HCNQA: Enhancing 3D VQA with Hierarchical Concentration Narrowing Supervision

Fangjing Wang; Feng Zheng; Jianuo Zhu; Qilin Huang; Shengli Zhou; Yanfu Zhang

REVIEW 4 major objections 6 minor 27 references

HCNQA: Enhancing 3D VQA with Hierarchical Concentration Narrowing Supervision

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Supervising three checkpoints of spatial narrowing makes 3D VQA models reason before answering and cuts shortcut behavior roughly in half compared with an answer-centric baseline.

desk verdict HCNQA has a genuinely new three-phase supervision idea, but the evidence for its shortcut-suppression claim is undermined by inconsistent baselines and an unvalidated pseudo-label pipeline. read the letter →

arxiv 2507.01800 v1 pith:XALJDTPY submitted 2025-07-02 cs.CV cs.MM

Shengli Zhou , Jianuo Zhu , Qilin Huang , Fangjing Wang , Yanfu Zhang , Feng Zheng This is my paper

classification cs.CVcs.MM

keywords 3Dvisualquestionansweringhierarchicalsupervisionreasoningpathwayregularizationshortcutsuppressionspatialobject-levelmasksScanQAhuman-like

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

HCNQA argues that a 3D visual question-answering model should be supervised not only on the final answer but also on three intermediate checkpoints of spatial concentration: narrowing from a whole scene to a block of interest, then to objects of interest, then to the target object. The paper claims that forcing the model to commit to these masks at each phase prevents it from learning superficial shortcuts that match question patterns to answers, and that this makes the model perform spatial reasoning before answering. If the claim holds, it offers a cheap way to make 3D VQA more reliable and more interpretable, adding only about 7.30% FLOPs. The reported results show the method beating answer-centric baselines on ScanQA and degrading about half as much under a synonym-perturbation shortcut test.

What carries the argument

The load-bearing object is the Hierarchical Supervision Module (HSM), a sequence of three lightweight MLP submodules that predict object-level masks: Blocks of Interest (BoI, a cell in a 5x5 spatial grid that contains target or anchor objects), Objects of Interest (OoI, the target plus anchor objects), and Object of Target (OoT, the object directly answering the question). Each submodule's feature is concatenated with the base scene feature and passed to the next, so the model must commit to earlier masks before computing later ones. The masks are trained with weighted binary cross-entropy to counter class imbalance, and the final OoT mask reweights object tokens before a Modular Co-Attention Network (MCAN) VQA head generates the answer.

What would settle it

Take a random sample of ScanQA questions and have independent human annotators produce BoI, OoI, and OoT masks from each question and scene; if HCNQA's errors concentrate on the questions where the heuristic masks disagree with human masks, or if retraining on human masks does not improve accuracy over the heuristic masks, then the claim that the heuristic pathway itself is the mechanism of improvement fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that supervising a general, question-independent three-phase reasoning pathway—coarse grounding (which room block matters), fine grounding (which objects matter), and inference (which object is the target)—is sufficient to keep a 3D VQA model from taking shortcuts, and that the resulting model answers more accurately. The evidence is that HCNQA reaches 27.01 EM@1 on ScanQA's Test w/ obj split, versus 25.94 for the answer-centric 3D-VisTA baseline, and that on a synonym-substitution perturbation test the model's accuracy drop is 0.24 versus 0.47, roughly half the baseline's degradation. The paper also shows in ablations that each supervised phase contributes, and that fully supervised models outperform partially supervised ones.

Load-bearing premise

The load-bearing premise is that the heuristic rules that turn ScanQA's object_ids into Blocks of Interest, Objects of Interest, and Object of Target labels describe, for every question, the same correct reasoning pathway a human would follow; if those labels are wrong for a question, the model is trained to commit to an erroneous intermediate conclusion and the error can propagate to the final answer.

