acceptodds
Under review as a conference paper at ICLR 2027

FuncRoom-Agent: Sequential Feed-Forward 3D Functional Indoor Scene Generation

Abstract

We introduce Function-Room Generation, a new indoor 3D scene generation setting that creates rooms supporting explicit functional goals rather than merely visually plausible layouts. Existing agentic and executable methods improve controllability, but often depend on costly test-time generate–evaluate–revise loops, making functional room generation slow and computationally expensive. We address this challenge with three technical contributions. First, we design a recursive domain-specific language to effectively organize the hierarchical object compositions required by functional rooms, from room structure and major furniture to dense support-surface and nested small objects. It represents rooms as staged executable programs with explicit geometric and functional relations. Second, we propose a sequential feed-forward scene construction framework that distills recursive construction traces into a scene construction expert. At inference time, the expert writes executable DSL code stage by stage, and a deterministic executor directly instantiates each stage without teacher agents, online critics, or iterative repair. Third, we introduce ScenePRM, an execution-grounded process reward framework that improves the expert through reinforcement learning with functional, geometric, relational, and future-constructability feedback. We further establish a function-oriented benchmark and show state-of-the-art performance on both general indoor scene generation and function-room generation, achieving stronger functional completeness, relation correctness, geometric executability, and generation efficiency.

Primary area:
generative models
Keywords:
3D scene generation, agentic RL, function room generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_z665OpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces FuncRoom-Agent, which is used for generating 3D indoor scenes based on functional requirements. The method builds the scene hierarchically through recursive DSL and uses the construction trajectory generated by the agent for visual-context SFT. Then, it further optimizes through ScenePRM-guided RL. The inference stage adopts sequential feed-forward construction and no longer performs explicit test-time critique-and-repair. The experiments are conducted on SceneEval and the author's proposed Function-Room benchmark.

Critical strengths
  1. The generation of indoor scenes has been further expanded from visual rationality to functional completeness and usability, and the problem setting has certain practical significance.
  2. Recursive DSL can naturally represent the hierarchical structure among furniture, support relationships, wall objects, and nested small objects.
Critical weaknesses
  1. Hierarchical scene representation, DSL/code-based scene generation, execution feedback, and agent refinement have all been extensively explored in recent works (Sage, SceneWeaver, scenethesis, scenesmith). It is hoped to clarify the novelty of this paper.
  2. There is currently no clear ablation study conducted for different input modalities.
Questions for authors
  1. Are the Function-Room automatic evaluation indicators stable under different VLM evaluators? And what is the correlation between them and human ratings?
  2. It is not clear whether different methods use the same asset library, retrieval/generation backend, rendering setup and external models.
  3. It is claimed that this method has achieved an improvement in efficiency, as it avoids the iterative repair during test time. However, the complete process still takes approximately 25 minutes per scene, while the closest alternative, Code as Room, takes 33 minutes.
  4. The scale of the Function-Room benchmark is relatively small, with only 50 test sets.
  5. Function-Room metrics relies on VLM for its judgment, and ScenePRM itself also depends on GPT-based critic. There is a lack of sensitivity analysis under different VLM evaluators, as well as a correlation analysis between automatic metrics and human judgment.
1 comment
AuthorsOpenReview ↗

Response to Reviewer z665: DSL Novelty, Input Ablations, Evaluation, and Runtime

Thank you for the constructive review. We address the critical weaknesses (W) and the questions (Q) below.

W1. Novelty of the Our DSL

Representation. Our contribution is not recursion alone: our DSL yields execution-grounded training traces. Each fragment builds one functional group from committed parents and can be executed and committed on its own. Each step records the pre-action state and render, the action, and the grounded result. Recursion defines stages for sampling K candidates from the same state.

Training. We use these traces for SFT on state–action pairs and stage-level RL with ScenePRM rewards. SAGE, SceneWeaver, Scenethesis, and SceneSmith prompt general-purpose models for scene generation.

Inference. Conditioned on the current scene state and render, our trained generator completed all 50 tasks with one candidate per step and no iterative critique or repair. SAGE, SceneWeaver, and SceneSmith use iterative feedback, and Scenethesis can replan after judging spatial coherence.

Table 3 in the main paper shows that our recursive agentic construction performs better than both one-shot construction with a flat DSL and staged construction on Function-Room.

