acceptodds
Under review as a conference paper at ICLR 2027

LivingFlow: Agentic Reinforcement Learning for Circulation-Aware Scene Refinement

Abstract

Comfortable living spaces require both functional furniture arrangements and smooth circulation—the movement of people between functional locations during everyday activities. A collision-free layout can still force people through tight passages or unnecessary detours, even when every destination is reachable. We introduce LivingFlow, an agentic reinforcement learning framework that iteratively rearranges furniture in existing indoor layouts. To specify circulation requirements, a frozen large language model generates natural-language activity chains describing how people use the space throughout a day; transitions between successive uses define fixed movement tasks that a geometric executor evaluates for furniture access, clearance, and route efficiency as the layout changes. Our framework adapts Group-in-Group Policy Optimization to optimize a trainable editing policy over complete tool cycles, connecting information gathering and editing decisions with executable feedback. Experiments on original 3D-FRONT layouts show improved layout quality and circulation over existing refiners, and confirm the benefits of activity-conditioned evaluation and executable feedback.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Agentic Reinforcement Learning, Indoor Layout, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YteJOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

LivingFlow optimizes existing 3D foreground furniture rooms to improve the circulation layout of movement/navigation-oriented interior scenes. An initial scene layout is generated using LLM, followed by optimization through editing operations such as asset positioning and scaling. A trainable Qwen3.8-27B editor queries registered assets and then edits or freezes optimized assets for candidates requiring movement, rotation, and resizing. After each edit, the geometry executor re-evaluates the effectiveness of reachability and path constraints. To efficiently solve the problem, this paper combines the GiGPO adaptive algorithm with a well-designed reward rule to optimize the model's layout adjustment capabilities.

Critical strengths

The paper develops a framework called LIVINGFLOW, which introduces a furniture layout optimization framework that senses pedestrian flow. By incorporating geometric feedback based on activity context, it improves existing layouts, thus addressing to some extent a problem in the layout design field regarding how scenes can support downstream tasks. The article is clearly structured, and the charts provide strong visual support.

Critical weaknesses
  1. LivingFlow can resize furniture, but the paper doesn't explicitly state whether LayoutVLM and other baseline methods can change object dimensions. If some input layouts can only be fixed by resizing, then baseline methods limited to movement and rotation might not solve these problems. Therefore, some performance gaps may stem from additional operations rather than the proposed methods themselves.

  2. The paper mentions that the 2D layout is fixed, and paths are recalculated after each edit. Recalculating paths doesn't necessarily change the required passage width. Is there a minimum width, or does it depend on the embodied or user common sense? This is crucial for clarifying that the optimization goal is to achieve functional circulation, not just a passable path.

  3. The paper showcases living rooms, living-dining rooms, and bedrooms, but it doesn't extend or discuss other indoor scenarios that require more circulation distribution, such as kitchens. This raises some concerns about the method's scalability.

Questions for authors
  1. Compared with previous reachability optimization methods, such as PhyScene [1] and RoomPlanner [2], and agent-based approaches like SAGE [3], this work integrates these ideas into a unified framework and introduces RL-based optimization. It would be helpful if the authors could further clarify the key technical contributions and explain how the proposed approach advances beyond the combination of existing methods.

  2. Is the required circulation width fixed across different scenarios, or does it vary depending on the activity and user? What factors determine whether a route is considered acceptable? It would be helpful to clarify how the proposed criteria capture the convenience and comfort of movement rather than simply ensuring that a valid path exists.

[1]:physcene:PhyScene: Physically Interactable 3D Scene Synthesis for Embodied AI [2]:RoomPlanner: Reachability-Aware View Sampling for Text-to-Room 3D Gaussian Splatting [3]:SAGE: Scalable Agentic 3D Scene Generation for Embodied AI

2 comments
AuthorsOpenReview ↗

Official Response to Reviewer YteJ (1/2)

Thank you for asking us to clarify the editing permissions and passage-width criterion, and for suggesting that we evaluate LivingFlow in kitchens.