Editorial extensions

If this is right

3D VQA models trained with checkpoint supervision on intermediate masks outperform answer-centric-only models on ScanQA, with EM@1 improving from 25.94 to 27.01 and gains across BLEU, ROUGE-L, Meteor, and CIDEr.
Supervised reasoning checkpoints suppress superficial shortcuts: under synonym perturbation, HCNQA's EM@1 drop is 0.24 versus 0.47 for the baseline, which is about half the degradation.
The hierarchical supervision adds only about 7.30% FLOPs, so the reasoning-pathway regularization is computationally cheap relative to its accuracy benefit.
Ablation results show that removing any of the three supervision phases lowers performance, implying that each phase's checkpoint contributes to the final result.
The method underperforms on questions with short or implicit reasoning chains, where the multi-stage mechanism adds complexity and can propagate early mispredictions, a limitation the paper explicitly discusses.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

(Editorial inference) The same three-phase narrowing idea could transfer to other 3D grounding tasks or embodied question answering, where object-level checkpoints can be generated from existing annotations with little extra cost.
(Editorial inference) Because the paper's masks come from heuristic rules on ScanQA annotations, one testable extension is to make the masks learnable or human-annotated for a subset of questions; if human masks improve accuracy further, the heuristic pathway, not the supervision mechanism, is the current bottleneck.
(Editorial inference) The shortcut-suppression result suggests intermediate checkpoint supervision could act as a general debiasing regularizer for vision-language models, for example on biased VQA splits where answer distributions correlate with question wording.
(Editorial inference) The paper's framing against underthinking in slow-thinking LLMs suggests that checkpoint supervision could be applied to longer chain-of-thought generation as a commitment mechanism, though this goes beyond what the paper tests.

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.

Desk Editor's Note

read the letter

Here's my read. HCNQA's contribution is the BoI-to-OoI-to-OoT supervision ordering, which is distinct from the flat detection/classification auxiliaries in 3D JCG, ViL3DRef, and MVT. The architecture follows naturally: each HSM stage consumes the previous stage's output, so the model is pushed to commit to a coarse area before finer objects. That is a coherent idea, and the paper does some honest things: it includes a shortcut-perturbation test and a discussion section that concedes the framework can hurt on short reasoning chains.

The soft spots are real. The baseline number for 3D-VisTA changes from 25.94 (Table 3) to 23.25 (Table 4). If these are different splits, the paper never says so, and the shortcut test compares degradation percentages across different absolute levels, which is confusing at best. The ablation in Table 2 is the clearest problem: the full four-head row (22.95 EM@1) is not the best row; the three-head row with CG/FG/IF but without VQA hits 22.98. That directly contradicts the claim that all checkpoints plus the answer head work best. One point on EM@1 is close to noise, and there are no error bars or significance tests anywhere.

The load-bearing assumption is the pseudo-label generation. Section 3.1 says the masks come from "heuristic rules on ScanQA's annotation" and then immediately notes that annotation "itself does not have a clear meaning." The rules are not specified—no pseudocode, no thresholds, no examples. The stress-test note is right that this matters more than usual because the model is forced to commit to these masks at three checkpoints; if the masks encode a wrong pathway, the supervision actively teaches a bad intermediate behavior. The paper's own Section 4.7 admits early-stage mispredictions can mislead. So the causal claim—that enforcing this pathway suppresses shortcuts—is plausible but untested. The Table 1 "annotation quality" experiment shows these masks are learnable and sometimes help, but it does not validate that they match any human reasoning process.

What's solid: the idea is positioned within the literature, the loss and architecture are described well enough to reimplement if the pseudo-label code is released, and the comparison against answer-centric methods is fair in spirit. I would not call this a flawed paradigm; I would call it a promising extension whose main evidence is not yet clean. Readers working on 3D vision-language grounding or auxiliary supervision will find the core idea useful.

For peer review: I'd send it out. The novelty and the honest limitations justify referee time, but any acceptance should be conditional on unifying the baseline numbers, fixing the ablation story, adding error bars, and—most importantly—releasing the exact pseudo-label generation rules so the pathway can be independently checked.

Referee Report

4 major / 6 minor

Summary. The paper proposes HCNQA, a 3D VQA model that augments answer-centric supervision with three hierarchical object-mask supervision objectives (BoI, OoI, OoT), intended to force the model to perform coarse-to-fine spatial grounding before answering. Masks are generated by unpublished heuristic rules applied to ScanQA's object_ids annotations. The model is built on 3D-VisTA-style encoders plus a sequential Hierarchical Supervision Module (HSM) and a MCAN VQA head. Experiments on ScanQA report EM@1 of 27.01 vs 25.94 for 3D-VisTA on Test w/ obj, and a smaller degradation under synonym perturbation (0.24 vs 0.47). Ablations and annotation-quality tables are also presented.

