FLOORA: A Human-Aligned Domain-Specific Language Model for Architectural Design

Sahand Rezaei-Shoshtari; Patryk Wozniczka; Shu Ishida; Gregg Streuber; Farnoosh Javadi; Jeffrey Landes; Angela Ju; Muhammad Azam

cs.LGSep 28, 2026

FLOORA: A Human-Aligned Domain-Specific Language Model for Architectural Design

Authors: Sahand Rezaei-Shoshtari, Patryk Wozniczka, Shu Ishida, Gregg Streuber, Farnoosh Javadi, Jeffrey Landes, Angela Ju, Muhammad Azam, +9 more

Organizations: Autodesk Research

Abstract

Foundation models are powerful generators, but many engineering domains require structured representations that general-purpose systems handle poorly. We introduce FLOORA (Floor Layout Optimization with RL Alignment), a family of small domain-specific language (DSL) models for architectural layout generation. With specialized data and alignment, our 0.6B model outperforms much larger frontier models, achieving VLM judge win rates up to 92.0% on out-of-distribution real-world buildings and 96.0% on synthetic buildings. Human evaluations further corroborate these results, with FLOORA selected as the best model in 89.3% of evaluations. FLOORA combines a token-efficient DSL, custom tokenization, domain-specific pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) with learned human-preference and verifiable rewards. This pipeline improves architectural and geometric validity, supported by extensive empirical evaluation and ablation studies. Although focused on architecture, our results suggest that similar domain-specific recipes may be useful in other engineering domains with structured, verifiable outputs. Datasets, models, and inference code are available at https://github.com/AutodeskAILab/floora.

Figures & tables

Figure 1: Overview of our DSL pre-training, human feedback collection, and post-training pipeline.

Figure 2: Qualitative comparison of FLOORA-0.6B and frontier models on a real-world test sample. FLOORA outperforms the baselines on geometric and functional checks and architectural quality.

Panel of Figure 3 — Figure 3: High-level recipe for training and post-training a domain-specific language model.

Panel of Figure 4 — Figure 4: (a) t-SNE visualization of massing polygons across synthetic, OSM, and SFT datasets, indicating the distribution shift between synthetic and real-world massing geometries. (b) Tokenized sequence length distributions for the pre-training dataset using the DSL tokenizer and the original Qwen3 tokenizer, showing that the DSL tokenizer reduces sequence lengths by 54%, on average.

Panel of Figure 5 — Figure 5: Structured prompt and completion format used for DSL pre-training and post-training.

Panel of Figure 6 — Figure 6: Pass@5 for FLOORA models across model sizes and training stages, where success requires all checks to pass. Shading shows 95% CIs across 3 training and 5 evaluation seeds.

Figure 7: Reward model validation accuracy, reported with 95% CIs.

Figure 8: Pairwise VLM judge outcomes for each GRPO (RM+VR) model against its matched GRPO (VR only) counterpart. Bars show the fraction of prompts where RM+VR wins, ties, or loses. As model size increases, the RM+VR variant is increasingly preferred. Error bars indicate 95% Wilson intervals over the evaluation prompts.

Panel of Figure 9 — Figure 9: (a) Pass@k comparison of FLOORA-0.6B (Qwen3) and frontier baselines, where success requires passing both geometric and functional checks. (b) Pairwise VLM judge outcomes for FLOORA-0.6B (Qwen3) against frontier baselines. Bars show the fraction of prompts for which FLOORA wins, ties, or loses. Our model substantially outperforms all evaluated frontier baselines.

Figure 10: Human evaluation results across 100 test inputs with 95% CIs. FLOORA is selected as the best model in 89.3% of evaluations.

Appendix figures & tables41 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 11 — Figure 11: Examples of DSL representation of multifamily residential buildings, illustrating the massing, vertical cores, living units, corridors, and corresponding 2D floor-plan layout.

Stage	Description	Samples	Share of corpus
1,2	TileGPT-derived layouts with scale jitter	3.9M	95%
3	Inference + repair (admitted)	0.2M	5%
Total before rotation augmentation		4.1M	100%
Total after rotation augmentation		82M

Appendix

Table 1: Synthetic corpus composition by generation stage. Counts are before rotation augmentation unless noted.

