579 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Generating Actionable Robot Knowledge Bases by Combining 3D Scene Graphs with Robot Ontologies , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  4. [4]

    Proceedings of the beyond vision and language: integrating real-world knowledge (LANTERN) , pages=

    On the role of scene graphs in image captioning , author=. Proceedings of the beyond vision and language: integrating real-world knowledge (LANTERN) , pages=

  5. [5]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Auto-encoding scene graphs for image captioning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  6. [6]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Gqa: A new dataset for real-world visual reasoning and compositional question answering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  7. [7]

    International journal of computer vision , volume=

    Visual genome: Connecting language and vision using crowdsourced dense image annotations , author=. International journal of computer vision , volume=. 2017 , publisher=

    2017

  8. [8]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Image retrieval using scene graphs , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  9. [9]

    2023 , eprint=

    SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models , author=. 2023 , eprint=

    2023

  10. [10]

    2024 , eprint=

    LLaVA-OneVision: Easy Visual Task Transfer , author=. 2024 , eprint=

    2024

  11. [11]

    Naval research logistics quarterly , volume=

    The Hungarian method for the assignment problem , author=. Naval research logistics quarterly , volume=. 1955 , publisher=

    1955

  12. [12]

    arXiv preprint arXiv:2505.04769 , year=

    Vision-language-action models: Concepts, progress, applications and challenges , author=. arXiv preprint arXiv:2505.04769 , year=

    work page arXiv

  13. [13]

    Sensors , VOLUME =

    Das, Murat and Hussain, Zawar and Nawaz, Muhammad , TITLE =. Sensors , VOLUME =. 2026 , NUMBER =

    2026

  14. [14]

    2023 , eprint=

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control , author=. 2023 , eprint=

    2023

  15. [15]

    2023 , eprint=

    PaLM-E: An Embodied Multimodal Language Model , author=. 2023 , eprint=

    2023

  16. [16]

    2022 , eprint=

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances , author=. 2022 , eprint=

    2022

  17. [17]

    2023 , eprint=

    Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding , author=. 2023 , eprint=

    2023

  18. [18]

    2024 , eprint=

    Monkey: Image Resolution and Text Label Are Important Things for Large Multi-modal Models , author=. 2024 , eprint=

    2024

  19. [19]

    2023 , eprint=

    Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V , author=. 2023 , eprint=

    2023

  20. [20]

    2026 , eprint=

    _0 : A Vision-Language-Action Flow Model for General Robot Control , author=. 2026 , eprint=

    2026

  21. [21]

    2024 , eprint=

    Octo: An Open-Source Generalist Robot Policy , author=. 2024 , eprint=

    2024

  22. [22]

    2025 , eprint=

    LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models , author=. 2025 , eprint=

    2025

  23. [23]

    2020 , eprint=

    Object-Centric Learning with Slot Attention , author=. 2020 , eprint=

    2020

  24. [24]

    2025 , eprint=

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success , author=. 2025 , eprint=

    2025

  25. [25]

    Imagine while Reasoning in Space: Multimodal Visualization-of-Thought

    Imagine while reasoning in space: Multimodal visualization-of-thought , author=. arXiv preprint arXiv:2501.07542 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2505.11409 (2025) 28 Z

    Visual Planning: Let's Think Only with Images , author=. arXiv preprint arXiv:2505.11409 , year=

    work page arXiv

  27. [27]

    GoT-R1: Unleashing Reasoning Capability of MLLM for Visual Generation with Reinforcement Learning

    Got-r1: Unleashing reasoning capability of mllm for visual generation with reinforcement learning , author=. arXiv preprint arXiv:2505.17022 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2501.05452 , year=

    Refocus: Visual editing as a chain of thought for structured image understanding , author=. arXiv preprint arXiv:2501.05452 , year=

    work page arXiv

  29. [29]

    arXiv preprint arXiv:2505.00684 , year=

    Visual test-time scaling for gui agent grounding , author=. arXiv preprint arXiv:2505.00684 , year=

    work page arXiv

  30. [30]

    arXiv preprint arXiv:2507.00008 , year=

    DiMo-GUI: Advancing Test-time Scaling in GUI Grounding via Modality-Aware Visual Reasoning , author=. arXiv preprint arXiv:2507.00008 , year=

    work page arXiv

  31. [31]

    arXiv preprint arXiv:2411.13591 , year=

    Improved gui grounding via iterative narrowing , author=. arXiv preprint arXiv:2411.13591 , year=

    work page arXiv

  32. [32]

    2024 , eprint=

    Expanding Scene Graph Boundaries: Fully Open-vocabulary Scene Graph Generation via Visual-Concept Alignment and Retention , author=. 2024 , eprint=

    2024

  33. [33]

    2025 , eprint=

    EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents , author=. 2025 , eprint=

    2025

  34. [34]

    2025 , eprint=

    Reflective Planning: Vision-Language Models for Multi-Stage Long-Horizon Robotic Manipulation , author=. 2025 , eprint=

    2025

  35. [35]