Weakness 1: Potential advantage from different resizing permissions

Response. We added an audit of LayoutVLM’s edit permissions (Appendix D.2). LayoutVLM optimizes translation and yaw, and all 1,510 audited items retain their original dimensions. RoomPlanner, SAGE, PhyScene, OID-PPO, and HAISOR also use fixed furniture sizes in our evaluations.

With furniture dimensions fixed, the main-table LivingFlow checkpoint reaches 53.76% MPR and LayoutVLM reaches 37.01%, a difference of 16.75 percentage points. We applied the same ranking rule to each method’s available layouts, including the input. LayoutVLM contributes its returned layout, and LivingFlow contributes the states saved during editing. The two methods also have different permissions for other edits. The shared-executor searches in Table 2 use the same move, rotate, and resize operations as LivingFlow.

Method with fixed furniture dimensionsMPR (%)SR (%)Valid (%)
LayoutVLM37.016.8424.74
LivingFlow (Resize disabled at inference)53.7614.2161.05

Table 2 also reports the fixed-size policy trained and evaluated without Resize, with layouts ranked by the same rule as LivingFlow.

Weakness 2 & Question 2: Fixed passage width and its relation to comfort

Response. We recalculate routes after each edit using the same 60 cm clearance threshold across activities and rooms. A movement passes when the required furniture access locations can be selected normally and a route connects them at the specified clearance (Section 3.2 and Appendix A.3). The threshold approximates the onset of shoulder rotation in the aperture experiment cited in Section 4.5.

We added a passage-width sensitivity analysis that varies the threshold while holding the layouts and movement tasks from Tables 1 and 3 fixed (Section 4.5, Figure 4). The scores at 60 cm match the main comparisons.

RoomsWidths testedLivingFlow MPR relative to input
3D-FRONT living rooms, living–dining rooms, and bedrooms50–90 cmHigher at every tested width
HSSD kitchens and bathrooms40–80 cmHigher at every tested width; at 80 cm, the gains are 0.43 and 2.19 points, respectively

The main gain is that more previously infeasible movements become feasible. Among movements that passed in both the input and refined layouts, paths were 2.91 cm shorter on average (Appendix F.1, Table 8). The blinded study reports how two participants judged comfort for everyday use from floor plans (Table 9).

AuthorsOpenReview ↗

Official Response to Reviewer YteJ (2/2)

Weakness 3: Limited evidence of generalization to kitchens

Response. We added a transfer experiment on 47 kitchens and 114 bathrooms from HSSD (Section 4.4, Table 3). We applied the same 3D-FRONT LivingFlow policy to both room types without further training. All baselines were adapted to the same editing task, with nine fine-tuned on HSSD. LivingFlow leads on MPR, SR, and Valid in both room types. Relative to the input, MPR improves by 8.04 percentage points in kitchens and 7.23 in bathrooms. Bathroom SR also rises from 12.28% to 25.44%.

All values are percentages.

MethodKitchen MPRKitchen SRKitchen ValidBathroom MPRBathroom SRBathroom Valid
Input layout31.572.132.1356.2012.2827.19
DiffuScene16.272.132.1340.886.1417.54
LEGO-Net20.942.132.1347.137.8918.42
InstructScene11.942.132.1336.497.0220.18
LayoutVLM4.140.000.001.940.004.39
SceneReVis28.792.132.1345.479.6522.81
ReSpace12.910.000.0014.390.880.88
RoomPlanner19.032.134.2636.7113.1635.09
SAGE29.892.134.2638.347.8928.07
PhyScene13.282.132.1340.146.1416.67
OID-PPO22.952.132.1328.103.5114.04
HAISOR29.612.132.1350.4411.4025.44
iARCS16.042.132.1341.446.7316.08
LivingFlow39.614.2610.6463.4325.4444.74

Question 1: Technical contribution beyond existing reachability and agent-based methods

Response. LivingFlow refines existing furnished layouts by turning transitions between consecutive furniture uses in activity chains into fixed movement tasks and recomputing access and routes after each edit to guide the next decision (Sections 3.1–3.3). Our GiGPO adaptation assigns credit over a complete tool cycle that groups the editor's queries with the resulting edit or stop decision (Section 3.5). Table 2 compares editing strategies under the shared executor.

