Pith. sign in

REVIEW 5 major objections 6 minor 46 references

The paper claims that verifying each furniture placement before it joins the layout—rather than generating the whole room at once—cuts structural violations from over 90% to under 17%.

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

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

T0 review · deepseek-v4-flash

2026-08-01 09:06 UTC pith:HNQLY5PB

load-bearing objection A worthwhile iterative multi-agent layout generation idea and a potentially useful benchmark, but the empirical claims need artifact release and a few fixes before I'd trust the headline numbers. the 5 major comments →

arxiv 2607.20866 v1 pith:HNQLY5PB submitted 2026-07-23 cs.CV

Agentic Designer: Progressive Multi-Agent Collaboration for Structure-Aware Interior Layout Generation

Zhijing Yang , Haocheng Lin , Zhihua Xu , Haojie Li , Keze Wang , Liang Lin , Tianshui Chen This is my paper
classification cs.CV
keywords interior layout generationmulti-agent collaborationprogressive consensus mechanismstructure-aware generationLLM spatial reasoninggeometric constraint verificationlayout benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that the dominant one-shot paradigm for interior layout generation—diffusion models or a single LLM pass from constraints to a complete layout—is the wrong abstraction when architectural constraints are strict. Its central fix is a multi-agent loop, the Progressive Consensus Mechanism, in which a Generator proposes one furniture item, an Evaluator checks it against walls, doors, windows, and already-placed items, and a Refiner corrects flagged attributes before the item is committed to the layout history. The claim is that this verify-before-commit protocol prevents error accumulation: every subsequent placement conditions on a structure-valid history, and the paper reports scene violation rates falling from over 90% for baselines to 16.85% for living rooms and 7.14% for bedrooms, with boundary violations cut by an order of magnitude. A sympathetic reader should care because the paper also provides a benchmark (InStruct, 18,853 manually verified samples) and structure-centric metrics that would let the field test this claim directly, rather than relying on visual impression.

Core claim

On the paper's own terms, the discovery is that structural validity in layout generation is a property of the decision process, not just the generative model. Agentic Designer turns generation into a stepwise loop: the Generator predicts one candidate object (category, position, size, orientation) from the room structure and the history of accepted objects; the Evaluator diagnoses four violation types—partial boundary intrusion, complete boundary violation, object collision, and orientation misalignment; the Refiner adjusts only the flagged geometric attributes while preserving the intended object. The Progressive Consensus Mechanism commits an object to history only after it passes or is re

What carries the argument

The load-bearing machinery is the Progressive Consensus Mechanism (PCM), a Generator→Evaluator→Refiner loop. The room is encoded as a program-like coordinate representation of walls, doors, and windows; the Generator samples candidate objects; the Evaluator returns a four-bit Boolean violation report; the Refiner performs targeted corrections when any bit is set. PCM enforces a 'verified history'—every object that becomes context for the next step has been checked and, if needed, repaired—so errors cannot cascade the way they do in ordinary autoregressive generation. The paper's Eq. (2) states the invariant and Eq. (3) states the causal consequence: cleaner history makes the next prediction

Load-bearing premise

The argument rests on the Evaluator's coordinate-level checks being trustworthy enough that the 'verified history' is actually clean; the paper concedes the Evaluator can mistake a deep external recess for valid interior space, and if such false certifications are common, the claimed error-prevention advantage collapses on exactly the complex floor plans the benchmark is designed around.

What would settle it

Take the test subset of rooms with deep external recesses or concave boundaries, and run the Evaluator alone on proposals deliberately placed in the exterior nook that is locally coordinate-valid; compute its false-acceptance rate. If that rate is high, the verified history of Eq. (2) is not clean on complex rooms, and the central claim reduces to local correction rather than genuine error prevention. A simpler proxy comparison: SVR on the recess-heavy subset versus the rectangular subset—a large gap driven by Evaluator accepts outside the interior would directly test the mechanism's boundary.