Parameter	Value
Boundary sample points $N$	64
Maximum samples per dataset	10,000
t-SNE perplexity	40
t-SNE iterations	1,000
t-SNE initialization	PCA

Appendix

Table 2: Hyperparameters used for the t-SNE analysis of massing footprints.

Figure 12: Samples from the synthetic dataset. Each sample contains the building massing, metadata, and a procedurally generated space layout consisting of cores, corridors, and living units.

Figure 13: Samples from the OSM dataset. Each sample contains only the building massing and metadata, without ground-truth space layouts, and therefore cannot be used for pre-training.

Figure 14: Samples from the SFT dataset. Each sample contains the building massing, metadata, and an architect-edited space layout consisting of living units and, when needed, cores and corridors.

Figure 16: Model-generated outputs are ranked relative to one another, with ties permitted.

Figure 17: The highest-ranked model-generated output is subsequently edited by the labeler.

Category	Reason	Count
Skipped	Rating $-1$	61
Skipped	Inconsistent pairs	575
Skipped	Failed geometric checks	287
Errors	Failed DSL encodings	227
Total		1,150

Appendix

Table 3: Number of human feedback datapoints removed during data filtering and processing.

Dataset	Before Aug.	After Aug.
SFT completions	10.4k	135.4k
Pairwise preferences	90.6k	1.2M

Appendix

Table 4: Number of datapoints in the human-feedback dataset after filtering, before and after rotation augmentation.

Tokenizer	Mean	Median	P95	Max	Before Aug.	After Aug.
DSL Tokenizer	235.9	219.0	373.0	785	880M	17.6B
Qwen3 Tokenizer	516.6	475.0	844.0	1,902	1.9B	38.5B

Appendix

Table 5: Tokenization statistics by tokenizer. “Before Aug.” and “After Aug.” denote total token counts before and after augmentation.

Base Model	Effective Params	Learning Rate
Qwen3-0.6B	440M	3e-4
Qwen3-1.7B	1.4B	4e-4
Pythia-70M	20M	2e-4
Pythia-160M	90M	6e-5 1 1 1 A lower learning rate was used because training was unstable at higher learning rates.
Pythia-410M	310M	3e-4
Pythia-1.4B	1.2B	3e-4

Appendix

Table 6: Effective parameter counts after resizing the embedding layers to match the DSL tokenizer vocabulary, and peak learning rates used during pre-training.

Model	SFT	RM	GRPO
Qwen3-0.6B	2e-4	8e-5	6e-5
Qwen3-1.7B	2e-4	3e-4 2 2 2 A higher learning rate was used because the model was under-trained at lower learning rates.	8e-5
Pythia-410M	2e-4	8e-5	4e-5
Pythia-1.4B	2e-4	8e-5	4e-5

Appendix

Table 7: Peak learning rates used during post-training.

Panel of Figure 18 — Figure 18: Example VLM judge reasoning for representative pairwise outcomes.

Panel of Figure 19 — Figure 19: Pass@1/3/5 performance of the base models on OSM and synthetic test sets. Pass@k is considered achieved only when both the geometric correctness and functional compliance checks pass. Shaded regions represent 95% CIs across 5 evaluation seeds.

Panel of Figure 20 — Figure 20: Pre-training validation across model scales. (a) Final validation loss achieved by each model. (b) Validation loss throughout pre-training.

Panel of Figure 21 — Figure 21: SFT validation loss across model scales. (a) Final validation loss achieved by each model. (b) Validation loss throughout training. Shaded regions represent 95% CIs across 3 training seeds.

Panel of Figure 22 — Figure 22: Reward model validation accuracy across model scales. (a) Final validation accuracy achieved by each model. (b) Validation accuracy throughout reward model training. Shaded regions represent 95% CIs across 3 training seeds.

Panel of Figure 23 — Figure 23: Learning curves for GRPO reward components across model scales, including geometric correctness, functional compliance, the normalized reward model score, and the soft overlong penalty. Shaded regions show 95% CIs across 3 training seeds.

Panel of Figure 24 — Figure 24: Pass@1/3/5 performance of the base and fine-tuned models on OSM and synthetic test sets. Pass@k is considered achieved only when both the geometric correctness and functional compliance checks pass. Shaded regions represent 95% CIs across 3 training and 5 evaluation seeds.