Activity ablation. We added the w/o Activity comparison in Table 2, using the same return rule as LivingFlow. The ablated editor sees geometric feedback during editing. The fixed evaluation movements are used to rank its saved layouts after editing. In living rooms, w/o Activity reaches 85.71% Valid and 37.71% MPR, compared with 83.67% Valid and 60.33% MPR for LivingFlow. Geometrically valid layouts can still leave required movements infeasible.

3D-FRONT room typeLivingFlow MPR / SR / Valid (%)w/o Activity MPR / SR / Valid (%)
Living room60.33 / 16.33 / 83.6737.71 / 4.08 / 85.71
Living–dining room74.73 / 24.00 / 92.0051.37 / 8.00 / 90.00
Bedroom74.14 / 41.76 / 97.8043.13 / 19.78 / 85.71
Reviewer_ng4COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This method refines existing furniture layouts so that people can move between the furniture they use during daily activities.

A frozen LLM generates activity chains that define movement tasks, a geometric executor checks access and walking routes after every edit and an LLM editor trained with a GiGPO variant moves, rotates or resizes one object at a time.

Experiments on 190 3D-FRONT rooms report higher movement pass rate and geometric validity than six layout methods.

Critical strengths

It intuitively makes sense to evaluate a layout by whether people can move between the furniture they actually use, rather than by collisions and reachability alone, and the activity-chain formulation makes this verifiable.

The ablations in table 2 cover the right factors (activity choice, circulation feedback, sequential interaction, RL). the w/o RL control uses the same prompts, tools and budgets as the trained model, so the RL gain is cleanly isolated.

Critical weaknesses
  1. Lacking evidence that shows the refined rooms are better

The test movement tasks come from the same frozen GPT and circulation is scored by the authors’ own geometric rules and thresholds (table 3). Nothing checks whether the edited rooms look sensible and functional. A small user study comparing input, w/o RL, and LivingFlow layouts would make the central claim credible.

  1. Missing baselines that does geometry and walkability optimization:

MPR and SR are scored on walking routes plus geometry but none of the baselines optimizes walking space or accessibility. The comparison therefore makes it hard to know whether the proposed approach is actually better. Methods that do optimize free space, reachability, or walkability exist and they are not compared against. For example: HAISOR (Sun et al., 2024) and PhyScene (Yang et al., 2024a), OID-PPO (Yoon et al., 2025), iARCS (Adhikari et al., 2026). Running one of these on the same movement tasks would be a fair test of the activity-chain idea.

References

  1. Sun et al., HAISOR: Human-Aware Indoor Scene Optimization via Deep Reinforcement Learning, ACM Transactions on Graphics, 2024.
  2. Yang et al., PhyScene: Physically Interactable 3D Scene Synthesis for Embodied AI, CVPR 2024.
  3. Yoon et al., OID-PPO: Optimal Interior Design using Proximal Policy Optimization by Transforming Design Guidelines into Reward Functions, arXiv 2508.00364, 2025.
  4. Adhikari et al., iARCS: Iterative Agentic RL for Controllable 3D Scene Generation, arXiv 2608.06161, 2026.
Questions for authors

NA

3 comments
AuthorsOpenReview ↗

Official Response to Reviewer ng4C

Reviewer ng4C

Thank you for suggesting a user study and pointing us to baselines that optimize walkability and accessibility. We added both to the revised manuscript.

