Pith. sign in

REVIEW 3 major objections 3 minor 54 references

Roomer: Reflective Object-Grounded Model Editing and Repair for 3D Indoor Layout Synthesis

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Roomer claims that residual local violations in generated indoor layouts can be repaired by verification-gated local edits that preserve already-valid regions.

desk verdict A genuinely careful empirical systems paper on verification-gated local repair for 3D indoor layouts, with one real weakness: the practical-usability metric is also the repair objective, and the professional validation doesn't cover Roomer's own outputs. read the letter →

arxiv 2608.01973 v1 pith:7264FBIP submitted 2026-08-03 cs.RO cs.CV

Lingwei Dang , Ziyan Qiu , Jiajia Cheng , Shishuo Shang , Zhenhao Zhang , Yufei Zhu , Qingxin Xiao , Pan Liu
show 4 more authors
Shenghui Huang Yun Hao Juntong Li Qingyao Wu
This is my paper · ORCID
classification cs.ROcs.CV
keywords 3Dindoorlayoutsynthesislocalrepairobject-groundedattributionverification-gatedcommitcontrolled-corruptiondatasetvision-languageplannerpracticalspatialusabilityevaluation
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

Roomer's central claim is that the residual failures left by indoor-layout generators are sparse, local, and attributable to a small set of furniture objects, so they can be repaired by targeted local edits instead of full-scene regeneration. The paper presents a reflective repair loop: measured violations are bound to implicated objects, a vision-language planner proposes a structured local edit, a deterministic solver generates verified candidates, and a patch is committed only if full-scene re-verification shows the target resolved, no new hard violations, protected relations preserved, and overall residual reduced. Training on a controlled-corruption dataset of 67,550 pairs and evaluating with a new protocol that adds practical-usability rules, the paper reports that Roomer lowers out-of-bounds and collision rates and improves usability on its own layouts and on outputs of four external generators. If true, this makes post-generation local repair a reliable and generator-agnostic complement to layout synthesis.

What carries the argument

The carrying mechanism is the Roomer loop: RoState (an object-addressable canonical layout), RoReview (instance-grounded violation evidence that attributes each measured failure to concrete objects and roles), and a schema-constrained StatePatch whose action, target, and parameter seed are proposed by the planner and instantiated by a deterministic solver. The load-bearing identity is the verification gate: a candidate patch is accepted only if it resolves the target issue, introduces no new hard-violation keys, preserves all protected satisfied relations, keeps structural validity, and strictly decreases the family-balanced residual. This gate is what converts the planner's suggestion into a safe commit and supplies the 'preserve valid regions' part of the claim.

What would settle it

Take a held-out set of Roomer-repaired outputs and unrepaired baselines, have fresh interior-design experts rate them blind, and check whether approval rises with Practical on Roomer's own outputs; if it does not, the usability claim fails. Alternatively, add a sixth common usability rule family such as kitchen work-triangle clearance or bath-fixture spacing to the evaluator and rerun the repair loop; if the added rule reverses the before/after improvement, the five frozen rules missed a dominant failure class and the claim that Roomer improves practical usability generically fails.

Watch

Extended reading notes

Core claim

The paper's discovery claim is that layout repair is better modeled as verification-gated local state repair than as regeneration: most violations are caused by a few objects, and a deterministic full-scene check can decide which local edit is safe. Roomer operationalizes this with RoState, RoReview, and StatePatch: the first is an object-addressable layout encoding; the second converts each measured rule failure into a tuple of violation type, implicated entities, relational roles, and geometric measurements; the third is the structured edit (MOVE, ROTATE, SCALE, INSERT, DELETE, REPLACE) proposed by a geometry-conditioned vision-language planner. The deterministic solver instantiates a finite candidate set and commits the first patch passing the five-clause verification gate. On a frozen cohort of 1,100 held-out rooms, the method improves hard validity, target resolution, strict safe repair, and non-target preservation, while raising the Practical usability score from 72.50% to 82.98% and transferring to four external generators.