Panel of Figure 25 — Figure 25: Pairwise VLM judge outcomes for each FLOORA GRPO (RM+VR) model against its matched GRPO (VR only) counterpart. Bars show the fraction of prompts where RM+VR wins, ties, or loses. As model size increases, the RM+VR variant is increasingly preferred. Error bars indicate 95% Wilson intervals over the evaluation prompts.

	Functional Compliance			Geometric Correctness			Total Reward
Model	@1	@3	@5	@1	@3	@5	@1	@3	@5
FLOORA-0.6B (Qwen3) – Base	34.2 $\pm$ 0.4	51.2 $\pm$ 0.8	58.8 $\pm$ 1.2	27.8 $\pm$ 0.3	38.5 $\pm$ 0.6	43.2 $\pm$ 0.8	23.8 $\pm$ 0.1	33.4 $\pm$ 0.4	38.1 $\pm$ 0.7
FLOORA-1.7B (Qwen3) – Base	35.6 $\pm$ 0.4	53.3 $\pm$ 0.5	61.2 $\pm$ 0.4	28.7 $\pm$ 0.1	40.5 $\pm$ 0.4	45.5 $\pm$ 0.6	24.9 $\pm$ 0.3	35.8 $\pm$ 0.6	40.7 $\pm$ 0.9
FLOORA-70M (Pythia) – Base	26.6 $\pm$ 0.4	47.6 $\pm$ 0.6	56.8 $\pm$ 0.8	4.3 $\pm$ 0.2	8.7 $\pm$ 0.5	11.1 $\pm$ 0.6	4.2 $\pm$ 0.2	8.6 $\pm$ 0.4	11.0 $\pm$ 0.6
FLOORA-160M (Pythia) – Base	26.3 $\pm$ 0.3	43.8 $\pm$ 0.5	52.3 $\pm$ 0.8	13.1 $\pm$ 0.2	21.3 $\pm$ 0.2	24.9 $\pm$ 0.2	12.4 $\pm$ 0.1	20.2 $\pm$ 0.3	23.7 $\pm$ 0.5
FLOORA-410M (Pythia) – Base	34.6 $\pm$ 0.4	52.3 $\pm$ 0.6	60.6 $\pm$ 0.7	26.7 $\pm$ 0.3	36.9 $\pm$ 0.5	41.5 $\pm$ 0.5	23.0 $\pm$ 0.1	32.2 $\pm$ 0.4	36.8 $\pm$ 0.7

Appendix

Table 8: Pass@1/3/5 performance of the base models on the OSM test set, reported with 95% CIs across 5 evaluation seeds. Best value within each model family is shown in bold.

	Functional Compliance			Geometric Correctness			Total Reward
Model	@1	@3	@5	@1	@3	@5	@1	@3	@5
FLOORA-0.6B (Qwen3) – Base	90.9 $\pm$ 0.9	94.3 $\pm$ 0.9	95.2 $\pm$ 0.9	96.9 $\pm$ 0.4	98.4 $\pm$ 0.3	98.5 $\pm$ 0.3	90.2 $\pm$ 1.0	94.0 $\pm$ 1.0	94.9 $\pm$ 1.0
FLOORA-1.7B (Qwen3) – Base	91.3 $\pm$ 0.9	94.2 $\pm$ 0.8	95.0 $\pm$ 0.8	97.2 $\pm$ 0.4	98.4 $\pm$ 0.3	98.5 $\pm$ 0.3	90.7 $\pm$ 0.9	93.9 $\pm$ 0.8	94.7 $\pm$ 0.8
FLOORA-70M (Pythia) – Base	70.2 $\pm$ 1.0	90.6 $\pm$ 0.7	93.8 $\pm$ 0.6	32.0 $\pm$ 0.6	54.8 $\pm$ 0.8	64.0 $\pm$ 0.7	31.0 $\pm$ 0.6	53.2 $\pm$ 0.9	62.2 $\pm$ 0.9
FLOORA-160M (Pythia) – Base	79.2 $\pm$ 0.6	91.9 $\pm$ 0.6	93.8 $\pm$ 0.5	69.6 $\pm$ 0.5	88.3 $\pm$ 0.6	92.3 $\pm$ 0.8	66.1 $\pm$ 0.3	84.7 $\pm$ 0.5	88.9 $\pm$ 0.9
FLOORA-410M (Pythia) – Base	91.1 $\pm$ 0.7	94.7 $\pm$ 0.7	95.6 $\pm$ 0.7	96.0 $\pm$ 0.3	98.3 $\pm$ 0.3	98.4 $\pm$ 0.3	89.6 $\pm$ 0.7	94.2 $\pm$ 0.7	95.1 $\pm$ 0.8

Appendix

Table 9: Pass@1/3/5 performance of the base models on the synthetic test set, reported with 95% CIs across 5 evaluation seeds. Best value within each model family is shown in bold.