Weakness 1: Limited evidence of everyday usability beyond the shared evaluator

Response. We added a blinded floor-plan study of perceived everyday comfort (Appendix F.1, Table 9). Both participants chose LivingFlow more often than the comparator in each comparison. One lay participant and one designer independently compared anonymous, same-scale floor plans from the 3D-FRONT test split, using the LivingFlow layouts reported in Table 1. Each participant saw one pair per room per comparison, with shuffled order and balanced left/right placement.

ParticipantCompared with LivingFlowLivingFlow chosenComparator chosenAbout the sameUnsure
Lay participantInput layout11820502
DesignerInput layout11032480
Lay participantLivingFlow (w/o RL)62486020
DesignerLivingFlow (w/o RL)8054506

Preference margins were larger against the input layouts. The study measures perceived comfort from floor plans.

Held-out movements and the evaluator. We added a held-out movement evaluation on frozen layouts (Appendix F.2, Table 10). The holdout tests transfer to new movement requirements using the same task generator and geometric checks. The interval excludes zero for the subset with no shared movement legs.

Held-out movement subsetRoomsTasksw/o RL MPR (%)LivingFlow MPR (%)Difference (pp) [paired 95% CI]
All held-out movements1901,85773.8877.35+3.47 [−0.24, +7.32]
No movement legs shared with editing tasks17293471.8676.89+5.02 [+0.66, +9.55]

Weakness 2: Missing baselines that optimize walkability or accessibility

Response. We added HAISOR, PhyScene, OID-PPO, and iARCS to the comparison on the same 3D-FRONT test rooms and movement requirements (Table 1). LivingFlow leads these baselines on MPR, SR, and Valid in each room type. Appendix D.2 gives the baseline configurations. OBR and CNR report out-of-bounds and collision rates.

Each cell reports OBR / CNR / MPR / SR / Valid in percent.

MethodLiving roomLiving–dining roomBedroom
HAISOR8.66 / 25.81 / 29.75 / 2.04 / 18.375.01 / 27.62 / 45.77 / 10.00 / 26.0019.94 / 30.13 / 29.72 / 7.69 / 18.68
PhyScene38.77 / 38.80 / 16.90 / 0.00 / 0.0025.34 / 48.25 / 19.13 / 0.00 / 2.0054.53 / 33.12 / 15.17 / 0.00 / 0.00
OID-PPO9.86 / 37.63 / 21.84 / 0.00 / 10.205.42 / 29.29 / 36.20 / 2.00 / 16.0023.59 / 43.37 / 18.84 / 3.30 / 12.09
iARCS38.53 / 67.80 / 10.26 / 0.00 / 0.0054.91 / 35.32 / 16.93 / 0.00 / 0.0036.25 / 42.16 / 14.07 / 1.10 / 6.59
LivingFlow2.42 / 11.37 / 60.33 / 16.33 / 83.671.32 / 15.23 / 74.73 / 24.00 / 92.000.44 / 2.84 / 74.14 / 41.76 / 97.80

We also checked whether the result held when each baseline could return the input or another available candidate. We applied the same ranking rule to the available candidates for each method on the same test rooms. LivingFlow uses the same layouts as Table 1.

MethodOBR (%)CNR (%)MPR (%)SR (%)Valid (%)
HAISOR15.3324.2638.547.3725.79
PhyScene21.1324.5239.436.3223.68
OID-PPO15.7924.6338.246.8425.79
iARCS18.9025.5638.006.3224.21
LivingFlow1.188.3070.7330.5392.63

The search controls in Table 2 give all policies access to the same activities and editing feedback.

I would like to thank authors for adding other baselines.

Two concerns remain:

  1. The human study has two raters. Against the untrained editor, the lay rater's result is 62 wins, 48 losses, which is not distinguishable from chance. Only the designer shows a preference, and no agreement statistic or test is reported. This does not yet show that the refined rooms are better. It should be validated with larger population.