MethodIterative repair at inferenceTraining of a dedicated scene generatorProcess-level RLDedicated Function-Room task and evaluation
SAGE✓———
SceneWeaver✓———
SceneThesis✓———
SceneSmith✓———
Code as Room✓———
FuncRoom-Agent—✓✓✓

W2. Input modality ablation

Input (SFT+RL)FZC ↑FOC ↑FOP ↑ACC ↑SUP ↑NAV ↑OOB ↓COL ↓
Full input82.586.365.769.168.8100.011.214.5
Without render75.483.164.660.765.899.714.317.8
Without scene state73.280.557.858.955.399.413.416.6
Without history79.583.764.867.068.1100.012.314.9
Text prompt only70.674.753.455.853.398.216.920.6

The ablations show that renders, scene state, and history each contribute to the full model’s performance. Removing renders only during RL yields 78.7 / 82.2 / 63.5 (FZC / FOC / FOP).

Q1/Q5. Evaluator robustness and human agreement

ScenePRM uses GPT-5.2. With evaluation inputs and prompts fixed, all three test VLMs gave the same method rankings for each metric.

Test evaluatorFuncRoom-Agent FZC / FOC / FOPRank on the three metrics
GPT-5.582.5 / 86.3 / 65.7First on all three
DeepSeek V4.1 Flash82.0 / 84.9 / 64.4First on all three
Gemini 3.5 Flash83.8 / 85.2 / 65.3First on all three
MetricMean agreement across participants ↑Mean Cohen’s κ ↑Kendall’s τ-b for method rankings ↑Spearman’s ρ across scenes ↑
FZC92.8%0.8490.8670.925
FOC93.3%0.8320.7330.937
FOP95.7%0.8411.0000.952

Fifteen participants independently assessed FZC, FOC, and FOP items. We computed item-level agreement and Cohen’s κ between GPT-5.5 and each participant, then averaged across participants. For each method, we computed Spearman’s ρ between GPT-5.5 and human-derived scores across 50 scenes, then pooled the six correlations using Fisher’s z transformation.

Q2. Baseline settings

Methods shared functional requirements, local hardware, final Blender settings, and five-view evaluation. Asset backends varied.

MethodInput adaptationLLM/VLMAsset backend
SAGE, SceneSmithDirect text inputGPT-5.5TRELLIS
SceneWeaverDirect text inputGPT-5.5Infinigen
SceneCodeDirect text inputGPT-5.5Blender Code
Code as RoomText-to-top-down-view interfaceGPT-5.5Blender Code
FuncRoom-AgentDirect text inputQwen3.6-35B-A3BTRELLIS

Q3. Runtime

Pipeline componentFuncRoom-AgentCode as Room
Planning and LLM/VLM generation5 min7 min
Asset generation and retrieval15 min10 min
Rendering and execution5 min3 min
Online evaluation and repair0 min13 min
End-to-end runtime25 min33 min
Runtime excluding asset generation and retrieval10 min23 min

Excluding asset generation and retrieval, runtime drops from 23 minutes to 10 minutes, a reduction of 56.5%. Online evaluation and repair account for the 13-minute gap. FuncRoom-Agent spends 5 min more on assets and saves 24.2% end to end.

Q4. Benchmark size

Function-Room statisticCount
Test prompts50
Room categories50
Annotated functional zones213
Annotated functional object slots670
Mean functional zones per prompt4.26
Mean functional object slots per prompt13.40
Evaluated methods6
Generated scenes300

We compare six methods on the same 50 human-reviewed prompts. Prior work reports 22 scenes (SceneThesis), 30 prompts (SceneCode). We also evaluate 30 SceneEval prompts and 531 held-out 3D-FRONT scenes.

Reviewer_uco9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Function-Room Generation, which generates executable 3D rooms from functional intent. It introduces (1) a recursive DSL with five hierarchical levels for room construction; (2) recursive construction traces are generated with a multimodal code agent and then used for SFT distillation into Qwen3.6-35B-A3B, followed by deterministic execution without test-time repair; (3) ScenePRM, which scores partial scenes on realism, functionality, layout, and prompt alignment to guide GRPO training. The method reports SOTA results on SceneEval and a self-constructed Function-Room benchmark.

Critical strengths
  1. Solid benchmark design. The 50 function-room prompts are manually validated, with metrics covering functional completeness and geometric usability.

  2. Consistent evidence for the proposed method. The recursive DSL encodes executable object dependencies, while 17,386 construction trajectories enable feed-forward generation without test-time repair. Improvements in Tables 3 and 4 support the effectiveness of recursive representation and process rewards.

Critical weaknesses
  1. Potential evaluation bias. FZC/FOC/FOP rely on an unspecified VLM, while ScenePRM uses GPT-5.2 for reward evaluation, risking alignment between training and evaluation preferences.