	Functional Compliance			G eometric Correctness			Total Reward			$\Delta$ Total Reward (vs Base)
Model	@1	@3	@5	@1	@3	@5	@1	@3	@5	@1	@3	@5
FLOORA-0.6B (Qwen3) – SFT	45.5 $\pm$ 1.0	73.4 $\pm$ 0.9	82.7 $\pm$ 0.8	45.1 $\pm$ 0.9	71.3 $\pm$ 1.0	79.6 $\pm$ 0.8	37.4 $\pm$ 1.0	63.9 $\pm$ 1.2	73.9 $\pm$ 1.1	+13.7	+30.5	+35.8
FLOORA-0.6B (Qwen3) – GRPO (RM+VR)	85.6 $\pm$ 0.3	95.6 $\pm$ 0.2	97.5 $\pm$ 0.2	80.9 $\pm$ 0.7	92.3 $\pm$ 0.5	94.8 $\pm$ 0.5	79.3 $\pm$ 0.7	91.3 $\pm$ 0.5	94.0 $\pm$ 0.5	+55.5	+57.9	+55.9
FLOORA-0.6B (Qwen3) – GRPO (VR only)	84.0 $\pm$ 0.7	95.3 $\pm$ 0.4	97.2 $\pm$ 0.3	76.3 $\pm$ 0.5	90.1 $\pm$ 0.3	93.5 $\pm$ 0.3	74.7 $\pm$ 0.7	89.0 $\pm$ 0.5	92.5 $\pm$ 0.5	+50.9	+55.5	+54.3
FLOORA-1.7B (Qwen3) – SFT	46.1 $\pm$ 0.2	73.5 $\pm$ 0.3	82.5 $\pm$ 0.3	46.6 $\pm$ 0.2	72.1 $\pm$ 0.3	79.9 $\pm$ 0.3	38.8 $\pm$ 0.2	64.8 $\pm$ 0.3	74.2 $\pm$ 0.3	+13.9	+29.0	+33.5
FLOORA-1.7B (Qwen3) – GRPO (RM+VR)	85.8 $\pm$ 0.4	95.7 $\pm$ 0.2	97.4 $\pm$ 0.2	78.2 $\pm$ 0.1	90.5 $\pm$ 0.2	93.4 $\pm$ 0.2	76.9 $\pm$ 0.1	89.7 $\pm$ 0.2	92.8 $\pm$ 0.3	+52.0	+54.0	+52.0

Appendix

Table 10: Pass@1/3/5 performance of the post-trained models on the OSM test set, reported with 95% CIs across 3 training and 5 evaluation seeds. The final column reports percentage-point changes in Total Reward relative to the base checkpoint. Best absolute value within each model family is shown in bold.