Significance. If the mechanism is validated, the work would make a useful contribution by showing that checkpoint supervision of intermediate spatial grounding can suppress shortcuts in 3D VQA, with only 7.30% FLOPs overhead. The paper's strengths are the clear problem motivation, the inclusion of a shortcut-perturbation evaluation, and the release of code. However, the central evidence is currently weakened by the unspecified mask-generation rules, an inconsistent baseline number between Tables 3 and 4, and an ablation in which full supervision is not the best configuration. These issues prevent the paper from establishing the claimed causal mechanism.

major comments (4)

[§3.1] The exact heuristic rules used to convert ScanQA's object_ids annotations into BoI, OoI, and OoT masks are not given (no pseudocode, no thresholds, no examples), and the text states that the annotation 'itself does not have a clear meaning.' Because the HSM is trained to commit to these masks at three sequential checkpoints, any systematic error in the masks will propagate into the final answer; the paper itself concedes in Section 4.7 that early-stage mispredictions can mislead on short reasoning chains. The authors should fully specify the rules, release the generated masks, and validate them against human judgments or an oracle; otherwise the EM@1 gain over 3D-VisTA cannot be attributed to the claimed reasoning pathway rather than to generic auxiliary supervision.
[§4.4, §4.5 (Tables 3 and 4)] The 3D-VisTA baseline is 25.94 EM@1 in Table 3 but 23.25 in Table 4. If these are different splits or evaluation settings, the shortcut-degradation comparison is not controlled; if they are the same setting, at least one number is wrong. This matters because the shortcut-suppression claim is a central piece of evidence. The authors must reconcile the numbers and report the perturbation experiment on the same split and model checkpoint used in Table 3.
[§4.3 (Table 2)] The full-supervision row (CG+FG+IF+VQA, EM@1 22.95) is not the best configuration; the three-head combination without the VQA head reaches 22.98, and the difference to CG-only (22.33) is small. This undermines the claim that all three phases and the full pipeline contribute, and no significance tests are reported. Please provide multiple runs with standard deviations or confidence intervals and a paired test, and discuss whether the differences are within noise.
[§4.2 (Table 1)] The annotation-quality verification shows that training with BoI, OoI, or OoT improves EM@1 relative to no localization supervision, but trainability is not evidence that the masks encode a human-like reasoning pathway. No agreement metric between the heuristic masks and human annotations is reported, and the γ ratio only compares improvements over object_ids. Please add direct validation, e.g., inter-annotator agreement, oracle-mask upper bounds, or per-question inspection of mask quality.

minor comments (6)

[§4.6, Fig. 4] The section heading and figure caption say 'Quantitative results' but the content is qualitative; rename to 'Qualitative results'.
[§3.3, Eq. (1)] Equation (1) is missing a normalization convention: the prefactor (c0+c1)/N and the per-term division by c0 and c1 are introduced without explaining the effective class balance; please clarify the derivation.
[§4.1] The loss weights λ_cg=0.2, λ_fg=0.3, λ_if=0.5 and grid size S=5 are chosen without sensitivity analysis; report at least a small sweep over these hyperparameters.
[References] References [24] and [25] are the same paper (Ye and Kovashka) and should be merged.
[Abstract/Code] The code link in the abstract has no commit hash or license; for reproducibility, please pin a version.
[§4.3, Table 2] Table 2's checkmark rows are ambiguous because three rows have identical '✓ ✓ ✓' symbols; label each configuration explicitly (e.g., 'CG+FG' vs 'CG+IF' vs 'FG+IF').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the intermediate BoI/OoI/OoT masks are auxiliary supervisions derived from ScanQA annotations, not re-expressions of the final answer, and the reported gains are evaluated against an independent baseline and a shortcut-perturbation test.