Load-bearing premise

The entire practical-usability conclusion rests on the five author-defined Practical rule families (bedside clearance, dining-table clearance, living-room functional angle, door-swing proxy, and walkable connectivity) being a fair stand-in for what makes a layout usable, and the expert validation that anchors these rules was run on baseline layouts rather than on Roomer's repaired outputs.

Editorial extensions

If this is right

  • Indoor layout generators no longer need to be globally perfect; a verification-gated post-generation repair stage can absorb their residual errors.
  • Practical usability should be measured separately from distributional quality and collision or out-of-bounds checks, because the three can move independently.
  • Deterministic candidate search plus full-scene verification can compensate for imperfect vision-language proposals, making repair robust to planner noise.
  • Roomer is generator-agnostic: any layout that can be converted into RoState can be run through the same repair loop, as demonstrated on four external generators.
  • Because rejected candidates are rolled back, repair is safe by construction: a failed edit leaves the committed layout unchanged.

Reading between the lines

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

  • The verification-gated local-repair pattern is not tied to residential furniture; it could transfer to other structured generation domains where errors are sparse and measurable, such as circuit-board placement or warehouse layout.
  • Roomer's ceiling is set by its rule library and by the action–issue coverage of its controlled-corruption data; adding a new usability rule family after training would likely require new paired supervision because the planner learns specific action–issue mappings.
  • The professional validation anchors the Practical metric on baseline layouts rather than on Roomer's own repaired outputs, so a direct blind expert study on Roomer-repaired scenes is the natural missing test of the usability claim.
  • A stricter test of the framework would be to treat the five Practical rule families as an open, versioned library and check whether repair still improves a newly added rule family without retraining the planner.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes Roomer, a post-generation repair module for 3D indoor layouts. It converts a complete layout into an object-addressable RoState, detects rule-based violations, constructs an object-grounded RoReview, and uses a geometry-conditioned VLM planner to propose a StatePatch action. A deterministic solver instantiates candidate edits, and each candidate is committed only if a full-scene verification gate confirms target resolution, no new hard violations, preservation of protected relations, structural validity, and global decrease of a family-balanced residual. The planner is trained on Roomer-CC, a controlled-corruption dataset built from 3D-FRONT layouts. The paper also introduces Roomer-Eval, which adds five rule families for practical spatial usability to standard distributional and physical metrics. Experiments on a frozen 1,100-room cohort and on outputs of four external generators report consistent gains in physical validity and Practical, with ablations showing the value of RoReview, geometry conditioning, deterministic search, and verification-gated commit.

Significance. Roomer addresses a real and under-served problem: generated indoor layouts often contain local, object-level violations that full-scene regeneration disrupts. The empirical design is a clear strength: the frozen common-1100 cohort, density-controlled baseline selection that avoids evaluation metrics, commit-prefix pinning and seed records for baselines, byte-identical initialization for ablations, and rollback semantics are all described at a level that supports the internal claims. If the practical-usability claim can be externally anchored, Roomer would be a useful generator-agnostic post-processing module. However, the current manuscript does not yet provide that external anchor, and the main quantitative tables lack uncertainty estimates. No code, data, or checkpoints are released, which limits reproducibility.