  2. In Table 1, PhyScene and iARCS reach close to ~0% in all room types where the untouched input is 20% valid. A method that optimizes physical plausibility having 0% valid raises concern. This points to a problem in how the baselines were adapted.

So, I would like to keep my rating.

AuthorsOpenReview ↗

Official Response to Reviewer ng4C

Thank you for the follow-up.

  1. Both participants showed a clear preference for LivingFlow over the original layouts. Against w/o RL, both participants also chose LivingFlow more often, although the preference margins were smaller. Professional experience from designer may influence more sensitivity to differences in furniture arrangement and everyday usability, potentially contributing to the observed preference pattern.

  2. Regarding PhyScene and iARCS, refining an existing furnished layout introduces additional preservation and consistency requirements. A method must improve the requested property while respecting the supplied architecture and furniture inventory. Changing one placement can improve one relationship while introducing a collision, boundary violation, or doorway obstruction elsewhere. LEGO-Net explicitly distinguishes rearrangement from generation and documents boundary violations and object penetrations after rearrangement. SceneReVis similarly reports substantial geometric violations for some baselines in goal-oriented scene optimization.

Our Valid metric requires all hard geometric checks to pass, including unobstructed doorways. This makes whole-room validity sensitive to newly introduced violations. LivingFlow’s advantage also persists when the baselines can retain the original layout: applying a common ranking criterion to each method’s available candidates, including the input, gives PhyScene and iARCS 23.68% and 24.21% Valid, compared with 92.63% for LivingFlow. Their corresponding MPR values are 39.43%, 38.00%, and 70.73%.

Reviewer_rSzdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LivingFlow, an agentic RL framework that refines existing furnished rooms to improve activity-driven circulation, i.e., movement between functional locations during everyday activities. A frozen GPT-5.5 generates cached per-room activity chains whose transitions define fixed movement tasks. After every move/rotate/resize edit, a geometric executor re-evaluates access and routes using grid A* with a 60 cm passage width plus hard boundary, overlap and door-keepout checks. A Qwen3.8-27B editor is trained with a GiGPO adaptation that treats the queries and the closing edit or Finish as one "complete tool cycle", rewarded by hand-designed execution-grounded rewards. On 190 original 3D-FRONT test rooms, LivingFlow reports the best boundary and collision rates, movement pass rate (MPR), success rate and validity against six adapted baselines (MPR +27–44 pp over the unedited input), with ablations on activity targets, sequential interaction and RL training.

Critical strengths
  1. A practical, under-addressed objective with a carefully engineered environment. Judging a layout by whether the movements of everyday activities can actually be executed goes beyond collision-free placement. The executor also closes several reward-hacking channels: actions must be registered, credit requires fresh and comparable evidence, and Finish is verified by re-execution. The full prompts, thresholds and training settings are given in the appendix.

  2. Matched controls and candid reporting. The w/o RL control shares initialization, prompts, tools and budgets. The sequential-vs-one-shot ablation shows large gaps (e.g., Bedroom MPR 74.14 vs. 31.64). The evaluation keeps failed episodes and initially passing rooms, and the appendix states its limits openly (e.g., the unestimated bounds in Prop. 3 and the per-category reversals in Fig. 6).

Critical weaknesses
  1. The headline results do not match the paper's own final checkpoint. Appendix E.2 evaluates the update-20 checkpoint (the full budget in Table 6) of the single training run on the same 190 rooms, with the same action budget and validity criterion. It finds 151/190 valid layouts (35/49, 40/50, 76/91). Tables 1–2, however, report LivingFlow Valid of 83.67/92.00/97.80, i.e., 176/190. The Base checkpoint reproduces the w/o RL row exactly (131/190), so the two protocols agree, yet no reported checkpoint reaches the headline row. Under that consistent protocol the RL validity gain roughly halves (+20 vs. +45 rooms). MPR and SR are computed on the same returned layouts, so they are equally unverified, and all results come from one run with greedy decoding and no variance.