	Functional Compliance			Geometric Correctness			Total Reward			$\Delta$ Total Reward (vs Base)
Model	@1	@3	@5	@1	@3	@5	@1	@3	@5	@1	@3	@5
FLOORA-0.6B (Qwen3) – SFT	57.0 $\pm$ 1.3	86.5 $\pm$ 0.9	93.6 $\pm$ 0.5	49.3 $\pm$ 1.5	79.7 $\pm$ 1.4	88.8 $\pm$ 0.9	48.7 $\pm$ 1.5	79.2 $\pm$ 1.4	88.5 $\pm$ 0.9	-41.5	-14.8	-6.4
FLOORA-0.6B (Qwen3) – GRPO (RM+VR)	97.4 $\pm$ 0.2	98.5 $\pm$ 0.1	98.6 $\pm$ 0.1	96.8 $\pm$ 0.2	98.2 $\pm$ 0.2	98.3 $\pm$ 0.1	96.8 $\pm$ 0.2	98.1 $\pm$ 0.2	98.3 $\pm$ 0.2	+6.6	+4.1	+3.3
FLOORA-0.6B (Qwen3) – GRPO (VR only)	97.9 $\pm$ 0.2	98.4 $\pm$ 0.1	98.5 $\pm$ 0.1	97.5 $\pm$ 0.1	98.2 $\pm$ 0.1	98.3 $\pm$ 0.1	97.3 $\pm$ 0.1	98.1 $\pm$ 0.2	98.2 $\pm$ 0.2	+7.1	+4.0	+3.2
FLOORA-1.7B (Qwen3) – SFT	57.1 $\pm$ 0.3	86.0 $\pm$ 0.3	92.9 $\pm$ 0.3	50.7 $\pm$ 0.3	80.3 $\pm$ 0.3	88.9 $\pm$ 0.2	50.1 $\pm$ 0.3	79.9 $\pm$ 0.3	88.7 $\pm$ 0.2	-40.6	-14.0	-6.0
FLOORA-1.7B (Qwen3) – GRPO (RM+VR)	97.6 $\pm$ 0.2	98.5 $\pm$ 0.1	98.6 $\pm$ 0.1	97.0 $\pm$ 0.2	98.2 $\pm$ 0.1	98.3 $\pm$ 0.1	97.0 $\pm$ 0.2	98.1 $\pm$ 0.1	98.2 $\pm$ 0.1	+6.3	+4.2	+3.6

Appendix

Table 11: Pass@1/3/5 performance of the post-trained models on the synthetic test set, reported with 95% CIs across 3 training and 5 evaluation seeds. The final column reports percentage-point changes in Total Reward relative to the base checkpoint. Best absolute value within each model family is shown in bold.

	Functional Compliance			Geometric Correctness			Total Reward
Model	@1	@3	@5	@1	@3	@5	@1	@3	@5
FLOORA-0.6B (Qwen3) – GRPO (RM+VR)	84.8	95.3	97.2	79.5	91.6	94.0	78.1	90.8	93.5
Claude Opus 4.8	46.1	73.6	83.0	36.4	60.8	72.0	24.7	45.0	55.6
Claude Sonnet 4.6	19.3	38.4	49.4	10.0	20.2	26.8	5.1	10.7	14.5
Gemini 2.5 Flash Image	37.6	62.0	71.7	25.5	44.3	53.8	11.1	20.3	26.0
Gemini 3.5 Flash	52.6	76.6	83.9	42.0	69.1	79.7	33.3	57.4	68.2

Appendix

Table 12: Pass@1/3/5 comparison of the FLOORA-0.6B (Qwen3) – GRPO (RM+VR) and frontier baselines on the OSM test set. Best absolute value is shown in bold. Results are obtained on a single seed.

	Functional Compliance			Geometric Correctness			Total Reward
Model	@1	@3	@5	@1	@3	@5	@1	@3	@5
FLOORA-0.6B (Qwen3) – GRPO (RM+VR)	97.4	98.3	98.4	96.5	97.8	98.0	96.5	97.7	97.9
Claude Opus 4.8	66.5	92.2	97.1	38.6	62.9	72.9	30.9	54.2	64.5
Claude Sonnet 4.6	39.0	70.2	82.5	8.1	15.5	20.4	6.9	13.5	17.9
Gemini 2.5 Flash Image	64.1	90.7	96.7	15.6	29.7	38.0	12.9	24.0	30.6
Gemini 3.5 Flash	75.3	96.4	99.2	39.9	66.2	77.0	39.1	65.2	76.0

Appendix

Table 13: Pass@1/3/5 comparison of the FLOORA-0.6B (Qwen3) – GRPO (RM+VR) model and frontier baselines on the synthetic test set. Best absolute value is shown in bold. Results are obtained on a single seed.

Panel of Figure 26 — Figure 26: Pass@k comparison of the FLOORA-0.6B (Qwen3) GRPO (RM+VR) model and frontier baselines, where success requires passing both geometric and functional checks. Our model substantially outperforms all evaluated frontier baselines.

Panel of Figure 27 — Figure 27: Pairwise VLM judge outcomes for the FLOORA-0.6B (Qwen3) GRPO (RM+VR) model against frontier baselines. Bars show the fraction of prompts for which the FLOORA model wins, ties, or loses. Our model is preferred over all evaluated frontier baselines. Error bars indicate 95% Wilson intervals over the evaluation prompts.