major comments (3)
  1. [Verification gate (Eq. 3) and Roomer-Eval (Eq. 4)] The abstract and Experiments section claim that Roomer improves practical usability, but the Practical metric used to measure usability is also the objective that the repair loop optimizes. The verification gate's aggregate residual V in Eq. (3) is family-balanced over hard, content, relational, and practical rules (Appendix F, Eq. (A21)), and the practical families are exactly the five rule families in Roomer-Eval (Appendix J). Consequently the 10.48-point Practical gain in Table 1 and the before/after gains in Table 2 partly measure the system's success at optimizing the evaluation instrument. The Conclusions concede that the system is limited to its predefined residential rule set. To support the usability claim, the paper needs an external anchor that is not used in repair: for example, professional judgments on Roomer-repaired outputs, or a held-out rule family never seen by the verifier.
  2. [Appendix M, Professional Validation] The professional validation samples 90 layouts from baseline generators only and is disjoint from the Roomer repair cohort. It establishes a monotonic relationship between Practical strata and professional approval on the baseline distribution, but it does not certify that Roomer's repaired outputs receive comparable approval. A repaired layout may reach High Practical by local edits that satisfy the frozen rules while displacing a failure to an aspect outside the five families, such as ergonomics or aesthetics. The sentence in the Experiments section — 'the higher Practical scores achieved by Roomer reflect improvements that are aligned with professional usability judgments' — is therefore not directly supported. I request a professional validation set that includes Roomer-repaired outputs stratified by Practical level, evaluated with the same blinding and the same Cochran–Armitage test.
  3. [Tables 1–3; Appendix I] Main quantitative claims are reported as single point estimates without confidence intervals or standard errors for OOB, COL, Practical, Target Resolution, and Strict Safe Repair. For example, Table 2 reports DiffuScene-RS Practical improving from 45.09% to 47.73%, and Table 3 reports several ablation differences of two to five points; without uncertainty estimates it is impossible to tell whether these differences are meaningful. Appendix I reports bootstrap variance for FID and KID only, and the appendix states that OOB surface sampling has no fixed seed, tying the reported values to a single execution. Please provide bootstrap confidence intervals or standard errors for all central metrics in the main tables, and report variance over repeated OOB evaluations.
minor comments (3)
  1. [Tables and typesetting] Several tables have formatting problems in the supplied text: Table A9 shows concatenated numerical values such as '57.425.1021.73', and the caption of Table A9 and Table 1 use 'oncommon-1100' without a space. Please ensure the camera-ready version renders all columns and numbers cleanly.
  2. [Reproducibility artifacts] No code, checkpoints, or data are released. Given the complexity of Roomer-CC and the Roomer-Eval evaluator, an artifact release would materially support reproducibility and would help readers verify the unusual care taken in the experimental protocol.
  3. [Appendix M, inter-rater reliability] Krippendorff's alpha is reported as 0.65 with confidence interval [0.54, 0.75], which is moderate agreement. Please state whether the Cochran–Armitage result is robust when any single evaluator is excluded, and justify the 'at least seven Yes votes' threshold.

Circularity Check

1 steps flagged · score 4.0 of 10

Practical-usability gain is the repair objective: Roomer's verification gate V and Roomer-Eval's Practical metric share the same five rule families, so the reported usability improvement is partially self-fulfilling.

  1. self definitional [Main paper: Eq. (3) verification gate and Eq. (4) Practical definition; Roomer-Eval protocol; Appendix F Eq. (A21); Appendix M professional validation.]
    "Let H be the stable hard-violation key set, Zt keep the protected satisfied relations, C the structural-validity predicate, and V the family-balanced residual over hard, content, relational, and practical repair rules. ... Although the same frozen definitions guide Roomer during repair, final scores are recomputed from each method’s complete output without reusing repair-time detections or decisions."

    In Eq. (3), a candidate is committed only if V_t^p ≤ V_t − ε_V, where V is the family-balanced residual over hard, content, relational, and practical repair rules. The same practical rule families define the Practical metric in Eq. (4), and the paper explicitly concedes that 'the same frozen definitions guide Roomer during repair.' Consequently, the Practical improvements reported in Tables 1–2 are the quantity the solver is optimizing, not an independent measurement of usability. Appendix M's professional validation samples only baseline layouts, not Roomer's repaired outputs, so it cannot independently certify that optimizing these rules yields professionally usable rooms.

full rationale

Roomer's derivation chain is otherwise self-contained and independently grounded. The planner is trained on Roomer-CC with room-level held-out splits; common-1100 is excluded from training and model selection; external-generator outputs are frozen; and the deterministic solver and verification gate are specified with explicit action domains and fallback sequences. The physical-validity metrics OOB and COL are assembled-mesh checks explicitly stated to be independent of the two-dimensional geometric detectors used to construct RoReview and verify repair candidates, so the physical-validity claim is not circular. There is no load-bearing self-citation and no imported uniqueness theorem; the only overlap between objective and measurement is the five Practical rule families that appear both in the verification residual V (Eq. 3, Appendix F Eq. A21) and in the Practical metric (Eq. 4). The paper's own sentence 'the same frozen definitions guide Roomer during repair' concedes the overlap. The professional validation in Appendix M anchors Practical to human judgment only on 90 baseline layouts, explicitly disjoint from the Roomer repair cohort, so it does not close the circularity for repaired outputs. This yields partial circularity with score 4: the practical-usability prediction is partly the optimization target, while physical validity and distributional-quality results remain independent.

