cs.CLMar 18, 2026
SaveText-to-Stage: Spatial Layouts from Long-form Narratives
Organizations: Rice University · Meta FAIR · Meta Reality Labs Research
Abstract
In this work, we probe the ability of a language model to demonstrate spatial reasoning from unstructured text, mimicking human capabilities and automating a process that benefits many downstream media applications. Concretely, we study the narrative-to-play task: inferring stage-play layouts (scenes, speaker positions, movements, and room types) from text that lacks explicit spatial, positional, or relational cues. We then introduce a dramaturgy-inspired deterministic evaluation suite and, finally, a training and inference recipe that combines rejection SFT using Best-of-N sampling with RL from verifiable rewards via GRPO. Experiments on a text-only corpus of classical English literature demonstrate improvements over vanilla models across multiple metrics (character attribution, spatial plausibility, and movement economy), as well as alignment with an LLM-as-a-judge and subjective human preferences.
Figures & tables
Figure 1: We propose a Dramaturg Spatializer that transforms long-form narrative into structured stage layouts. The model assigns dialogue and characters to canonical positions on a discrete grid—depth ( front, middle, back ) and lateral placement ( left, center, right )—with temporal blocking to produce plausible staging from text. This automates a “text-to-play” process for applications such as game previsualization and spatial audiobooks.
Figure 2: Overview of our Spatializer training pipeline. Left: A strong teacher LLM produces pseudo stage-play annotations via Best-of- sampling and rejection SFT under our deterministic dramaturgy evaluator. Right: The resulting Spatializer-SFT is post-trained with RL from verifiable rewards (GRPO): for each passage, we sample candidate spatializations, score them with the deterministic evaluator, compute group-relative advantages (with KL regularization to a frozen reference model), and update the policy to obtain Spatializer-GRPO . The model outputs a reasoning trace (e.g., per-quote placement plan) and the corresponding stage-grid layout.
| Model | Deterministic suite components | AVG | |||||
| Quote Attr. | Alias Res. | Stage Pos. | Char. Pos. | Move. Coh. | Scene Trans. | ||
| Proprietary models | |||||||
| Claude Sonnet 3.5 | 96.0 | 97.0 | 87.0 | 85.0 | 77.0 | 51.0 | 82.0 |
| GPT-4.1 | 94.0 | 99.0 | 65.0 | 84.0 | 90.0 | 51.0 | 81.0 |
| GPT-4o | 95.0 | 99.0 | 68.0 | 83.0 | 84.0 | 51.0 | 80.0 |
| Claude Sonnet 4 | 96.0 | 97.0 | 85.0 | 83.0 | 69.0 | 50.0 | 80.0 |
Table 1: Leaderboard across models. Metrics include Quote Attr. (Quote Attribution), Alias Res. (Alias Resolution), Stage Pos. (Stage Positioning Validity), Char. Pos. (Character Positioning), Move. Coh. (Movement Coherence), and Scene Trans. (Scene Transitions). Top scores are bolded.
Figure 3: Additional analyses. (a) Distribution of per-example AVG scores over GRPO, SFT, and Baseline models. RL training shifts mass and reduces degenerate failures, indicating improved reliability beyond mean score gains. (b) Deterministic macro score (%) versus GRPO steps when swapping the backbone (Qwen3 vs. LLaMA 3.3); the horizontal line denotes the best proprietary API baseline (83.45%). (c) ROC curve for a logistic regression that predicts whether listeners prefer audio over from the deterministic dramaturgy score difference . The classifier achieves with Brier score .