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

If this is right

  • Treating layout generation as a sequential verify-and-commit process, rather than a one-shot mapping, reduces scene violation rates from roughly 90% or higher to 16.85% (living rooms) and 7.14% (bedrooms) on the paper's own benchmark, with boundary violation counts cut by about an order of magnitude.
  • A clean, verified history is itself a causal driver of quality: when the same Generator is evaluated on its raw proposals only, providing a verified history lowers SVR from 20.00% to 17.97% in living rooms and from 21.14% to 17.38% in bedrooms (Table 8).
  • Stepwise evaluation-and-refinement beats a single global post-hoc pass: replacing PCM with global refinement raises SVR from 7.14% to 17.74% in bedrooms, showing error prevention is not the same as error correction.
  • The mechanism converges quickly: with Nmax=1 refinement pass per object the system already attains its headline numbers, allowing up to 4 passes only modestly lowers SVR (to 8.43% living rooms) while average refinement steps plateau near 0.08–0.10 per object; latency is about 2.2×–2.4× a one-shot generator.
  • The same verified loop transfers to a different data distribution (3D-FRONT) without fine-tuning, cutting SVR to 14.81% (living rooms) and 8.33% (bedrooms) compared with 86–93% for baselines.

Where Pith is reading between the lines

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

  • Beyond the paper: the verify-before-commit loop is a generic control structure for any generation task where early errors propagate—long-form code synthesis, multi-step planning, molecular or layout generation—so the mechanism's value may not be specific to furniture.
  • Beyond the paper: the paper's own failure analysis predicts a testable upgrade: replacing or augmenting the coordinate-only Evaluator with a topology-aware visual verifier should differentially improve SVR on rooms with deep external recesses; one could measure false-acceptance rates before and after such an upgrade.
  • Beyond the paper: because the Refiner is trained on noise-injected ground truth, the system's ceiling tracks the Evaluator's certification accuracy; an adversarial probe of the Evaluator on locally valid but globally exterior placements would bound how much of the reported 16.85%/7.14% is genuine structural understanding versus room-geometry luck.
  • Beyond the paper: the default Nmax=1 trades some collision reduction for boundary compliance (in living rooms, global refinement reaches CS 2.54 while progressive reaches 3.27), suggesting an adaptive per-object iteration budget—spend more refinements on objects near boundaries or in dense regions—could improve both metrics simultaneously.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces Agentic Designer, an LLM-based multi-agent framework for structure-aware interior layout generation. A Generator proposes furniture objects sequentially; an Evaluator checks each candidate against four geometric violation types (partial/complete boundary intrusion, collision, orientation misalignment); and a Refiner corrects flagged objects. A Progressive Consensus Mechanism commits only evaluated/refined objects to the conditioning history, with the stated goal of preventing error accumulation. The paper also introduces InStruct, a benchmark with 18,853 manually verified parametric room layouts and a suite of structural metrics (SVR, BVC, CS, OR, ANF), plus an adapted 3D-FRONT testbed. Experiments report large gains over DiffuScene and SemLayout on InStruct and 3D-FRONT, with ablations on multi-agent collaboration, progressive versus global refinement, and history quality. The authors acknowledge in Section 6 that concave boundaries and deep external recesses can still confuse the coordinate-based Evaluator.

Significance. If the claims hold, the work would be a useful step toward constraint-aware, interpretable layout generation: the multi-agent verification loop is timely, the InStruct dataset is larger and more topologically diverse than many existing benchmarks, and the structured metrics address a real gap in evaluating structural adherence. The promised release of code, models, and data is a concrete strength, as is the inclusion of ablations that test the core mechanism. However, the central quantitative claims are currently weakened by fairness and circularity concerns, and the acknowledged failure mode on complex concave geometries is precisely the regime the benchmark emphasizes. The contribution is promising but not yet convincingly established.