Assumptions & free parameters 6 free parameters · 6 assumptions · 5 invented entities

All learned weights are fitted to Roomer-CC, but the load-bearing hand-set quantities are the usability thresholds and verification tolerances listed above. The axioms are domain assumptions about the validity of 3D-FRONT, the Practical rule families, the verifier, and the professional validation; these are reasonable but should be read as premises, not derived facts.

free parameters (6)
  • Bedside clearance ratio threshold = c_usable / w_bed >= 0.30
    Hand-set in Roomer-Eval (Eq. A5) and used both as a repair target and evaluation rule; changing it directly changes Practical and the verification gate.
  • Dining-table clearance threshold = min(c1, c2) >= 0.60 m
    Hand-set in Appendix J; defines the dining clearance rule used as a repair target and evaluation metric.
  • Living functional angle threshold = angle >= 135 degrees
    Hand-set in Appendix J; defines the living functional organization rule.
  • Walkable connectivity passage width = dilation 0.30 m, i.e., 0.60 m passage; component threshold 20 pixels
    Hand-set in Appendix J; determines walkable connectivity satisfaction and influences repair verification.
  • Door swing proxy rule = 90-degree sectors with radius equal to door width; any footprint intersection or contact fails
    Hand-set in Appendix J; determines the door obstruction rule.
  • Repair residual descent epsilon = epsilon_V = 1e-6
    Hand-set in Appendix F; part of the verification gate V_tp <= V_t - epsilon_V.
assumptions (6)
  • domain assumption Valid 3D-FRONT layouts that satisfy target room specifications are reliable ground-truth starting points for repair supervision.
    Roomer-CC corrupts these layouts and assumes the inverse patch restores true validity; any latent invalidity in 3D-FRONT propagates into training targets. Stated in the Roomer-CC section.
  • domain assumption The five Practical rule families are a sufficient operationalization of practical spatial usability.
    Roomer-Eval, Appendix J; the professional validation supports the metric's alignment with human judgment, but only on 90 baseline layouts.
  • domain assumption The deterministic verifier correctly enumerates hard violations and protected satisfied relations.
    Eq. (3) and Appendix F; if the verifier misses a violation type or misassigns dependencies, the no-new-violation guarantee would not hold.
  • domain assumption Professional approval by ten interior-design evaluators with a seven-of-ten threshold is a valid ground-truth signal.
    Appendix M; this grounds Practical but is itself a small-sample human judgment.
  • domain assumption Stable object references preserve identity across edits, so RoReview attribution remains correct.
    Appendix A; if reference stability fails, the planner could edit the wrong object.
  • domain assumption Roomer-CC's inverse-corruption supervision teaches repair behavior that generalizes to outputs of unrelated generators.
    The transfer experiments support this empirically for four generators, but the training signal still comes only from corruptions of 3D-FRONT rooms.
invented entities (5)
  • RoState
    purpose: Canonical object-addressable layout representation for planner and verifier.
    Internal representation; no external falsifiable handle.
  • RoReview
    purpose: Binds measured violations to implicated objects and roles.
    Internal diagnostic structure; outputs are only tested indirectly.
  • StatePatch
    purpose: Structured local edit schema with action, target, and parameter seed.
    Internal action representation.
  • Roomer-CC
    purpose: Controlled-corruption repair training dataset with inverse StatePatch targets.
    Not released; cannot be externally reproduced or falsified yet.
  • Roomer-Eval / Practical independent evidence
    purpose: Unified evaluation protocol and five rule families for practical usability.
    Professional human judgments in Appendix M correlate with Practical levels, providing an external anchor, though on baseline layouts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Roomer: Reflective Object-Grounded Model Editing and Repair for 3D Indoor Layout Synthesis." pith.science (2026). https://pith.science/paper/7264FBIP