Figure 4: Challenging qualitative cases for spatialization diagnostics. Left: case probing consistency between generated reasoning and emitted layout. Right: case probing continuity when new characters enter an already-populated scene.
| Model | Quote Attr. | Alias Res. | Stage Pos. | Char. Pos. | Move. Coh. | Scene Trans. | AVG |
| Spatializer-GRPO | 91.75 | 99.09 | 92.11 | 92.54 | 98.25 | 57.40 | 88.52 |
| -Only JSON | 82.40 | 93.00 | 76.62 | 83.00 | 89.91 | 48.03 | 78.83 |
| -Quote+Alias | 27.02 | 31.57 | 51.00 | 72.00 | 75.00 | 39.00 | 49.26 |
| -Stage Pos | 95.19 | 99.23 | 76.00 | 73.74 | 69.00 | 51.07 | 77.37 |
| -Character | 94.00 | 99.00 | 66.45 | 57.76 | 65.95 | 49.00 | 72.03 |
| -Movements | 93.87 | 99.37 | 77.74 | 88.02 | 65.00 | 51.07 | 79.18 |
Table 2: Reward-component ablations for the 8B spatializer trained with GRPO. We retrain variants using only JSON validity as reward or by removing one deterministic reward category at a time, and then evaluate using the full deterministic suite. Scores are percentages (higher is better). AVG is the macro-average over Quote Attr., Alias Res., Stage Pos., Char. Pos., Move. Coh., and Scene Trans. (excluding Valid JSON).
| Evaluator | Agreement with human pref. | |
| Spearman | Rank Acc. | |
| LLM-as-Judge (Abs.) | 0.88 | 86.7 |
| LLM-as-Judge (Rel.) | 0.86 | 86.7 |
| Deterministic | 0.95 | 93.3 |
Table 3: External validity of automated evaluators against human pairwise preferences on binaural renderings. We report Spearman correlation ( ) between evaluator-induced system rankings and human win-rate rankings, and rank accuracy (Rank Acc.) over all system pairs. Best results are shown in bold .
| Model | Dramaturgy Prompt | Minimal Prompt |
| Zero-shot models | ||
| Llama-4-Scout | 83.39% | 83.25% ↓0.14 |
| Llama-3.3-70B-Instruct | 83.39% | 82.05% ↓1.34 |
| Llama-4-Maverick | 83.14% | 82.07% ↓1.07 |
| gemma-3-12b-it | 80.00% | 78.31% ↓1.69 |
| Qwen3-8b | 78.16% | 77.82% ↓0.34 |
Table 4: Prompt-sensitivity ablation. We compare the deterministic macro score (AVG) under our full prompt (explicit dramaturgy constraints) versus a minimal prompt that gives only a high-level task description. Drop is computed as (dramaturgy prompt) (minimal prompt), so larger values indicate stronger dependence on explicit prompt constraints.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B.1: Extra results: test-time selection and training-data scaling. (a) Best-of- at inference time for Qwen3-8B : we sample candidate spatializations per passage and select the highest-scoring one using three ranking signals— Same Model (self-judging with the LLM-as-a-judge prompt), Bigger Model (GPT-4.1 as judge with the same prompt), and our deterministic dramaturgy evaluator ( Oracle ). Increasing improves the selected output for all scorers, with the largest gains under the deterministic oracle. (b) Data scaling for Spatializer-SFT : we fine-tune the same base model with increasing numbers of supervised training examples and evaluate the resulting model on the same test set. Performance rises monotonically with dataset size, with diminishing marginal gains at larger scales.
| Stage I | Stage II | |
| Config | SFT | GRPO |
| Training Hyper-Parameters | ||
| Base model | Qwen3-8B | Qwen3-8B |
| Optimizer | AdamW | AdamW |
| Learning rate | ||
| Training epochs | 1 | 15 |
Table D.1: Training recipes for Spatializer . Stage I : supervised fine-tuning (SFT) on PDNC-Fanfic . Stage II : GRPO post-training with a deterministic dramaturgy reward on PDNC . means are used for the prompt while are used for the answer.
Figure E.1: Stage-Play Generation Prompt. We use the following prompt template to convert a quoted passage into a staged play with canonical speakers, dramaturgically grounded stage positions, and scene splits, outputting structured JSON.
Figure E.2: Minimal Dramaturgy Prompt. A simplified prompt template that performs canonical speaker mapping, coarse stage positioning, and scene splitting with minimal dramaturgical constraints, outputting structured JSON.
Figure E.3: LLM-as-a-Judge Prompt. We use the following prompt to score candidate spatializations (speaker attribution + stage positions + scene metadata) against ground-truth speakers and canonical mappings, producing per-criterion integer scores in JSON.