full rationale

The paper's core claim is that supervising intermediate masks (BoI, OoI, OoT) enforces a three-phase spatial reasoning pathway before the answer is produced. These masks are generated by heuristic rules over ScanQA's object_ids and question anchors, as stated in Section 3.1: 'we generate pseudo-labels for ground truths by applying heuristic rules on ScanQA's annotation (which itself does not have a clear meaning).' Although OoT is conceptually identical to ScanQA's object_ids (the target object), BoI and OoI are coarser object sets derived from the target and anchors, so they are not merely the final answer relabeled. The answer supervision comes from the same ScanQA benchmark, but that is standard supervised training, not circularity. The empirical evaluation compares against 3D-VisTA, which already uses object_ids supervision, and includes a synonym-perturbation shortcut test that is independent of the pseudo-label construction. No equation-level reduction, no fitted parameter renamed as a prediction, and no load-bearing self-citation chain is present; the paper cites no prior work by its own authors. The main weakness is that the heuristic pseudo-labels are not fully specified or validated, which is a reproducibility and assumption-risk concern rather than a circular-derivation concern. Therefore the derivation chain is self-contained with respect to circularity, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The central claim rests on hand-chosen hyperparameters (loss weights, grid size, attention constant, MLP depth) and on unverified assumptions that the heuristic pseudo-label pathway is rational and that intermediate supervision suppresses shortcuts. The three invented mask entities are derived from existing annotations, so they have independent testable handles.

free parameters (4)

loss weights λ_cg, λ_fg, λ_if = 0.2, 0.3, 0.5
Chosen by hand for the total HSM loss in Eq. 1; no sensitivity analysis reported.
BoI grid size S = 5
Hand-chosen for the 5x5 grid in Section 4.1.
VQA head attention constant = 1
Added to avoid ignoring objects not selected by OoT (Section 3.2).
HSM MLP depth = 4 layers
Architecture choice for the pre-HSM feature extractor; no comparison with other depths.

assumptions (5)

ad hoc to paper The three-phase concentration narrowing pathway (BoI to OoI to OoT) is a general and rational reasoning pathway for 3D VQA.
Introduced in Section 3.1 as 'mimicking humans' way', but no cognitive validation is provided; the pathway is defined by the heuristic label rules.
domain assumption Supervising intermediate checkpoints suppresses shortcuts and improves answer accuracy.
Central hypothesis of the paper; supported only by the reported experiments, not by theory or external validation.
domain assumption ScanQA object_ids annotations reliably identify the target and anchors for every question.
Used to generate pseudo-labels; if the annotation is noisy or ambiguous, the intermediate supervision labels are wrong.
domain assumption Object segmentation (ground truth or Mask3D predictions) is accurate and consistent across experiments.
Stated in Section 4.1; inconsistent segmentation would corrupt the mask supervision.
domain assumption Pretraining on Masked Language Modeling, Masked Object Modeling, and Scene-Text Matching following 3D-VisTA transfers to the HSM training.
The model inherits the 3D-VisTA pretraining recipe; the paper does not isolate the effect of pretraining on the new supervision.

invented entities (3)

Block of Interest (BoI) mask independent evidence
purpose: Intermediate supervision for coarse grounding; a grid-level mask of blocks containing target or anchors.
New construct defined from ScanQA annotations, testable through its effect on VQA performance.
Objects of Interest (OoI) mask independent evidence
purpose: Intermediate supervision for fine grounding; object-level mask of target plus anchors.
New construct defined from ScanQA annotations, testable through its effect on VQA performance.
Object of Target (OoT) mask independent evidence
purpose: Intermediate supervision for inference; target object mask used to re-weight object features in the VQA head.
New construct defined from ScanQA annotations, testable through its effect on VQA performance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HCNQA: Enhancing 3D VQA with Hierarchical Concentration Narrowing Supervision." pith.science (2026). https://pith.science/paper/XALJDTPY

@misc{pith2026250701800,
  author       = {Pith},
  title        = {Pith review of: HCNQA: Enhancing 3D VQA with Hierarchical Concentration Narrowing Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XALJDTPY}},
  note         = {Machine review of arXiv:2507.01800}
}

read the original abstract

3D Visual Question-Answering (3D VQA) is pivotal for models to perceive the physical world and perform spatial reasoning. Answer-centric supervision is a commonly used training method for 3D VQA models. Many models that utilize this strategy have achieved promising results in 3D VQA tasks. However, the answer-centric approach only supervises the final output of models and allows models to develop reasoning pathways freely. The absence of supervision on the reasoning pathway enables the potential for developing superficial shortcuts through common patterns in question-answer pairs. Moreover, although slow-thinking methods advance large language models, they suffer from underthinking. To address these issues, we propose \textbf{HCNQA}, a 3D VQA model leveraging a hierarchical concentration narrowing supervision method. By mimicking the human process of gradually focusing from a broad area to specific objects while searching for answers, our method guides the model to perform three phases of concentration narrowing through hierarchical supervision. By supervising key checkpoints on a general reasoning pathway, our method can ensure the development of a rational and effective reasoning pathway. Extensive experimental results demonstrate that our method can effectively ensure that the model develops a rational reasoning pathway and performs better. The code is available at https://github.com/JianuoZhu/HCNQA.