  2. The evaluation does not isolate the claimed contributions. At test time LivingFlow queries the same executor and cached specification that compute MPR/SR/Valid. The six baselines see neither, and they are worse than the unedited input on Valid and CNR in all 18 entries. A non-learned search over the same interface and executor (as in the cited Yu et al., 2011, or HAISOR) is missing. Even within the paper's own controls, the untrained agent already obtains 68–89% of the MPR gain over the input. In Bedroom, RL does not change the share of valid layouts in which every movement passes (38/89 = 43% vs. 27/61 = 44% w/o RL; input 8/19 = 42%), so the SR gain tracks geometric repair rather than better circulation. Route quality (detours, path length), the stated motivation, is never measured. Edit magnitude is not measured either, although the shown examples shrink furniture to roughly 55–65% of its original area (Fig. 8).

Questions for authors
  1. Which run, checkpoint and returned-layout rule produced Tables 1–2, and how do they relate to the update-20 checkpoint in App. E.2 (151/190 valid)? Please report MPR/SR/Valid for that checkpoint and across several training and evaluation seeds.

  2. How does LivingFlow compare with a non-learned search (e.g., simulated annealing or executor-selected best-of-N) over the same move/rotate/resize interface and executor objective? How does LivingFlow score on activity specifications it did not query at test time?

  3. How much do the gains depend on resizing? Please report displacement and resize statistics, route-quality metrics (path length, detour ratio), and results with Resize disabled.

4 comments
AuthorsOpenReview ↗

Official Response to Reviewer rSzd (1/2)

Thank you for examining how we obtained the main results and for asking for comparisons that separate the effects of the editing interface, training, and resizing.

Weakness 1 & Question 1: Discrepancy with the final checkpoint and variation across seeds

Response. We evaluated every checkpoint from Update 0 through Update 20 of the seed-42 run on all 70 validation rooms. Update 9 achieved the highest validation MPR (68.16%) and Valid (100%), with an SR of 20%. We therefore selected Update 9 for evaluation on the 190 test rooms. Tables 1 and 2 report its ranked layouts. The original Appendix E.2 reported the final layouts at Update 20 from the same run.

Seed-42 checkpointReturn ruleMPR (%)SR (%)Valid (%)
Update 9Ranked state within one trajectory70.7330.5392.63 (176/190)
Update 9Final state64.3723.6876.32 (145/190)
Update 20Ranked state within one trajectory67.7628.9589.47 (170/190)
Update 20Final state61.8923.1679.47 (151/190)

At Update 9, ranking the saved states adds 31 valid rooms relative to returning the final state. Under the same ranking rule, Update 9 has six more valid rooms than Update 20. We rank the input and saved states within one trajectory by geometric validity, then by the number of passing movements, and return the earliest state in a tie. WebGPT also uses candidate ranking at inference time (Nakano et al., 2021). WebGPT

Training seeds. To assess how much the headline scores vary across training runs, we report Update 9 for seeds 42, 43, and 44 using the same layout return rule as Tables 1 and 2. Seed 42 provides the reference for the main-table results.

Training seedMPR (%)SR (%)Valid (%)
4270.7330.5392.63
4368.7329.4792.63
4469.4530.5391.05

Inference-seed reruns. We reran LivingFlow on all 190 rooms in the 3D-FRONT test split with inference seeds 42, 43, and 44. All three reruns used the checkpoint from Tables 1 and 2 (training seed 42, Update 9), greedy decoding, and the same layout return rule as the main tables. Inference-seed reruns. We reran LivingFlow on all 190 rooms in the 3D-FRONT test split with inference seeds 42, 43, and 44. All three reruns used the checkpoint from Tables 1 and 2 (training seed 42, Update 9), greedy decoding, and the same layout return rule as the main tables.