major comments (5)
  1. [§5.1.1, Table 2] The main comparison on InStruct is confounded by training domain. Agentic Designer is fine-tuned on the InStruct training split, while DiffuScene and SemLayout are run from their official checkpoints without any InStruct fine-tuning, as stated in §4.4. Table 2 therefore measures adaptation advantage as much as method quality. To support the claim of state-of-the-art performance, the authors should fine-tune or train the baselines on the same InStruct training data (or provide a zero-shot/few-shot comparison labeled as such) and report both settings.
  2. [§4.3, Tables 1–3] The Occupancy Ratio (OR) direction is internally inconsistent. Table 1's caption says lower values are better for OR, while Table 2 and the §5.1.1 text treat higher OR as better ("more realistic Occupancy Ratio"); Table 3 says lower is better. Because OR appears in the core quantitative comparisons and ablations, this inconsistency makes the direction of the reported OR improvements ambiguous. Please define OR precisely and state once whether high or low occupancy is preferred, then apply the same convention throughout.
  3. [§3.2.2, §3.4.2, §5.1] There is a circularity risk in the headline SVR/BVC/CS gains. The Evaluator is fine-tuned to detect exactly the four violation types that the benchmark metrics count (boundary intrusion, collision, orientation), and the Refiner is trained to denoise exactly those same perturbation types. Thus, the full system is optimized on the same error categories used for evaluation, while the baselines are not. The authors should provide external validation: e.g., a geometric checker that does not share the training-perturbation design, human-annotated violation labels, and results on violation types not included in Evaluator/Refiner training. Table 8 is a good step, but the Generator is still trained in a pipeline whose history is produced by the same Evaluator/Refiner, so it does not fully dispel the concern.
  4. [§6, Eq. (2), Fig. 6] The acknowledged failure on deep external recesses directly undermines the "verified history" premise stated in Eq. (2). If the Evaluator can certify coordinates in an external recess as valid interior space, then the conditioning history is not guaranteed clean, and the strict structural-adherence claim fails exactly on the complex non-rectangular rooms that InStruct emphasizes. Moreover, the reported SVR/BVC metrics likely use coordinate-level boundary checks analogous to the Evaluator's, so the metrics may share the same blind spot. Please quantify this: report SVR/BVC stratified by room complexity (e.g., number of wall segments or concavity index), use a robust connected-interior/point-in-polygon check, and state how often the Evaluator accepts placements outside the true connected interior. Without this, the headline 16.85% SVR may overstate true structural validity.
  5. [Tables 2–9, §5.3] No error bars, confidence intervals, or significance tests are reported. The main tables compare single point estimates without variance, and the user study uses 31 participants and 30 scenes. Given the small sizes in Table 3 (300 rooms) and the small user-study sample, the word "significantly" is not supported. Please report standard deviations over multiple runs/seeds and appropriate statistical tests for the perceptual scores.
minor comments (6)
  1. [Algorithm 1, §5.5] With the default Nmax=1, an object that fails evaluation is refined once and committed without a post-refinement re-check. This means the accepted history can contain objects the Refiner failed to fix. Please state this explicitly and consider reporting the default setting with post-refinement verification, or clarify why the one-shot commit is acceptable.
  2. [§4.4, Table 3] The table is labeled "3D-FRONT test set," but the text says 300 rooms were randomly sampled from a curated subset of adapted 3D-FRONT. Please clarify the relationship between this subset and the original 3D-FRONT train/test split, especially since the baselines were originally trained on 3D-FRONT data.
  3. [§4.1.1] The InStruct dataset is constructed from real-world floor plans, but the source, licensing, and annotation-quality agreement are not described in enough detail. If Appendix A contains this, please point to it explicitly; otherwise add a data-provenance statement and inter-annotator agreement.
  4. [§3.4, §5.6] For reproducibility, please specify the sampling temperature, random seed, and decoding parameters used for the Generator/Evaluator/Refiner, and report latency with variance. The current latency table gives only single mean values.
  5. [§4.3.1, Table 1] The header "CS (IoU)↓" is confusing: Collision Score is described as a scaled average pairwise IoU where lower is better, but the parenthetical "IoU" suggests higher overlap is better. Please rename or add a footnote to avoid ambiguity.
  6. [Abstract, §7] The abstract and conclusion say layouts "strictly adhere" to architectural constraints. Given the Section 6 limitation, please soften this to "substantially improve" or "more closely adhere," and state the known failure mode in the abstract-level summary.