Figures

Figures reproduced from arXiv: 2507.01800 by the authors.

**Figure 1.** In answer-centric approaches, the absence of supervision on the reasoning pathway can lead models to adopt shortcuts, such as predicting answers before identifying relevant objects. As shown in the figure, during training, the model formed an incorrect shortcut between table and chairs, not adopting the correct reasoning pathway marked with dashed lines. This faulty correlation ultimately led to a failure during te… view at source ↗

**Figure 2.** The architecture of the HCNQA model. The Hierarchical Supervision Module (HSM) is boxed out in the model. It includes the coarse grounding, fine grounding, and inference modules. The modules in HSM generate the results of hierarchical concentration narrowing (i.e., BoI, OoI, and OoT), respectively. The information from each module also serves as a reference for the subsequent module. A VQA head is placed after the … view at source ↗

**Figure 3.** The architecture of the Hierarchical Supervision Module (HSM). The arrows in the diagram indicate the direction of data flow throughout the processing stages. HSM The Hierarchical Supervision Module (HSM) consists of 3 submodules, namely coarse grounding, fine grounding, and inference. These modules correspond to the reasoning phases in Section 3.1. HSM is designed as a sequential structure to ensure that the model… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Quantitative results. In the middle column, objects predicted to be in object_ids by the answer-centric model 3D-VisTA are marked in blue. The right column shows the predictions of our model: all masked objects belong to BoI, green and purple objects are part of OoI, …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages

[1]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answer- ing for spatial scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19129–19139 (June 2022)

work page 2022
[2]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16464–16473 (June 2022)

work page 2022
[3]

In: NeurIPS (2022)

Chen, S., Tapaswi, M., Guhur, P.L., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. In: NeurIPS (2022)

work page 2022
[4]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 26428–26438 (June 2024)

work page 2024
[5]

In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV)

Dancette, C., Cadène, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1574–1583 (October 2021)

work page 2021
[6]

DeepSeek-AI: Deepseek-r1: Incentivizing reasoning capability in llms via reinforce- ment learning (2025),https://arxiv.org/abs/2501.12948

arXiv 2025
[7]

ArXivabs/2306.02329 (2023), https://api

Delitzas, A., Parelli, M., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Multi-clip: Contrastive vision-language pre-training for ques- tion answering tasks in 3d scenes. ArXivabs/2306.02329 (2023), https://api. semanticscholar.org/CorpusID:259076122

arXiv 2023
[8]

In: North American Chapter of the Association for Computational Linguistics (2019)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)

work page 2019

Show all 27 references

[9]

NeurIPS (2023)

Hong,Y.,Zhen,H.,Chen,P.,Zheng,S.,Du,Y.,Chen,Z.,Gan,C.:3d-llm:Injecting the 3d world into large language models. NeurIPS (2023)

2023
[10]

Proceedings of Machine Learning Research, vol

Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., Li, Q., Zhu, S.C., Jia, B., Huang,S.:Anembodiedgeneralistagentin3dworld.In:Salakhutdinov,R.,Kolter, Z.,Heller,K.,Weller,A.,Oliver,N.,Scarlett,J.,Berkenkamp,F.(eds.)Proceedings of the 41st International Conference on Ma...

2024
[11]

Huang,S.,Chen,Y.,Jia,J.,Wang,L.:Multi-viewtransformerfor3dvisualground- ing.In:ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPattern Recognition (CVPR). pp. 15524–15533 (June 2022)

2022
[12]

In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR)

Jin, Z., Hayat, M., Yang, Y., Guo, Y., Lei, Y.: Context-aware alignment and mu- tual masking for 3d-language pre-training. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 10984–10994 (June 2023) 12 S. Zhou et al

2023
[13]

Longpre, S., Hou, L., Vu, T., Webson, A., Chung, H.W., Tay, Y., Zhou, D., Le, Q.V., Zoph, B., Wei, J., Roberts, A.: The flan collection: Designing data and meth- ods for effective instruction tuning (2023),https://arxiv.org/abs/2301.13688

2023 arXiv
[14]

In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=IDJx97BC38

Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: Sqa3d: Sit- uated question answering in 3d scenes. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=IDJx97BC38

2023
[15]

Proceedings of the AAAI Conference on Artificial In- telligence 38(5), 4261–4268 (Mar 2024).https://doi.org/10.1609/aaai.v38i5

Mo, W., Liu, Y.: Bridging the gap between 2d and 3d visual question answering: A fusion approach for 3d vqa. Proceedings of the AAAI Conference on Artificial In- telligence 38(5), 4261–4268 (Mar 2024).https://doi.org/10.1609/aaai.v38i5. 28222, https://ojs.aaai.org/index.php/AA...