Inference seedMPR (%)SR (%)Valid (%)
4268.1728.4291.05
4367.4331.5888.42
4467.7630.5389.47

Across the three reruns, the observed ranges spanned 0.74 percentage points in MPR, 3.16 in SR, and 2.63 in Valid. In 162 rooms (85.26%) from our earlier paired reruns, at least 90% of the furniture in each room differed by at most 20 cm in position, 15° in orientation, and 15% in size along either axis.

Weakness 2 & Question 2: Missing shared-executor comparisons and evidence on held-out movements

Response. We added random search, greedy search, and simulated annealing under a shared executor (Section 4.3, Table 2). LivingFlow exceeds greedy search by 12.14 MPR points (paired 95% CI [8.55, 15.77]) and leads all three search baselines on MPR, SR, and Valid in every room type. All methods started from the same rooms and used the same activities, move/rotate/resize interface, and executor. Search used three runs per room, each with up to 15 candidate executions and 120 menu queries. The table aggregates rooms equally.

Random search accepts every successfully executed edit. Greedy search and simulated annealing use a score that rewards geometric validity and the fraction of passing movements and penalizes geometric violations. The return rule ranks the input and all successfully executed candidates by validity, then passing movement count, with ties going to the earliest state.

PolicyMPR (%)SR (%)Valid (%)
Random search48.918.4232.11
Greedy search58.5913.5140.70
Simulated annealing55.6811.0534.39
LivingFlow70.7330.5392.63

Movements held out during editing. We added a held-out movement evaluation on frozen layouts (Appendix F.2, Table 10). LivingFlow improves MPR on tasks with no movement legs shared with the editing tasks.

Held-out movement subsetRoomsTasksw/o RL MPR (%)LivingFlow MPR (%)Difference (pp) [paired 95% CI]
All held-out movements1901,85773.8877.35+3.47 [−0.24, +7.32]
No movement legs shared with editing tasks17293471.8676.89+5.02 [+0.66, +9.55]
AuthorsOpenReview ↗

Official Response to Reviewer rSzd (2/2)

Task-specific RL and geometric repair.

We added a comparison between LivingFlow and its untuned editor under the same return rule (Appendix F.3, Table 11). The untuned editor already reaches 67.50% MPR. The clearest gain associated with training is geometric validity.

PolicyOBR (%)CNR (%)MPR (%)SR (%)Valid (%)
LivingFlow without task-specific RL3.4010.2267.5026.3287.89
LivingFlow1.188.3070.7330.5392.63

Question 3: Contribution of resizing, route quality, and edit magnitude

Response. We added measurements of movement coverage, path length, and detour ratio (Appendix F.1, Table 8). The main gain is that more required movements become feasible. Path length and detour ratio are measured on the 1,463 movements that pass in both layouts.

MeasureResult
Passing in the input layout1,544 / 4,414
Passing after LivingFlow refinement3,093 / 4,414
Newly passing1,630
No longer passing81
Mean path-length change on paired routes−2.91 cm
Mean detour-ratio change on paired routes+0.017

Furniture edits. We added measurements of displacement, resizing, and footprint change by room type (Appendix F.4, Table 12).

Room typeMean displacement (cm)Items resized (%)Summed footprint change (%)
Living room15.3917.92−14.58
Living–dining room14.0315.22−13.33
Bedroom23.2534.06−23.63
Overall18.8024.94−18.59

Effect of Resize. We reran the main-table LivingFlow checkpoint with Resize disabled at inference, using the same layout return rule. MPR dropped by 16.97 points. We added this comparison to Appendix F.4 (Table 13).

LivingFlow configurationMPR (%)SR (%)Valid (%)
Resize enabled70.7330.5392.63
Resize disabled at inference53.7614.2161.05

Thank you for the detailed responses. Selecting the best validation checkpoint, reporting ranked versus final states, and adding seed variation, shared-executor search baselines, held-out movements, path-length/detour statistics, edit magnitude, and the Resize-disabled ablation address the main factual points raised in my previous comments.