@misc{pith2026260801973,
  author       = {Pith},
  title        = {Pith review of: Roomer: Reflective Object-Grounded Model Editing and Repair for 3D Indoor Layout Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7264FBIP}},
  note         = {Machine review of arXiv:2608.01973}
}
read the original abstract

Existing indoor layout generators produce globally plausible layouts yet may retain local violations such as collisions, out-of-bounds placements, obstructed openings, and blocked circulation. Most prior work focuses on full-scene synthesis or scene-level optimization, with limited support for identifying responsible objects and locally repairing affected regions. We present Roomer, a reflective repair framework that casts these violations as sparse, object-grounded repair problems. Roomer encodes layouts as ``RoState'' and uses ``RoReview'' to bind measured violations to implicated objects. A geometry-conditioned vision-language model planner proposes a structured local edit, while a deterministic solver validates it and generates a finite set of candidate edits when needed. Each candidate is committed only if full-scene verification confirms that it resolves the target violation without new hard violations or broken protected constraints. We train the planner on Roomer-CC, a controlled-corruption dataset that pairs faulty layouts with object-grounded violation evidence and known-feasible inverse StatePatches. Since existing benchmarks rarely assess whether physically valid layouts are usable, we introduce Roomer-Eval to assess distributional quality, physical validity, and practical usability. Experiments show that Roomer repairs residual violations while preserving valid regions, improves physical validity and usability, and transfers across external generators.

Figures

Figures reproduced from arXiv: 2608.01973 by the authors.

Figure 1
Figure 1. Motivation for Roomer. A generated layout may ap [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of Roomer. A generated layout is converted into an object-addressable RoState and evaluated to construct [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Representative final-layout comparisons on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Round-wise convergence and cost–quality trade-off on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 28 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 54 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    2021 , url =

    Huan Fu and Bowen Cai and Lin Gao and Lingxiao Zhang and Jiaming Wang and Cao Li and Qixun Zeng and Chengyue Sun and Rongfei Jia and Binqiang Zhao and Hao Zhang , publisher =. 2021 , url =. doi:10.1109/ICCV48922.2021.01075 , timestamp =

    2021

  5. [13]

    Maybank and Dacheng Tao , title =

    Huan Fu and Rongfei Jia and Lin Gao and Mingming Gong and Binqiang Zhao and Stephen J. Maybank and Dacheng Tao , title =. Int. J. Comput. Vis. , volume =. 2021 , url =. doi:10.1007/S11263-021-01534-Z , timestamp =

    2021 doi

  6. [15]

    2011 , url =

    Paul Merrell and Eric Schkufza and Zeyang Li and Maneesh Agrawala and Vladlen Koltun , title =. 2011 , url =. doi:10.1145/2010324.1964982 , timestamp =

    2011

  7. [16]

    Funkhouser and Pat Hanrahan , title =

    Matthew Fisher and Daniel Ritchie and Manolis Savva and Thomas A. Funkhouser and Pat Hanrahan , title =. 2012 , url =. doi:10.1145/2366145.2366154 , timestamp =

    2012

  8. [17]

    2015 , url =

    Matthew Fisher and Manolis Savva and Yangyan Li and Pat Hanrahan and Matthias Nie. 2015 , url =. doi:10.1145/2816795.2818057 , timestamp =

    2015

  9. [18]

    2018 , url =

    Siyuan Qi and Yixin Zhu and Siyuan Huang and Chenfanfu Jiang and Song. 2018 , url =. doi:10.1109/CVPR.2018.00618 , timestamp =

    2018

  10. [19]

    2019 , url =

    Manyi Li and Akshay Gadi Patil and Kai Xu and Siddhartha Chaudhuri and Owais Khan and Ariel Shamir and Changhe Tu and Baoquan Chen and Daniel Cohen. 2019 , url =. doi:10.1145/3303766 , timestamp =

