Technical Summary: LOOM – Controllable Narrative Rendering for Enhanced Assisted Writing

Problem Statement

Despite the proficiency of Large Language Models (LLMs) in basic writing assistance (grammar correction, fluency enhancement), their application to creative writing is hindered by a "binary failure." Current systems oscillate between two extremes:

1. Remedial Polishing: Safe, surface-level editing that fails to enhance the reader's experience or transform the perceptual texture of the text.
2. Destructive Expansion: Uncontrolled plot expansion where models, when prompted for expressivity, hallucinate new events or alter the story's causal structure.

This dilemma represents a critical trade-off between narrative fidelity (preserving the original story events) and descriptive intensity (enhancing atmosphere, mood, and sensory detail). Existing frameworks lack the granularity to decouple the story (what happens) from the discourse (how it is told), often conflating stylistic enrichment with plot alteration.

Methodology: The LOOM Framework

The authors propose LOOM, an assisted writing framework grounded in narratological theory (specifically the distinction between story and discourse). LOOM operationalizes a three-layer pipeline using a role-play-based semiotic chain-of-thought to enforce precise control over narrative intent and rendering density.

1. Perception Quota Layer (Bounded Allocation)

This layer acts as an "actuarial budgeter." Instead of generating content immediately, it determines where and how much detail to inject.

• Input: Raw text $T$, narrative intent $I$, and a global sensory budget $\Omega_{total}$.
• Mechanism: The system extracts event-carrying verbs $V$ and assigns a quota vector $q_i$ for each verb across seven sensory dimensions (Visual, Auditory, Olfactory, Gustatory, Tactile, Interoceptive, Kinesthetic).
• Constraints: The allocation must satisfy a global budget constraint ($\sum \|q_i\|_1 = \Omega_{total}$) and a local cap ($\|q_i\|_1 \le \Omega_{max}$) to prevent any single event from being overwhelmed. Central verbs receive higher allocations; peripheral verbs receive zero.

2. Meaning Making Layer (Semiotic Transformation)

This layer functions as a "narrative semiotician," transforming abstract numerical quotas into concrete, intent-aligned semantic symbols.

• Mechanism: For every non-zero quota, the model performs a semantic mapping $\Phi$ driven by a semiotic chain-of-thought. It reasons how a physical sensation embodies the abstract intent.
• Output: It generates semantic atoms (minimal conceptual units, e.g., "stale air" or "distant thump") rather than full sentences. Crucially, it produces a reasoning trace $J_{i,j}$ that serves as the cognitive bridge between the concrete carrier and the abstract intent, ensuring the detail is thematically relevant.

3. Narrative Rendering Layer (Microsurgical Injection)

This layer executes the integration of semantic atoms into the original text.

• Mechanism: Unlike standard generative rewriting which reshapes entire paragraphs, this layer applies a span-constrained injection operator $\Psi$. It identifies the precise narrative gap surrounding a specific verb and weaves the semantic atoms into the local context.
• Constraints: The process adheres to a strict "Show, Don't Tell" instruction and a negative constraint against adding new causal actions. This ensures modifications remain structurally subordinate to the original event backbone.

Key Contributions

1. LOOM Framework: A structured, three-layer pipeline that operationalizes the narratological semiotic chain-of-thought. It enables precise control over both narrative intent and rendering density, separating perceptual material generation from syntactic insertion.
2. Theoretically Informed Evaluation Protocol: The authors design a multi-dimensional rubric (RPB, RMC, RSI) grounded in the story-discourse distinction. This explicitly measures expressive intensity and adherence to source events, addressing the limitations of standard metrics that conflate stylistic enrichment with plot alteration.
3. Comprehensive Evaluation: Through LLM-based metrics, human assessments, and ablation studies, the paper demonstrates that LOOM resolves the tension between factual integrity and descriptive intensity, outperforming state-of-the-art baselines.

Experimental Results

The evaluation utilized the ROCStories dataset (5-sentence narratives with minimal descriptive texture) and compared LOOM against standard models (Qwen2.5 series) and reasoning-enhanced models (Qwen3, o4-mini, GPT-5-thinking).

• Rendering Proportion Balance (RPB): Baseline models scored near 1.0 (indicating overwhelming expansion). LOOM achieved 2.83 (near the ideal 3), demonstrating effective control over verbosity.
• Rendering Method Compliance (RMC): LOOM achieved 4.83, significantly outperforming baselines (e.g., GPT-5-thinking at 4.29, Qwen2.5-7B at 2.45). This confirms LOOM's ability to enhance text without hallucinating new plot elements.
• Stylistic Integration (RSI): While GPT-5-thinking scored slightly higher (4.33 vs. 3.93), this was attributed to paragraph rewriting strategies that sacrificed fidelity. LOOM maintained high integration while preserving the original structure.
• Overall Performance: LOOM achieved the highest overall score (3.86) in automated evaluation and 3.77 in human evaluation, significantly outperforming the GPT-5-thinking baseline (2.77).
• Ablation Studies: Removing the Meaning Making layer caused a drop in fidelity (RMC), while removing the Narrative Rendering layer (replacing injection with rewriting) degraded coherence (RSI).
• Control Accuracy: The system demonstrated perfect adherence (1.00 accuracy) to both total quota and per-verb cap constraints, with token expansion scaling linearly with the assigned budget.

Significance and Claims

The paper claims that LOOM successfully resolves the fundamental tension in AI-assisted creative writing: the trade-off between factual integrity and descriptive intensity. By decoupling the generation of perceptual material from syntactic insertion, LOOM ensures that narrative enhancement occurs without violating the original event structure.

The authors assert that this work provides the first computational framework capable of precisely controlling enhancement within the perceptual layer. They conclude that integrating structural constraints (via the three-layer pipeline) is viable for reliable, creative human-AI co-writing, moving beyond "remedial polishing" to active "narrative rendering."

Limitations Acknowledged: The authors note that LOOM's reliance on linear story-discourse structures limits its immediate applicability to poetry or non-linear plots, and future work will investigate agent-based automation and distillation to address these constraints.