Panel of Figure 28 — Figure 28: Qualitative comparison between the FLOORA-0.6B (Qwen3) GRPO (RM+VR) model and frontier models on the OSM test set.

Panel of Figure 29 — Figure 29: Qualitative comparison between the FLOORA-0.6B (Qwen3) GRPO (RM+VR) model and frontier models on the synthetic test set.

Panel of Figure 30 — Figure 30: Qualitative comparison between the FLOORA-0.6B (Qwen3) Base, SFT, GRPO (RM+VR), and GRPO (VR only) models on the OSM test set.

Panel of Figure 31 — Figure 31: Qualitative comparison between the FLOORA-0.6B (Qwen3) Base, SFT, GRPO (RM+VR), and GRPO (VR only) models on the synthetic test set.

Panel of Figure 32 — Figure 32: Ablation of the RM normalization scale $\alpha$ during GRPO training with the FLOORA-0.6B (Qwen3) model. Moderate scales $\alpha\in[1,4]$ stabilize training and achieve high verifiable rewards. Shaded regions show 95% CIs across 3 training seeds.

Setting	Reward Model	Geometric Correctness	Functional Compliance	Soft Overlong Penalty
No normalization	28.719 $\pm$ 5.376	0.383 $\pm$ 0.192	0.761 $\pm$ 0.224	0.000 $\pm$ 0.000
$\alpha=1$	0.994 $\pm$ 0.002	0.853 $\pm$ 0.026	0.951 $\pm$ 0.016	0.000 $\pm$ 0.000
$\alpha=2$	0.990 $\pm$ 0.012	0.846 $\pm$ 0.017	0.944 $\pm$ 0.002	0.000 $\pm$ 0.000
$\alpha=3$	0.988 $\pm$ 0.011	0.851 $\pm$ 0.027	0.948 $\pm$ 0.007	0.000 $\pm$ 0.000
$\alpha=4$	0.982 $\pm$ 0.015	0.850 $\pm$ 0.017	0.943 $\pm$ 0.012	0.000 $\pm$ 0.000
$\alpha=10$	0.954 $\pm$ 0.049	0.771 $\pm$ 0.032	0.922 $\pm$ 0.026	0.000 $\pm$ 0.000

Appendix

Table 14: Final RM and verifiable reward scores for different normalization scales

\alpha

, reported with 95% CIs across 3 training seeds.

Panel of Figure 33 — Figure 33: Tokenizer ablation for FLOORA-0.6B (Qwen3) over optimizer steps. Shaded regions show 95% CIs across 3 training seeds, except pre-training which uses a single seed.

Panel of Figure 34 — Figure 34: Tokenizer ablation during GRPO (RM+VR) training with the FLOORA-0.6B (Qwen3) model. Shaded regions show 95% CIs across 3 training seeds.

	Total Reward
Model	@1	@3	@5
FLOORA-0.6B (Qwen3) – Base [DSL tok.]	23.8 $\pm$ 0.1	33.4 $\pm$ 0.4	38.1 $\pm$ 0.7
FLOORA-0.6B (Qwen3) – Base [Orig. tok.]	19.6 $\pm$ 0.4	27.3 $\pm$ 0.5	31.1 $\pm$ 0.5
FLOORA-0.6B (Qwen3) – GRPO (RM+VR) [DSL tok.]	79.3 $\pm$ 0.7	91.3 $\pm$ 0.5	94.0 $\pm$ 0.5
FLOORA-0.6B (Qwen3) – GRPO (RM+VR) [Orig. tok.]	71.6 $\pm$ 2.9	83.2 $\pm$ 2.0	86.8 $\pm$ 1.6
FLOORA-0.6B (Qwen3) – GRPO (VR only) [DSL tok.]	74.7 $\pm$ 0.7	89.0 $\pm$ 0.5	92.5 $\pm$ 0.5

Appendix

Table 15: Pass@1/3/5 performance of FLOORA-0.6B (Qwen3) across training stages using the original and DSL tokenizers on the OSM test set, reported with 95% CIs across 3 training and 5 evaluation seeds. Best absolute value is shown in bold.

