acceptodds
Under review as a conference paper at ICLR 2027

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

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 included in the supplemental.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
domain-specific language models, rlhf, rlvr

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YCxmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents FLOORA. It is a small domain‑specific language‑model family for multi‑family residential architectural floor‑layout generation. The authors design a compact, parsable domain‑specific language to represent building massing and floor‑plan geometry, which is far more token‑efficient than standard IFC BIM formats. The authors release datasets, model weights and inference code, and suggest this domain‑specialized recipe may transfer to other engineering structured‑generation tasks.

Critical strengths
  1. Most prior architectural generative work uses either pure rule‑based optimization or only human‑preference reward. However, this paper carefully normalizes reward‑model outputs to prevent learned reward from overwhelming hard constraint rewards, achieving both architectural subjective quality and formal geometric validity.

  2. The authors find that reward‑model benefits depend on model capacity, which is an important empirical observation for domain‑specific RLHF.

  3. Compared with verbose industry IFC standards, the DSL achieves ~15× text compression and 54% shorter average token sequences.

Critical weaknesses
  1. While labeling protocols and filtering steps are thoroughly described, the raw human feedback dataset (SFT edited layouts and pairwise preference data) is not released.

  2. The paper hypothesizes the recipe generalizes to other engineering domains, but there is no preliminary empirical cross‑domain evidence to back this claim.

  3. Readers cannot assess how transferable this full pipeline would be for other structured engineering generation problems.

Questions for authors
  1. For the reward‑model normalization hyper‑parameter α=3 selected in GRPO: in practical deployment across different model sizes, do you observe that the optimal α value shifts with model capacity? Could you give practical guidance for picking α when adapting this hybrid‑reward setup to new domain tasks?

  2. Regarding the unreleased human feedback dataset: are there licensing or privacy constraints preventing release? Would it be feasible to release anonymized?

Reviewer_oF8hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FLOORA, a family of small language models that generate floor-plan layouts for multifamily residential buildings from a given building massing. It introduces a compact text-based domain-specific language for building geometry, with a parser that validates outputs and converts them back to geometry. Models are pretrained on synthetic data, fine-tuned on architect-edited layouts, and then aligned with reinforcement learning that combines a learned architect-preference reward and rule-based validity rewards.

Critical strengths
  1. The DSL is compact, human-readable, and works with a deterministic parser. The tokenizer ablation is concrete: sequences are much shorter, training cost drops substantially, and OSM pass@5 improves by about 6.5–7.2 points.

  2. Small models are cheap to deploy and the FLOORA family of small domain specific models are an important contribution.

Critical weaknesses
  1. The pass@k metric uses the same verifier logic as the RL reward and the data-admission filter. The VLM judge’s prompt penalizes overlaps and out-of-massing spaces, which are among the validity constraints FLOORA is trained to optimize.

  2. Gemini 3.5 Flash is both the judge and one of the compared baselines, creating possible self-preference effects.

Questions for authors

Comment: The figures and associated texts are way to small to read Please increase the sizes.

Question: Can the VLM-judge results be reproduced with a judge that is not among the compared models, or with a multi-judge ensemble? What is the judge-architect agreement specifically on FLOORA-vs-frontier pairs?

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

This paper presents FLOORA, a domain-specialized language-model framework for generating multifamily residential floor layouts from building footprints. It uses a compact, parseable architectural DSL, large-scale synthetic pre-training, architect-edited supervised fine-tuning (SFT), and GRPO with verifiable geometric/functional rewards and a learned architect-preference reward model. The authors evaluate Qwen3- and Pythia-based models across sizes on synthetic inputs and real-world OpenStreetMap building footprints. They report constraint-based pass@k, VLM pairwise preferences, and a blinded study with independent architects. Results show substantial post-training gains and strong performance from a compact 0.6B model, including against the tested frontier general-purpose models under the specified DSL prompting protocol.

Critical strengths
  1. The paper presents a coherent and well-developed pipeline combining a compact architectural DSL, domain-specific pre-training, architect feedback, and reinforcement learning. The tokenizer ablation demonstrates substantial efficiency gains (54% shorter sequences and 66% fewer GPU-hours).
  2. The evaluation is extensive, covering multiple model families, real-world building footprints, reward ablations, and independent architect assessments. The strong performance of a 0.6B model on this specialized task is a meaningful empirical result.
Critical weaknesses
  1. The contribution of expert SFT to the final GRPO performance is not fully isolated. All GRPO variants are initialized from SFT checkpoints. A Base → GRPO (VR-only) ablation would help clarify the additional contribution of architect-edited supervision to the final performance.
  2. The frontier-model comparison does not fully separate DSL compliance from layout quality. Although the paper includes VLM and human evaluations, it remains unclear how much of the automatic performance gap arises from DSL parsing failures rather than layout-quality failures. Reporting parsing success rates and performance conditional on successful parsing would help clarify this distinction.
Questions for authors
  1. SFT contribution: Have the authors evaluated GRPO (VR-only) initialized directly from the base model under a matched training budget? If so, how does it compare with GRPO initialized from the SFT checkpoint?
  2. Frontier-model comparison: What are the DSL parsing success rates of the frontier baselines? How much of the constraint-performance gap remains when restricting the comparison to successfully parsed layouts?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.