Circularity Check

0 steps flagged

No load-bearing circularity; the paper's formal inequality is close to definitional, but independent controls (Table 8, 3D-FRONT, user study) keep the central claim grounded.

full rationale

The central claim is empirical rather than derivational: Agentic Designer's SVR/BVC gains are evaluated with geometric metrics, not with the Evaluator's own outputs. The Evaluator is fine-tuned to flag exactly the violation categories counted by SVR/BVC/CS (§3.2.2, §3.4.2, §4.3.1), so the headline numbers partly reflect the training objective; however, this is a standard train/evaluate alignment rather than a fitted parameter renamed as a prediction. The paper explicitly anticipates the circularity objection in §5.4.3/Table 8, where metrics are computed strictly on the Generator's raw, unrefined proposals under verified vs. unverified histories, providing independent evidence for Eq. (3). Cross-dataset evaluation on 3D-FRONT (§5.1.2) and the user study (§5.3) further ground the claim outside the training distribution. Eq. (2)–(3) are not a derivation but a formal restatement of the mechanism's selection rule; V and Pvalid are not independently defined, so the inequality is largely a restatement of the filtering procedure. Section 6 concedes Evaluator false negatives on concave recesses, which weakens the scope of 'strict structural adherence' but is an acknowledged empirical limitation, not circularity. Self-citations [9], [18], [32] appear only as related work and are not load-bearing. Overall, no specific reduction of a prediction to its inputs is exhibited; score 2 reflects the minor definitional flavor of Eq. (3) and the evaluator/metric alignment, not genuine circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 1 invented entities

The central empirical claims rest on several unverified pillars: (i) the coordinate/code representation preserves the topology the Evaluator needs; (ii) the learned Evaluator is correct often enough for Eq. (2); (iii) the InStruct and adapted-3D-FRONT annotations are accurate; and (iv) the Refiner generalizes from synthetic noise to real Generator errors. The paper tests (ii) indirectly but concedes its failure mode in Section 6. No new physical entities are introduced; the End token is a model construct.

free parameters (5)
  • Nmax (max refinement iterations per object) = 1 (default); 2–4 studied
    Inference-time budget directly trades off SVR/CS; Table 9 shows Nmax=2 reduces living-room SVR from 16.85% to 10.04%, so reported headline results depend on this choice.
  • LoRA rank and alpha = r=32, alpha=64
    Chosen hyperparameters for all three agents; not swept, affects capacity of the fine-tuned agents.
  • Learning rate = 1e-4
    AdamW learning rate for fine-tuning; no sensitivity analysis provided.
  • Perturbation ranges for Evaluator/Refiner training noise
    Random translation/rotation noise injected into ground-truth layouts defines what the Evaluator learns to flag and the Refiner learns to fix; magnitudes are unstated and load-bearing for verification accuracy.
  • Furniture ordering rule for Generator training data = semantic importance and frequency
    Generator training data is ordered by this heuristic; it shapes the autoregressive distribution and termination behavior.
axioms (6)
  • domain assumption Coordinate-based, program-like textual representation fully captures room topology and constraint geometry.
    Used throughout §3.4.1; the paper itself qualifies this in Section 6, admitting external recesses are misread, so the axiom is only partially valid.
  • domain assumption The learned Evaluator's violation judgments are accurate enough that committed history satisfies Eq. (2): E(o_k,S,O_<k)≈0.
    Eq. (2) in §3.3; if false, the verified-context story collapses; Section 6 gives counterexamples.
  • domain assumption Improved validity of conditioning history strictly increases generator prediction validity (Eq. 3).
    Eq. (3) is asserted, not derived; Table 8 tests it but only on point estimates without error bars.
  • domain assumption Qwen2.5-Coder-7B pretraining transfers to geometric layout reasoning via LoRA fine-tuning.
    The entire framework rests on this backbone; no comparison to other backbones is given.
  • domain assumption The adapted 3D-FRONT annotations (walls/doors/windows) are topology-consistent and comparable to InStruct.
    Section 4.1.1; the authors reprocess raw meshes; quality of this adaptation affects Table 3 and all 3D-FRONT evaluations.
  • domain assumption The Refiner, trained on synthetic perturbations of ground-truth layouts, generalizes to the Generator's actual mistakes at inference.
    Training pairs are invalid/valid states derived from perturbed ground truth; the paper does not measure Refiner accuracy on out-of-distribution Generator errors.
invented entities (1)
  • End token (terminal furniture instance with category c=End) no independent evidence
    purpose: Lets the Generator decide when the layout is complete during autoregressive generation.
    An implementation device, not a physical claim; it functions like an end-of-sequence token and carries no external falsifiable prediction.