  2. Incomplete efficiency analysis. The 24.2% speedup may largely come from the asset pipeline (over 59% in Table.5) rather than eliminating test-time refinement.

Questions for authors
  1. Which VLM, model, and prompts are used for automatic functional evaluation? Is the evaluator independent of the GPT-5.2 critic used in ScenePRM? How well do automatic functional scores correlate with human judgments?

  2. Of the 25-minute runtime reported in Table 2, how much is spent on model calls, asset generation/retrieval, and execution, respectively? Are the asset sources, hardware, and timing protocol consistent with those used for Code as Room?

4 comments
AuthorsOpenReview ↗

Response to Reviewer uco9: Functional Evaluation and Runtime Breakdown

Thank you for the constructive review. We address evaluator bias and runtime below.

Q1. Functional evaluator, cross-VLM robustness, and human agreement

The Function-Room test evaluator uses GPT-5.5, while ScenePRM uses GPT-5.2 to score partial scenes during training. The test evaluator receives fixed, human-reviewed functional annotations and five rendered views (top, front, back, left, right). The prompt asks it to judge each item separately: whether a required zone is present (FZC), whether a valid object fills a required slot (FOC), and whether a generated object is relevant to the functional requirements (FOP). We will include the full prompt in the revision.

To check for shared model-family preferences, we reevaluated the same scenes with two additional VLMs, holding annotations, views, and instructions fixed. Each cell reports FZC / FOC / FOP.

MethodGPT-5.5DeepSeek V4.1 FlashGemini 3.5 Flash
SAGE43.6 / 48.7 / 56.544.3 / 47.5 / 52.644.7 / 48.2 / 53.3
SceneWeaver45.0 / 73.4 / 61.445.7 / 74.2 / 60.146.1 / 76.2 / 61.4
SceneSmith68.8 / 69.4 / 44.767.1 / 69.7 / 44.267.7 / 68.5 / 45.3
SceneCode47.5 / 50.3 / 58.346.8 / 47.9 / 58.848.9 / 51.6 / 60.1
Code as Room72.9 / 82.5 / 63.270.2 / 81.4 / 62.773.4 / 82.8 / 63.6
FuncRoom-Agent82.5 / 86.3 / 65.782.0 / 84.9 / 64.483.8 / 85.2 / 65.3

All three evaluators rank FuncRoom-Agent first on all three metrics. Across evaluators, it leads Code as Room by 9.6–11.8 percentage points in FZC and 2.4–3.8 in FOC.

MetricMean agreement across participants ↑Mean Cohen’s κ ↑Kendall’s τ-b for method rankings ↑Spearman’s ρ across scenes ↑
FZC92.8%0.8490.8670.925
FOC93.3%0.8320.7330.937
FOP95.7%0.8411.0000.952

Fifteen participants independently assessed FZC, FOC, and FOP items. We computed item-level agreement and Cohen’s κ between GPT-5.5 and each participant, then averaged across participants. For each method, we computed Spearman’s ρ between GPT-5.5 and human-derived scores across 50 scenes, then pooled the six correlations using Fisher’s z transformation.

We will clarify these model roles in the revision to avoid confusion.

Q2. Runtime breakdown and comparison settings

Both methods used the same functional requirements, local hardware, and final rendering and evaluation protocol. We timed each native pipeline end to end. FuncRoom-Agent took text directly and used TRELLIS for assets. Code as Room used its text-to-top-down-view interface and Blender Code. We report times with and without asset generation and retrieval to account for the different asset pipelines.

Pipeline componentFuncRoom-AgentCode as Room
Planning and LLM/VLM generation5 min7 min
Asset generation and retrieval15 min10 min
Rendering and execution5 min3 min
Online evaluation and repair0 min13 min
End-to-end runtime25 min33 min
Runtime excluding asset generation and retrieval10 min23 min

Excluding asset generation and retrieval for each method, runtime drops from 23 minutes to 10 minutes, a reduction of 56.5%. Planning, model generation, rendering, and execution total 10 minutes in both pipelines. The 13-minute gap comes from online evaluation and repair. FuncRoom-Agent spends 5 minutes more on assets and still reduces end-to-end runtime by 24.2%.

AuthorsOpenReview ↗

Follow-up on Rebuttal: Evaluation Reliability and Runtime Analysis

Dear Reviewer,

Thank you again for your positive and constructive assessment of our work.

We would like to kindly follow up on our November 14 rebuttal, where we provided additional cross-VLM evaluations, human agreement analyses, and a detailed runtime breakdown to address your questions.