These additions help clarify the results. However, the headline tables still depend on an intermediate checkpoint and a ranking rule that recovers many valid layouts only after the trajectory ends, while final states are substantially weaker. The untuned editor already captures most of the movement-pass-rate gain, and disabling Resize at inference produces a large drop, indicating that a substantial part of the improvement comes from reducing object footprints rather than from rearrangement or the learned policy alone. The contribution of the trained policy and the complete-cycle objective therefore remains only partially isolated. I thus maintain the original score of weak rejection.

AuthorsOpenReview ↗

Official Response to Reviewer rSzd

Dear Reviewer,

Thank you for clarifying your concerns. We respectfully wish to clarify three points.

Trajectory selection preserves solutions already reached during execution. Every selected layout is an actual visited state. Training benefits also persist in terminal outputs, with geometric validity increasing from 131/190 rooms for the untuned editor to 151/190 at Update 20. Three Update-20 runs achieve best-saved MPR of 67.76%–69.73%, showing that improvements persist beyond the intermediate checkpoint.

Fixed-footprint rearrangement provides substantial gains. The separately trained w/o Resize variant disables resizing during both training and inference. It improves MPR over the input layouts by 25.76, 17.11, and 14.84 percentage points across the three room categories. These results directly demonstrate effective refinement with fixed object footprints. The inference-time Resize removal experiment measures the full policy’s dependence on that action, including its interactions with other edits.

Policy training contributes under matched tools and selection. Training increases MPR from 67.50% to 70.73% and SR from 26.32% to 30.53%. On held-out movement tasks sharing no movement legs with those used during editing and selection, MPR increases from 71.86% to 76.89%. These comparisons establish an additional contribution from learning under the shared interface and return rule.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their constructive feedback. In response to your suggestions, we added the following experiments and analyses to the revised manuscript and clarified how we obtained the main results.

  1. We clarified the discrepancy in geometric validity (Appendix E.2). Tables 1 and 2 report the highest-ranked layouts at Update 9, with 176/190 geometrically valid rooms (92.63%). The original Appendix E.2 reported the final layouts at Update 20, with 151/190 valid rooms (79.47%). Both checkpoints came from the same training run. In the rebuttal below, we explain how we selected Update 9 using the 70-room validation split and report results across three training seeds and three inference seeds.

  2. We expanded the baseline comparison and added search controls under a shared executor (§§4.2–4.3, Tables 1–2). Random search, greedy search, and simulated annealing use the same initial layouts, movement tasks, executor, and editing operations as LivingFlow. On the 3D-FRONT test split, LivingFlow exceeds greedy search, the strongest search baseline, by 12.14 percentage points in movement pass rate (MPR). The new activity ablation also shows that a geometrically valid layout can still block everyday movements. Removing activity guidance yields one more valid living room (42/49 versus 41/49), while MPR falls from 60.33% to 37.71%.

  3. We added kitchens and bathrooms as new evaluation scenarios (§§4.4–4.5, Table 3 and Figure 4). We evaluated LivingFlow and the baselines on 47 kitchens and 114 bathrooms from HSSD, where cabinets and fixed fixtures constrain everyday routes. The 3D-FRONT-trained LivingFlow policy improves MPR over the input by 8.04 points in kitchens and 7.23 points in bathrooms without further training. We also tested how the gains change as the required passage width increases across all five room types.

  4. We measured what the edits change for everyday use (Appendices F.1, F.2, and F.4). Of 4,414 movement tasks, 1,630 became feasible after refinement and 81 ceased to be feasible. For home layout planning, this means more of the routes needed for everyday activities become usable. We also tested held-out movement tasks, measured furniture displacement and resizing, and reran the main checkpoint with Resize disabled. In a blinded floor-plan assessment, both participants chose LivingFlow more often than the comparison layout in each pairwise comparison when judging comfort for everyday use.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.