pith-pipeline@v1.3.0-alltime-deepseek · 20544 in / 16862 out tokens · 167944 ms · 2026-08-01T09:06:26.644130+00:00 · methodology

0 comments
read the original abstract

Generating realistic interior furniture layouts that strictly adhere to architectural constraints (e.g., walls, doors, and windows) remains a fundamental challenge in automated spatial design. Existing approaches, primarily based on one-shot generation using diffusion models or Large Language Models (LLMs), lack explicit mechanisms for intermediate geometric constraint verification, often resulting in structural collisions and functionally infeasible arrangements under complex room constraints. To address these challenges, we propose Agentic Designer, a progressive, multi-agent framework that formulates structure-aware interior layout generation as an iterative and constraint-verified decision process. By decomposing layout synthesis into modular stages of proposal, verification, and adjustment, the framework coordinates three specialized agents, a Generator, an Evaluator, and a Refiner, through a Progressive Consensus Mechanism. This mechanism enforces stepwise geometric validation and correction before each placement is committed, thereby preventing error accumulation. To facilitate this structure-aware paradigm and standardize evaluation, we establish InStruct, a comprehensive benchmark that integrates a dataset comprising over 18,000 high-quality, parametrically annotated samples with a novel suite of structure-centric metrics. Extensive quantitative evaluations, qualitative analyses, and user studies show that Agentic Designer significantly outperforms state-of-the-art methods, demonstrating substantial improvements in strict structural adherence and functional design coherence.

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

46 extracted references · 9 linked inside Pith

  1. [1]

    Planit: Planning and instan- tiating indoor scenes with relation graph and spatial prior networks,

    K. Wang, Y.-A. Lin, B. Weissmann, M. Savva, A. X. Chang, and D. Ritchie, “Planit: Planning and instan- tiating indoor scenes with relation graph and spatial prior networks,”ACM Transactions on Graphics (TOG), vol. 38, no. 4, pp. 1–15, 2019

    2019

  2. [2]

    Atiss: Autoregressive transformers for indoor scene synthesis,

    D. Paschalidou, A. Kar, M. Shugrina, K. Kreis, A. Geiger, and S. Fidler, “Atiss: Autoregressive transformers for indoor scene synthesis,”Advances in Neural Information Processing Systems, vol. 34, pp. 12 013–12 026, 2021

    2021

  3. [3]

    Layoutgpt: Compositional visual planning and generation with large language models,

    W. Feng, W. Zhu, T.-j. Fu, V . Jampani, A. Akula, X. He, S. Basu, X. E. Wang, and W. Y. Wang, “Layoutgpt: Compositional visual planning and generation with large language models,”Advances in Neural Information Processing Systems, vol. 36, pp. 18 225–18 250, 2023

    2023

  4. [4]

    Diffuscene: Denoising diffusion models for generative indoor scene synthesis,

    J. Tang, Y. Nie, L. Markhasin, A. Dai, J. Thies, and M. Nießner, “Diffuscene: Denoising diffusion models for generative indoor scene synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 507–20 518

    2024

  5. [5]

    Physcene: Physi- cally interactable 3d scene synthesis for embodied ai,

    Y. Yang, B. Jia, P . Zhi, and S. Huang, “Physcene: Physi- cally interactable 3d scene synthesis for embodied ai,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 262–16 272

    2024

  6. [6]

    S-inf: Towards realistic indoor scene synthesis via scene implicit neural field,

    Z. Liang, G. Xu, H. Wu, Y. Huang, W. Li, and L. Duan, “S-inf: Towards realistic indoor scene synthesis via scene implicit neural field,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 5, 2025, pp. 5173–5181