    2019 doi

  11. [20]

    2019 , url =

    Kai Wang and Yu. 2019 , url =. doi:10.1145/3306346.3322941 , timestamp =

    2019

  12. [21]

    Dhamo, Helisa and Manhardt, Fabian and Navab, Nassir and Tombari, Federico , booktitle=

  13. [22]

    2023 , url =

    Lin Gao and Jia. 2023 , url =. doi:10.1109/TPAMI.2023.3237577 , timestamp =

    2023

  14. [23]

    2021 , url =

    Xinpeng Wang and Chandan Yeshwanth and Matthias Nie. 2021 , url =. doi:10.1109/3DV53792.2021.00021 , timestamp =

    2021

  15. [24]

    Paschalidou, Despoina and Kar, Amlan and Shugrina, Maria and Kreis, Karsten and Geiger, Andreas and Fidler, Sanja , journal=

  16. [26]

    The Twelfth International Conference on Learning Representations,

    Chenguo Lin and Yadong Mu , publisher =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  17. [27]

    Chang , publisher =

    Xiaohao Sun and Divyam Goel and Angel X. Chang , publisher =. 2026 , url =. doi:10.1109/3DV69130.2026.00147 , timestamp =

    2026

  18. [28]

    Weixi Feng and Wanrong Zhu and Tsu. Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

    2023

  19. [29]

    2024 , url =

    Yue Yang and Fan. 2024 , url =. doi:10.1109/CVPR52733.2024.01536 , timestamp =

    2024

  20. [31]

    Computer Vision -

    Rao Fu and Zehao Wen and Zichen Liu and Srinath Sridhar , editor =. Computer Vision -. 2024 , url =. doi:10.1007/978-3-031-72933-1\_4 , timestamp =

    2024 doi

  21. [32]

    Wang and Angel X

    Hou In Ivan Tam and Hou In Derek Pun and Austin T. Wang and Angel X. Chang and Manolis Savva , publisher =. 2026 , url =. doi:10.1109/WACV61042.2026.00710 , timestamp =

    2026

  22. [33]

    Tuomas Kynk. Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada , pages =. 2019 , url =

    2019

  23. [34]

    Sutherland and Michael Arbel and Arthur Gretton , publisher =

    Mikolaj Binkowski and Danica J. Sutherland and Michael Arbel and Arthur Gretton , publisher =. 6th International Conference on Learning Representations,. 2018 , url =

    2018

  24. [35]

    CoRR , volume =

    Shuai Bai and Keqin Chen and Xuejing Liu and Jialin Wang and Wenbin Ge and Sibo Song and Kai Dang and Peng Wang and Shijie Wang and Jun Tang and Humen Zhong and Yuanzhi Zhu and Ming. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.13923 , eprinttype =. 2502.13923 , timestamp =

    2025 doi arXiv

  25. [36]

    CoRR , volume =

    Chenfei Wu and Jiahao Li and Jingren Zhou and Junyang Lin and Kaiyuan Gao and Kun Yan and Shengming Yin and Shuai Bai and Xiao Xu and Yilei Chen and Yuxiang Chen and Zecheng Tang and Zekai Zhang and Zhengyi Wang and An Yang and Bowen Yu and Chen Cheng and Dayiheng Liu and Deqi...

    2025 doi arXiv

  26. [37]

    Narasimhan and Yuan Cao , publisher =

    Shunyu Yao and Jeffrey Zhao and Dian Yu and Nan Du and Izhak Shafran and Karthik R. Narasimhan and Yuan Cao , publisher =. The Eleventh International Conference on Learning Representations,. 2023 , url =

    2023

  27. [38]

    Noah Shinn and Federico Cassano and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao , editor =. Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 20...

    2023

  28. [39]

    Aman Madaan and Niket Tandon and Prakhar Gupta and Skyler Hallinan and Luyu Gao and Sarah Wiegreffe and Uri Alon and Nouha Dziri and Shrimai Prabhumoye and Yiming Yang and Shashank Gupta and Bodhisattwa Prasad Majumder and Katherine Hermann and Sean Welleck and Amir Yazdanbakh...