As the discussion period is nearing its end, we would sincerely appreciate your feedback on whether our clarifications have addressed your concerns, or if there are any remaining questions we could help clarify.

Thank you for your time and consideration.

Dear AC,

Thank you for following up the discussion. The authors' additional evaluations and analyses have addressed most of my concerns.

Dear authors,

Thanks for the additional clarifications and experiments. Most of my concerns have now been addressed. therefore, I will maintain my score.

Reviewer_2L5HOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FuncRoom-Agent, a framework for generating functional 3D indoor scenes from natural-language requirements. It combines a recursive DSL, visual-context supervised fine-tuning, and ScenePRM-guided reinforcement learning to organize scene construction into dependent stages. At inference time, the model generates and executes code sequentially without online candidate search or iterative repair. Experiments demonstrate improvements in functional completeness, spatial quality, and generation efficiency.

Critical strengths

1. Practical relevance. The paper emphasizes functional usability rather than visual plausibility alone. The proposed Function-Room benchmark uses human-reviewed, fixed annotations to evaluate activity-specific functional zones and object requirements.

2. Coherent design. The recursive DSL captures spatial and support dependencies among scene objects. Conditioning subsequent generation on executed scene states and visual observations enables progressive construction of complex scenes.

3. Supporting experimental evidence. Ablations show that the complete ScenePRM training procedure outperforms the final-scene reward baseline on functional coverage and precision. An independent blinded user study provides complementary evidence of improved scene quality.

4. Improved inference efficiency. By shifting part of the planning and optimization process to training, the method avoids repeated online evaluation and repair. The reported full-scene generation time is 25 minutes, compared with 33 minutes for the fastest baseline.

Critical weaknesses

1. Insufficient attribution of ScenePRM improvements.

ScenePRM introduces both stage-level rewards and best-candidate continuation, selecting the highest-reward candidate among eight samples at each stage. Compared with the final-scene reward GRPO baseline, this changes reward granularity, candidate selection, and potentially computational budgets. Therefore, Table 4 does not sufficiently establish that the improvements primarily result from better process-level credit assignment.

2. Limited evidence of reliability without online repair.

The paper emphasizes single-candidate sequential construction but does not systematically report DSL execution success, full-scene completion rates, or failure handling. Despite strong functional metrics, the Function-Room results still show a collision rate of 14.5% and an out-of-bounds rate of 11.2%. The current evidence does not fully establish reliable construction without online repair.

3. Unclear independent contribution and novelty of the recursive DSL.

Prior work, such as HDSL, has explored hierarchical scene representations. Although Table 3 shows that recursive construction outperforms one-shot and staged construction, the comparison does not sufficiently control for generation steps, execution feedback, and computational budgets. The distinct contribution of the recursive representation over existing approaches therefore requires clearer justification.

4. Functional evaluation and baseline comparisons require further validation.

The paper uses human-reviewed annotations and an independent blinded user study, but agreement between automatic functional judgments and item-level human or geometric verification is not sufficiently quantified. Furthermore, baseline input adaptation, asset processing, and runtime configurations require clarification to establish fair quality and efficiency comparisons.

Questions for authors
  1. Can stage-level rewards and best-candidate continuation be evaluated separately under matched training budgets?
  2. What are the DSL execution success and full-scene completion rates across the 50 Function-Room test cases? How are execution failures, internal retries, and incomplete scenes handled?
  3. What is the main novelty of the recursive DSL compared with existing hierarchical representations? Could its independent contribution be evaluated under controlled generation and execution budgets?
  4. Could the authors report agreement between automatic functional metrics and human verification, and clarify baseline input adaptation and runtime configurations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2L5H: ScenePRM Ablation, Execution Reliability, DSL Novelty, and Evaluation

Thank you for the constructive review.

Q1. Stage rewards and candidate continuation

R1–R3 use final-scene reward GRPO. At equal fragment counts, R1 scores each final scene once, while R2 averages five scores per scene to match ScenePRM’s reward evaluation budget.