    2025

  7. [7]

    Semlayoutdiff: Semantic layout generation with diffusion model for indoor scene synthesis,

    X. Sun, D. Goel, and A. X. Chang, “Semlayoutdiff: Semantic layout generation with diffusion model for indoor scene synthesis,” pp. 1545–1555, 2026

    2026

  8. [8]

    Scenecraft: An llm agent for synthesizing 3d scenes as blender code,

    Z. Hu, A. Iscen, A. Jain, T. Kipf, Y. Yue, D. A. Ross, C. Schmid, and A. Fathi, “Scenecraft: An llm agent for synthesizing 3d scenes as blender code,” inForty-first International Conference on Machine Learning, 2024

    2024

  9. [9]

    Learning spatial-temporal coherent correlations for speech-preserving facial expression manipulation,

    T. Chen, J. Lin, Z. Yang, C. Qing, G. Wang, and L. Lin, “Learning spatial-temporal coherent correlations for speech-preserving facial expression manipulation,” IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2026

    2026

  10. [10]

    Automation in interior space planning: Utilizing conditional generative adversarial network models to create furniture layouts,

    H. Tanasra, T. Rott Shaham, T. Michaeli, G. Austern, and S. Barath, “Automation in interior space planning: Utilizing conditional generative adversarial network models to create furniture layouts,”Buildings, vol. 13, no. 7, p. 1793, 2023

    2023

  11. [11]

    Chat2layout: Interactive 3d furniture layout with a multimodal llm

    C. Wang, H. Zhong, M. Chai, M. He, D. Chen, and J. Liao, “Chat2layout: Interactive 3d furniture layout with a multimodal llm.”IEEE transactions on visualization and computer graphics, 2025

    2025

  12. [12]

    Matter- port3d: Learning from rgb-d data in indoor environ- ments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y. Zhang, “Matter- port3d: Learning from rgb-d data in indoor environ- ments,”arXiv preprint arXiv:1709.06158, 2017

    Pith/arXiv arXiv 2017

  13. [13]

    Semantic scene completion from a single depth image,

    S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and T. Funkhouser, “Semantic scene completion from a single depth image,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1746– 1754

    2017

  14. [14]

    Structured3d: A large photo-realistic dataset for struc- tured 3d modeling,

    J. Zheng, J. Zhang, J. Li, R. Tang, S. Gao, and Z. Zhou, IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 15 “Structured3d: A large photo-realistic dataset for struc- tured 3d modeling,” inEuropean Conference on Computer Vision. Springer, 2020, pp. 519–535

    2020

  15. [15]

    3d-front: 3d furnished rooms with layouts and semantics,

    H. Fu, B. Cai, L. Gao, L.-X. Zhang, J. Wang, C. Li, Q. Zeng, C. Sun, R. Jia, B. Zhaoet al., “3d-front: 3d furnished rooms with layouts and semantics,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 933–10 942

    2021

  16. [16]

    Fast and flexible indoor scene synthesis via deep convolutional genera- tive models,

    D. Ritchie, K. Wang, and Y.-a. Lin, “Fast and flexible indoor scene synthesis via deep convolutional genera- tive models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 6182– 6190

    2019

  17. [17]

    In- door scene generation from a collection of semantic- segmented depth images,

    M.-J. Yang, Y.-X. Guo, B. Zhou, and X. Tong, “In- door scene generation from a collection of semantic- segmented depth images,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 203–15 212

    2021

  18. [18]

    Contrastive decoupled representation learning and regularization for speech-preserving facial expression manipulation,

    T. Chen, J. Lin, Z. Yang, C. Qing, Y. Shi, and L. Lin, “Contrastive decoupled representation learning and regularization for speech-preserving facial expression manipulation,”International Journal of Computer Vision, vol. 133, no. 7, pp. 3822–3838, 2025

    2025

  19. [19]

    Grains: Generative recursive autoencoders for indoor scenes,

    M. Li, A. G. Patil, K. Xu, S. Chaudhuri, O. Khan, A. Shamir, C. Tu, B. Chen, D. Cohen-Or, and H. Zhang, “Grains: Generative recursive autoencoders for indoor scenes,”ACM Transactions on Graphics (TOG), vol. 38, no. 2, pp. 1–16, 2019