    2023

  29. [40]

    Bucher and Iro Armeni , booktitle=

    Martin JJ. Bucher and Iro Armeni , booktitle=. 2026 , url=

    2026

  30. [41]

    Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA,

    Martin Heusel and Hubert Ramsauer and Thomas Unterthiner and Bernhard Nessler and Sepp Hochreiter , editor =. Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA,. 2017 , url =

    2017

  31. [42]

    The Twelfth International Conference on Learning Representations,

    Zhibin Gou and Zhihong Shao and Yeyun Gong and Yelong Shen and Yujiu Yang and Nan Duan and Weizhu Chen , publisher =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  32. [43]

    2022 , url =

    Kurt Leimer and Paul Guerrero and Tomer Weiss and Przemyslaw Musialski , editor =. 2022 , url =. doi:10.1145/3550469.3555425 , timestamp =

    2022

  33. [44]

    2025 , url =

    Fan. 2025 , url =. doi:10.1109/CVPR52734.2025.02744 , timestamp =

    2025

  34. [45]

    Yandan Yang and Baoxiong Jia and Shujie Zhang and Siyuan Huang , editor =. Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2-7, 2025 / Mexico City, Mexico, Novembe...

    2025

  35. [46]

    Chang and Manolis Savva and Kai Wang and Hao Zhang , title =

    Ruiqi Wang and Qimin Chen and Daniel Ritchie and Angel X. Chang and Manolis Savva and Kai Wang and Hao Zhang , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2605.30819 , eprinttype =. 2605.30819 , timestamp =

    2026 doi arXiv

  36. [47]

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

    Noh, Seongrae and Seo, SeungWon and Park, Gyeong. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  37. [48]

    The Thirteenth International Conference on Learning Representations,

    Kaizhi Zheng and Xiaotong Chen and Xuehai He and Jing Gu and Linjie Li and Zhengyuan Yang and Kevin Lin and Jianfeng Wang and Lijuan Wang and Xin Eric Wang , publisher =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  38. [49]

    2026 , url =

    Lu Ling and Chen. 2026 , url =

    2026

  39. [50]

    2025 , url =

    Weilin Sun and Xinran Li and Manyi Li and Kai Xu and Xiangxu Meng and Lei Meng , editor =. 2025 , url =. doi:10.1609/AAAI.V39I7.32765 , timestamp =

    2025 doi

  40. [51]

    2026 , url =

    Chucheng Xiang and Ruchao Bao and Biyin Feng and Wenzheng Wu and Zhongyuan Liu and Yirui Guan and Ligang Liu , editor =. 2026 , url =. doi:10.1609/AAAI.V40I17.38452 , timestamp =

    2026 doi

  41. [52]

    2026 , url =

    Fr. 2026 , url =. doi:10.1609/AAAI.V40I4.37229 , timestamp =

    2026 doi

  42. [53]

    2026 , url =

    Hongchi Xia and Xuan Li and Zhaoshuo Li and Qianli Ma and Jiashu Xu and Ming. 2026 , url =

    2026

  43. [54]

    CoRR , volume =

    Yang Zhao and Shizhao Sun and Meisheng Zhang and Yingdong Shi and Xubo Yang and Jiang Bian , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.09432 , eprinttype =. 2602.09432 , timestamp =

    2026 doi

  44. [55]

    2025 , url=

    Lingwei Dang and Ruizhi Shao and Hongwen Zhang and Wei Min and Yebin Liu and Qingyao Wu , booktitle=. 2025 , url=

    2025

  45. [56]

    2607.17097 , archivePrefix=

    Lingwei Dang and Juntong Li and Zonghan Li and Hongwen Zhang and Liang An and Wei Min and Yebin Liu and Qingyao Wu , year=. 2607.17097 , archivePrefix=

    arXiv

  46. [57]

    2024 , doi =

    Yandan Yang and Baoxiong Jia and Peiyuan Zhi and Siyuan Huang , title =. 2024 , doi =

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.