Training settingGenerated fragmentsReward evaluationsTraining timeFZC ↑FOC ↑FOP ↑
R1. Final-scene reward GRPO, matched fragment count360,00072,00020 h76.874.252.7
R2. Final-scene reward GRPO, matched fragment and reward evaluation counts360,000360,00024 h77.275.455.0
R3. Final-scene reward GRPO, matched training time450,000450,00030 h78.476.957.8
R4. Stage-level rewards with continuation from a randomly selected executable candidate360,000360,00030 h80.282.662.2
R5. Best-candidate trajectories + SFT360,000360,00018 h77.274.853.4
D. Full ScenePRM360,000360,00030 h82.586.365.7

R4 vs. R2 shows gains from introducing stage-level rewards and their associated rollout procedure at matched fragment and reward-evaluation counts. D vs. R4 quantifies the additional gain from best-candidate continuation at matched rewards and budgets. Across all three metrics, the majority of the observed gain from R2 to D comes from introducing stage-level rewards together with their associated rollout procedure. D vs. R3 shows that full ScenePRM improves all three metrics at the same 30 h training budget. D vs. R5 shows that SFT on best-candidate trajectories alone does not reproduce the full ScenePRM gains on any of the three metrics.

Q2. Execution without online repair

Scene generation success and geometric quality (COL and OOB) are distinct evaluation dimensions.

Function-Room execution statisticResult
Full-scene generation success rate100% (50/50)
DSL fragment parsing success rate100%
Execution success rate for successfully parsed fragments100%
Online candidate searches, failed-fragment regenerations, and online repairs0 each
Incomplete or discarded scenes0

Each step executes one candidate. Parsing and execution had zero failures, measured over generated and parsed fragments, respectively. All 50 scenes were evaluated.

MethodCOL ↓OOB ↓
SAGE16.813.9
SceneWeaver38.112.5
SceneSmith18.923.1
SceneCode15.633.7
Code as Room37.115.7
FuncRoom-Agent14.511.2

Q3. Main novelty of the Our DSL

Our contribution is not recursion alone: our DSL is designed for training, with independently executable fragments and intermediate scene states that enable us to collect execution-grounded traces for SFT and stage-level RL. HDSL targets structured scene generation and localized editing. After training, our agent generates and executes one fragment per step at inference without retries or iterative repair, while HDSL uses a bounded generate–verify–revise loop during subtree expansion.

We used the same agent and 50 prompts with comparable step and token counts.

Construction settingSteps per sceneGenerated tokens per sceneFZC ↑FOC ↑FOP ↑
Flat sequential construction with rendering feedback52209,52039.336.539.7
Recursive construction with rendering feedback54205,57249.558.840.4

The full systems used the same 50 prompts, hardware, and evaluation protocol.

MethodFZC ↑FOC ↑FOP ↑End-to-end time ↓
HDSL (our reimplementation)45.347.439.237 min
FuncRoom-Agent82.586.365.725 min

Q4. Human agreement and runtime

MetricMean agreement across participants ↑Mean Cohen’s κ ↑Kendall’s τ-b for method rankings ↑Spearman’s ρ across scenes ↑
FZC92.8%0.8490.8670.925
FOC93.3%0.8320.7330.937
FOP95.7%0.8411.0000.952

Fifteen participants independently assessed FZC, FOC, and FOP items. We computed item-level agreement and Cohen’s κ between GPT-5.5 and each participant, then averaged across participants. For each method, we computed Spearman’s ρ between GPT-5.5 and human-derived scores across 50 scenes, then pooled the six correlations using Fisher’s z transformation.

All methods received the same standard text prompts directly, except that Code as Room routed them through its official text-to-image interface to generate the required top-down views. Both pipelines shared functional requirements, local hardware, and the final rendering and evaluation protocol.

Pipeline componentFuncRoom-AgentCode as Room
Planning and LLM/VLM generation5 min7 min
Asset generation and retrieval15 min10 min
Rendering and execution5 min3 min
Online evaluation and repair0 min13 min
End-to-end time25 min33 min
Time excluding asset generation and retrieval10 min23 min

Excluding asset generation and retrieval for each method, runtime drops from 23 minutes to 10 minutes, a reduction of 56.5%.

AuthorsOpenReview ↗

Follow-up on Rebuttal: Additional Experiments and Clarifications

Dear Reviewer,

Thank you again for your constructive feedback.

We would like to kindly follow up on our rebuttal submitted on November 14. We have conducted additional controlled experiments addressing your major concerns, particularly the attribution of ScenePRM improvements, execution reliability without online repair, and the independent contribution of our recursive DSL.

As the discussion period is approaching its end, we would greatly appreciate your feedback on whether these additional results have sufficiently addressed your concerns, or whether any important issues remain unresolved.

Thank you very much for your time and consideration.

Dear Authors,

Thank you for your detailed response and for more additional experiments. The new results and clarifications helped address my concerns, and I plan to raise my score accordingly.

Best regards, Reviewer

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.