    2019

  20. [20]

    Sceneformer: Indoor scene generation with transformers,

    X. Wang, C. Yeshwanth, and M. Nießner, “Sceneformer: Indoor scene generation with transformers,” in2021 International Conference on 3D Vision (3DV). IEEE, 2021, pp. 106–115

    2021

  21. [21]

    Lego-net: Learning regular rearrangements of objects in rooms,

    Q. A. Wei, S. Ding, J. J. Park, R. Sajnani, A. Poulenard, S. Sridhar, and L. Guibas, “Lego-net: Learning regular rearrangements of objects in rooms,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 037–19 047

    2023

  22. [22]

    Instructscene: Instruction-driven 3d indoor scene synthesis with semantic graph prior,

    C. Lin and Y. Mu, “Instructscene: Instruction-driven 3d indoor scene synthesis with semantic graph prior,” arXiv preprint arXiv:2402.04717, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Mixed diffusion for 3d indoor scene synthesis,

    S. Hu, D. M. Arroyo, S. Debats, F. Manhardt, L. Carlone, and F. Tombari, “Mixed diffusion for 3d indoor scene synthesis,”arXiv preprint arXiv:2405.21066, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    I-design: Personalized llm interior designer,

    A. C ¸elen, G. Han, K. Schindler, L. Van Gool, I. Armeni, A. Obukhov, and X. Wang, “I-design: Personalized llm interior designer,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 217–234

    2024

  25. [25]

    Anyhome: Open-vocabulary generation of structured and textured 3d homes,

    R. Fu, Z. Wen, Z. Liu, and S. Sridhar, “Anyhome: Open-vocabulary generation of structured and textured 3d homes,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 52–70

    2024

  26. [26]

    Holodeck: Language guided generation of 3d embod- ied ai environments,

    Y. Yang, F.-Y. Sun, L. Weihs, E. VanderBilt, A. Herrasti, W. Han, J. Wu, N. Haber, R. Krishna, L. Liuet al., “Holodeck: Language guided generation of 3d embod- ied ai environments,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 227–16 237

    2024

  27. [27]

    Code as policies: Language model programs for embodied control,

    J. Liang, W. Huang, F. Xia, P . Xu, K. Hausman, B. Ichter, P . Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,”arXiv preprint arXiv:2209.07753, 2022

    Pith/arXiv arXiv 2022

  28. [28]

    Do as i can, not as i say: Grounding language in robotic affordances,

    A. Brohan, Y. Chebotar, C. Finn, K. Hausman, A. Herzog, D. Ho, J. Ibarz, A. Irpan, E. Jang, R. Julianet al., “Do as i can, not as i say: Grounding language in robotic affordances,” inConference on robot learning. PMLR, 2023, pp. 287–318

    2023

  29. [29]

    Metagpt: Meta programming for a multi-agent collaborative framework,

    S. Hong, M. Zhuge, J. Chen, X. Zheng, Y. Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Linet al., “Metagpt: Meta programming for a multi-agent collaborative framework,” inThe Twelfth International Conference on Learning Representations, 2023

    2023

  30. [30]

    Mtr++: Multi- agent motion prediction with symmetric scene modeling and guided intention querying,

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “Mtr++: Multi- agent motion prediction with symmetric scene modeling and guided intention querying,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 3955–3971, 2024

    2024

  31. [31]

    Interaction pattern disentangling for multi- agent reinforcement learning,

    S. Liu, J. Song, Y. Zhou, N. Yu, K. Chen, Z. Feng, and M. Song, “Interaction pattern disentangling for multi- agent reinforcement learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 8157–8172, 2024

    2024

  32. [32]

    Mao: A framework for process model generation with multi- agent orchestration,

    L. Lin, Y. Jin, Y. Zhou, W. Chen, and C. Qian, “Mao: A framework for process model generation with multi- agent orchestration,”IEEE Transactions on Services Com- puting, 2025

    2025

  33. [33]