    Advances in neural information processing systems , volume=

    Fine-tuning large vision-language models as decision-making agents via reinforcement learning , author=. Advances in neural information processing systems , volume=

  36. [36]

    arXiv preprint arXiv:2402.14683 , year=

    Visual hallucinations of multi-modal large language models , author=. arXiv preprint arXiv:2402.14683 , year=

    work page arXiv

  37. [37]

    arXiv preprint arXiv:2411.18142 , year=

    Enhancing Visual Reasoning with Autonomous Imagination in Multimodal Large Language Models , author=. arXiv preprint arXiv:2411.18142 , year=

    work page arXiv

  38. [38]

    2024 , eprint=

    Multi-Object Hallucination in Vision-Language Models , author=. 2024 , eprint=

    2024

  39. [39]

    A Survey on Vision-Language-Action Models for Embodied AI

    A survey on vision-language-action models for embodied ai , author=. arXiv preprint arXiv:2405.14093 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  40. [40]

    Conference on robot learning , pages=

    Do as i can, not as i say: Grounding language in robotic affordances , author=. Conference on robot learning , pages=. 2023 , organization=

    2023

  41. [41]

    International conference on machine learning , pages=

    Language models as zero-shot planners: Extracting actionable knowledge for embodied agents , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  42. [42]

    Conference on Robot Learning , pages=

    Bc-z: Zero-shot task generalization with robotic imitation learning , author=. Conference on Robot Learning , pages=. 2022 , organization=

    2022

  43. [43]

    arXiv preprint arXiv:2407.20179 , year=

    Theia: Distilling diverse vision foundation models for robot learning , author=. arXiv preprint arXiv:2407.20179 , year=

    work page arXiv

  44. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Self-supervised learning from images with a joint-embedding predictive architecture , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  45. [45]

    arXiv preprint arXiv:2302.12766 , year=

    Language-driven representation learning for robotics , author=. arXiv preprint arXiv:2302.12766 , year=

    work page arXiv

  46. [46]

    2019 , eprint=

    Object Hallucination in Image Captioning , author=. 2019 , eprint=

    2019

  47. [47]

    2023 , eprint=

    Plausible May Not Be Faithful: Probing Object Hallucination in Vision-Language Pre-training , author=. 2023 , eprint=

    2023

  48. [48]

    2024 , eprint=

    HallE-Control: Controlling Object Hallucination in Large Multimodal Models , author=. 2024 , eprint=

    2024

  49. [49]

    2023 , eprint=

    Ferret: Refer and Ground Anything Anywhere at Any Granularity , author=. 2023 , eprint=

    2023

  50. [50]

    2024 , eprint=

    GROUNDHOG: Grounding Large Language Models to Holistic Segmentation , author=. 2024 , eprint=

    2024

  51. [51]

    2023 , eprint=

    Aligning Large Multimodal Models with Factually Augmented RLHF , author=. 2023 , eprint=

    2023

  52. [52]

    2024 , eprint=

    Detecting and Preventing Hallucinations in Large Vision Language Models , author=. 2024 , eprint=

    2024

  53. [53]

    2024 , eprint=

    OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation , author=. 2024 , eprint=

    2024

  54. [54]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  55. [55]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  56. [56]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  57. [57]

    arXiv preprint arXiv:2501.02189 , year=

    A survey of state of the art large vision language models: Alignment, benchmark, evaluations and challenges , author=. arXiv preprint arXiv:2501.02189 , year=

    work page arXiv

  58. [58]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    GEOBench-VLM: Benchmarking Vision-Language Models for Geospatial Tasks , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  59. [59]

    arXiv preprint arXiv:2403.13164 , year=

    VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning , author=. arXiv preprint arXiv:2403.13164 , year=

    work page arXiv

  60. [60]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Spatialvlm: Endowing vision-language models with spatial reasoning capabilities , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  61. [61]

    Visual Instruction Tuning , author=

  62. [62]

    and Chai, Joyce , title =

    Chen, Xuweiyi and Ma, Ziqiao and Zhang, Xuejun and Xu, Sihan and Qian, Shengyi and Yang, Jianing and Fouhey, David F. and Chai, Joyce , title =. 2025 , isbn =

    2025

  63. [63]

    arXiv preprint arXiv:2502.01969 , year=

    Mitigating object hallucinations in large vision-language models via attention calibration , author=. arXiv preprint arXiv:2502.01969 , year=

    work page arXiv

  64. [64]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  65. [65]

    2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

    MuJoCo: A physics engine for model-based control , author=. 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=. 2012 , organization=

    2012

  66. [66]

    Qwen2.5-VL Technical Report

    Qwen2.5-VL Technical Report , author=. arXiv preprint arXiv:2502.13923 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  67. [67]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  68. [68]

    2025 , eprint=

    Seed1.5-VL Technical Report , author=. 2025 , eprint=

    2025

  69. [69]

    2025 , eprint=

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , author=. 2025 , eprint=

    2025

  70. [70]

    2024 , url =

    OpenAI , title =. 2024 , url =

    2024