	Total Reward
Model	@1	@3	@5
FLOORA-0.6B (Qwen3) – Base [DSL tok.]	90.2 $\pm$ 1.0	94.0 $\pm$ 1.0	94.9 $\pm$ 1.0
FLOORA-0.6B (Qwen3) – Base [Orig. tok.]	88.9 $\pm$ 0.7	95.1 $\pm$ 0.6	96.2 $\pm$ 0.6
FLOORA-0.6B (Qwen3) – GRPO (RM+VR) [DSL tok.]	96.8 $\pm$ 0.2	98.1 $\pm$ 0.2	98.3 $\pm$ 0.2
FLOORA-0.6B (Qwen3) – GRPO (RM+VR) [Orig. tok.]	93.5 $\pm$ 1.3	97.6 $\pm$ 0.5	98.4 $\pm$ 0.3
FLOORA-0.6B (Qwen3) – GRPO (VR only) [DSL tok.]	97.3 $\pm$ 0.1	98.1 $\pm$ 0.2	98.2 $\pm$ 0.2

Appendix

Table 16: Pass@1/3/5 performance of FLOORA-0.6B (Qwen3) across training stages using the original and DSL tokenizers on the synthetic test set, reported with 95% CIs across 3 training and 5 evaluation seeds. Best absolute value is shown in bold.

Panel of Figure 35 — Figure 35: Human feedback data-scale ablation for FLOORA-0.6B (Qwen3) across SFT and GRPO (RM+VR) on the OSM and synthetic test sets. Shaded regions show 95% CIs across 3 training seeds and 5 evaluation seeds.

Explore similar work

Learning an Interior Layout Policy in a Domain Specific Language Action Space
Jul 31, 2026Yuhao Lu, Weichen Zhang, Wenyi Xiao +23D Layout Generation Layouts
Taobao & Tmall Group of Alibaba
Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans
Jun 9, 2026Fedor Rodionov, Aleksandar Cvejic, Michael Birsak +2Floor Plan Generation Affordances
King Abdullah University of Science and Technology (KAUST), Saudi Arabia · Miami University, United States of America
Design-MLLM: A Reinforcement Alignment Framework for Verifiable and Aesthetic Interior Design
Mar 4, 2026Yuxuan Yang, Xiaotong Mao, Jingyao WangMllm-Based Multimodal Large Language Models
Nanjing Forestry University, Nanjing, China · Université de Lorraine, Nancy, France · Laboratoire Réactions et Génie des Procédés, Nancy, France +2

Jul 31, 2026, cs.CVJ/K move · Enter open · S save

Learning an Interior Layout Policy in a Domain Specific Language Action Space

Yuhao Lu, Weichen Zhang, Wenyi Xiao, Haohui Chen+1

Taobao & Tmall Group of Alibaba

3D Layout Generation Layouts Floor Plan Generation Learning Policy Large Language Model Policy Optimization Spatial Reasoning+1

Open paper AskPDF arXiv

Indoor scene layout generation is a challenging task in interior design. Existing methods often oversimplify the task by reducing room conditions to coarse 3D bounding boxes and neglecting structural elements such as doors and windows. More fundamentally, many prior approaches formulate spatial reasoning as direct coordinate prediction, thereby casting interior layout design as continuous regression over raw geometric parameters, which hinders the model from learning the underlying reasoning logic of intelligent layout design. We propose \textbf{LayoutDSL}, a novel LLM-based framework for learning an interior layout policy in a domain-specific language (DSL) action space. The DSL provides an explicit symbolic representation of layout information and serves as a structured action space for layout reasoning, where each action corresponds to an interpretable design decision. Under this DSL-based policy learning paradigm, we construct 3D-FrontDSL, a dataset of room-structure annotations paired with synthetic DSL action sequences for supervised fine-tuning. To promote a more generalizable and scalable policy with verifiable feedback, we design rewards grounded in interior design principles and physical plausibility, and optimize the policy via reinforcement learning. Extensive experiments demonstrate that LayoutDSL substantially improves spatial plausibility and design logicality over strong baselines and existing methods.

Abstract

Explore similar work

Learning an Interior Layout Policy in a Domain Specific Language Action Space

Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

Design-MLLM: A Reinforcement Alignment Framework for Verifiable and Aesthetic Interior Design