    Self-refine: Iterative refinement with self- feedback, 2023,

    A. Madaan, N. Tandon, P . Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yanget al., “Self-refine: Iterative refinement with self- feedback, 2023,”URL https://arxiv. org/abs/2303.17651, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    Mo-mix: Multi- objective multi-agent cooperative decision-making with deep reinforcement learning,

    T. Hu, B. Luo, C. Yang, and T. Huang, “Mo-mix: Multi- objective multi-agent cooperative decision-making with deep reinforcement learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 10, pp. 12 098–12 112, 2023

    2023

  35. [35]

    Demystifying chains, trees, and graphs of thoughts,

    M. Besta, F. Memedi, Z. Zhang, R. Gerstenberger, G. Piao, N. Blach, P . Nyczyk, M. Copik, G. Kwa´sniewski, J. M ¨ulleret al., “Demystifying chains, trees, and graphs of thoughts,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  36. [36]

    Scannet: Richly-annotated 3d recon- structions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d recon- structions of indoor scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5828–5839

    2017

  37. [37]

    Habitat-matterport 3d dataset (hm3d): 1000 large- scale 3d environments for embodied ai,

    S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. Turner, E. Under- sander, W. Galuba, A. Westbury, A. X. Changet al., “Habitat-matterport 3d dataset (hm3d): 1000 large- scale 3d environments for embodied ai,”arXiv preprint arXiv:2109.08238, 2021

    Pith/arXiv arXiv 2021

  38. [38]

    Scannet++: A high-fidelity dataset of 3d indoor scenes,

    C. Yeshwanth, Y.-C. Liu, M. Nießner, and A. Dai, “Scannet++: A high-fidelity dataset of 3d indoor scenes,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12–22

    2023

  39. [39]

    Interi- ornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset,

    W. Li, S. Saeedi, J. McCormac, R. Clark, D. Tzoumanikas, Q. Ye, Y. Huang, R. Tang, and S. Leutenegger, “Interi- ornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset,”arXiv preprint arXiv:1809.00716, 2018

    Pith/arXiv arXiv 2018

  40. [40]

    Qwen2. 5-coder technical report,

    B. Hui, J. Yang, Z. Cui, J. Yang, D. Liu, L. Zhang, T. Liu, J. Zhang, B. Yu, K. Luet al., “Qwen2. 5-coder technical report,”arXiv preprint arXiv:2409.12186, 2024. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 16

    Pith/arXiv arXiv 2024

  41. [41]

    From system 1 to system 2: a survey of reasoning large language models,

    D. Zhang, Z.-Z. Li, M.-L. Zhang, J. Zhang, Z. Liu, Y. Yao, H. Xu, J. Zheng, X. Chen, Y. Zhanget al., “From system 1 to system 2: a survey of reasoning large language models,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  42. [42]

    Multi-step reasoning with large language models, a survey,

    A. Plaat, A. Wong, S. Verberne, J. Broekens, and N. Van Stein, “Multi-step reasoning with large language models, a survey,”ACM Computing Surveys, 2025

    2025

  43. [43]

    Elic- iting and improving the causal reasoning abilities of large language models with conditional statements,

    X. Liu, D. Yin, C. Zhang, D. Zhao, and Y. Feng, “Elic- iting and improving the causal reasoning abilities of large language models with conditional statements,” Computational Linguistics, pp. 1–38, 2025

    2025

  44. [44]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y. Shen, P . Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.”ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  45. [45]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,”arXiv preprint arXiv:1711.05101, 2017. Zhijing Yangreceived the B.S and Ph.D. degrees from the Mathematics and Computing Science, Sun Y at-sen University, Guangzhou China, in 2003 and 2008, respectively. He was a Visiting Research Scholar in the School of Computing, Informatics and Media...

    Pith/arXiv arXiv 2017

  46. [2024]

    Hundred Talents Program

    He is currently pursuing the master’s de- gree with the School of Information Engineering, Guangdong University of Technology, under the supervision of Prof. Zhijing Y ang. His research interests include multi-agent systems and gener- ative design. Zhihua Xureceived the B.E. degree from the School of Computer Science and Technology and the M.S. degree fro...

    2017