2024 doi
[16]

https://openai.com/index/ learning-to-reason-with-llms/ (2024), accessed: 2024-07-26

OpenAI: Learning to reason with llms. https://openai.com/index/ learning-to-reason-with-llms/ (2024), accessed: 2024-07-26

2024
[17]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

Parelli, M., Delitzas, A., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Clip-guided vision-language pre-training for question answering in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. ...

2023
[18]

In: Proceedings of the 31st International Con- ference on Neural Information Processing Systems

Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: deep hierarchical feature learn- ing on point sets in a metric space. In: Proceedings of the 31st International Con- ference on Neural Information Processing Systems. p. 5105–5114. NIPS’17, Curran Associates Inc., Red Hook, N...

2017
[19]

In: Meila, M., Zhang, T

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th Int...

2021
[20]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

2016
[21]

International Confer- ence on Robotics and Automation (ICRA) (2023)

Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. International Confer- ence on Robotics and Automation (ICRA) (2023)

2023
[22]

Sourulahti, S., Janssen, C.P., Jokinen, J.P.: Modeling rational adaptation of visual search to hierarchical structures (2024),https://arxiv.org/abs/2409.08967

2024 arXiv
[23]

Wang, Y., Liu, Q., Xu, J., Liang, T., Chen, X., He, Z., Song, L., Yu, D., Li, J., Zhang, Z., Wang, R., Tu, Z., Mi, H., Yu, D.: Thoughts are all over the place: On the underthinking of o1-like llms (2025),https://arxiv.org/abs/2501.18585

2025 arXiv
[24]

Ye, K., Kovashka, A.: A case study of the shortcut effects in visual commonsense reasoning. Proceedings of the AAAI Conference on Artificial Intelligence35(4), 3181–3189 (May 2021).https://doi.org/10.1609/aaai.v35i4.16428, https:// ojs.aaai.org/index.php/AAAI/article/view/16428

2021 doi
[25]

In: Proceedings of the Thirty-Fifth AAAI Conference on Artificial In- telligence (AAAI) (February 2021)

Ye, K., Kovashka, A.: A case study of the shortcut effects in visual commonsense reasoning. In: Proceedings of the Thirty-Fifth AAAI Conference on Artificial In- telligence (AAAI) (February 2021)

2021
[26]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

2019
[27]

In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV)

Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained trans- former for 3d vision and text alignment. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV). pp. 2911–2921 (October 2023)

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answer- ing for spatial scene understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19129–19139 (June 2022)

work page 2022

[2] [2]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Cai, D., Zhao, L., Zhang, J., Sheng, L., Xu, D.: 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16464–16473 (June 2022)

work page 2022

[3] [3]

In: NeurIPS (2022)

Chen, S., Tapaswi, M., Guhur, P.L., Schmid, C., Laptev, I.: Language conditioned spatial relation reasoning for 3d object grounding. In: NeurIPS (2022)

work page 2022

[4] [4]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 26428–26438 (June 2024)

work page 2024

[5] [5]

In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV)

Dancette, C., Cadène, R., Teney, D., Cord, M.: Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1574–1583 (October 2021)

work page 2021

[6] [6]

DeepSeek-AI: Deepseek-r1: Incentivizing reasoning capability in llms via reinforce- ment learning (2025),https://arxiv.org/abs/2501.12948

arXiv 2025

[7] [7]

ArXivabs/2306.02329 (2023), https://api

Delitzas, A., Parelli, M., Hars, N., Vlassis, G., Anagnostidis, S., Bachmann, G., Hofmann, T.: Multi-clip: Contrastive vision-language pre-training for ques- tion answering tasks in 3d scenes. ArXivabs/2306.02329 (2023), https://api. semanticscholar.org/CorpusID:259076122

arXiv 2023

[8] [8]

In: North American Chapter of the Association for Computational Linguistics (2019)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)

work page 2019