ReaderScope: From Offline Reader-Response Calibration to Online Controlled Story Generation¶
Conference: EMNLP 2026
OpenReview: https://aclanthology.org/2026.emnlp-main.1074/
Paper: ACL Anthology
Code: https://github.com/FengZhang758/ReaderScope
Area: LLM Reasoning
Keywords: Large Language Models, Controlled Story Generation, Reader-Response Theory, Trajectory Control, Closed-Loop Revision
TL;DR¶
Addressing the limitation that long-form story generators optimize plot coherence while neglecting the reader's evolving reading experience, ReaderScope calibrates four-dimensional reader-response trajectories (emotion, entropy, immersion, insight) offline across 20 reference novels and steers online generation through closed-loop reader-feedback revision.
Background & Motivation¶
A long-form narrative can follow a structurally sound outline, maintain strict character consistency, and adhere to causal logic, yet still fail as a satisfying literary experience: it frequently leaves human readers emotionally indifferent, unsurprised by plot turns, weakly immersed in the story world, and devoid of reflective resonance. Existing large language model (LLM) story generators predominantly operate under author-facing or text-facing paradigms—such as hierarchical outline expansion, external memory retrieval, or critic-based textual revision. These systems focus almost entirely on controlling narrative events, state persistence, and local syntactic fluency, while evaluating reader-oriented psychological impacts only as an afterthought once the entire text has already been drafted.
The core tension underlying this disconnect is that literary reading is fundamentally an unfolding, non-stationary cognitive dynamic: throughout a narrative, readers experience shifting emotional valences, calibrate anticipatory suspense around unresolved narrative branches, undergo immersive perceptual transportation, and synthesize moral or philosophical insights from character choices. Turning these reader-response dynamics into generation-time controls has historically hit a dual barrier: collecting dense human reader feedback during iterative drafting is prohibitively expensive, whereas distilling the reading experience into a single scalar quality score discards the temporal trajectory and multidimensional structure of narrative engagement. Without explicit, groundable trajectory targets, generation pipelines have no principled way to diagnose when suspense has stalled, when emotional progression has flatlined, or when descriptive pacing has broken immersion.
To bridge this control gap, this paper adapts literary Reader-Response Theory into an actionable computational pipeline, promoting simulated reader reactions from post-hoc evaluation proxies into generation-time steering targets. Core idea: calibrate multidimensional reader-response trajectories offline across 20 canonical novels across emotion, entropy, immersion, and insight, derive hybrid window constraints and elastic statistical envelopes, and guide online long-form generation via a closed-loop reader-panel feedback controller.
Method¶
ReaderScope formulates controllable narrative generation as a two-stage paradigm: offline trajectory calibration from reference literature and online reader-in-the-loop controlled composition. Rather than solely monitoring plot facts, the framework directly governs how the simulated audience processes and internalizes the evolving narrative.
Overall Architecture¶
In the offline calibration stage, a pool of 25 LLM reader agents—equipped with structured working, episodic, and semantic memories—reads 20 length-normalized classic novels (60 narrative units each), recording four-dimensional response signals. Trajectory clustering diagnostics reveal that emotion and immersion require local sliding-window constraints, whereas suspense entropy and insight exhibit consistent temporal structure warranting elastically aligned amplitude and slope envelopes. In the online generation stage, a Story Composer expands requirements into a five-act outline and drafts content in half-act units (6 paragraphs each). A 5-reader panel evaluates each unit on the fly, and a Feedback Controller assesses deviations against calibrated envelopes to generate targeted natural-language revision guidance within a fixed rewrite budget.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Story Requirements<br/>Premise / Setting / Cast"] --> B["Four-Dimensional Reader Response Modeling<br/>Retrieve-Respond-Update Cognitive Loop"]
B --> C["Dual-Track Calibration of Control Targets<br/>Sliding Windows + Elastic Statistical Envelopes"]
A --> D["Hierarchical Composition & Planning<br/>Five-Act Outline & Half-Act Windows"]
C --> E["Online Reader-in-the-Loop Revision<br/>Reader Panel Scoring + Violation Diagnosis & Rewrite"]
D --> E
E --> F["Controlled Coherent Long-Form Story"]
Key Designs¶
1. Four-Dimensional Reader Response Modeling: Deconstructing Literary Experience into Operational Trajectories Prior storytelling metrics often compress literary appreciation into uniform scalar ratings, ignoring the temporal multidimensionality of reader engagement. To address this, the authors mathematically operationalize four complementary response signals for simulated reader agent \(r\) at narrative unit \(t\). The Emotion signal \(A_{r,t}\) captures local affective shifts by identifying \(K_{r,t}\) emotional turning points with valence \(e_{r,t,k} \in [-1, 1]\) and averaging them as \(A_{r,t} = \frac{1}{K_{r,t}}\sum_{k=1}^{K_{r,t}} e_{r,t,k}\). The Entropy signal \(E_{r,t}\) quantifies suspense-related uncertainty by combining normalized Shannon entropy \(H_{r,t}^{\text{new}}\) over candidate answers of newly raised story questions with a resolution-impact score \(S_{r,t}^{\text{res}}\) for resolved historical questions, formulated as \(E_{r,t} = H_{r,t}^{\text{new}} + S_{r,t}^{\text{res}}\) where resolution rewards unexpectedness \(U_{r,t}\) while penalizing cognitive opacity \(D_{r,t}\). Immersion \(I_{r,t}\) adapts the 18-item Story World Absorption Scale (SWAS) across attention, transportation, emotional engagement, and mental imagery into a normalized composite in \([0, 1]\). Insight \(G_{r,t}\) balances 14 reflective meaning items \(Q_{r,t}\) with the absolute value-profile shift \(\Delta V_{r,t}\) across 10 Schwartz basic human values, yielding \(G_{r,t} = \frac{1}{2}(Q_{r,t} + \Delta V_{r,t})\). Operating within a retrieve-respond-update cognitive cycle, the reader models ensure that responses remain grounded in both individual reading personas and accumulated context.
2. Dual-Track Calibration of Control Targets: Accommodating Narrative Heterogeneity via Hybrid Envelopes Imposing a single global prototype or rigid curve across diverse literary works suppresses narrative originality. By analyzing 500 trajectories collected from 20 reference novels with Soft-DTW k-medoids clustering, the authors uncovered a sharp dichotomy: emotion and immersion maintain high cluster purity (\(>90\%\)) aligned with specific story identities (\(K \approx 20\)), whereas entropy and insight lose clustering structure after normalization and collapse to trivial binary splits. ReaderScope resolves this discrepancy through a dual-track strategy. For emotion and immersion, it avoids fixed paths and extracts local window tolerances: the maximum tolerated span of unbroken emotional monotonicity is calibrated to a ratio of 0.076 (approx. 4.5 narrative units), while the maximum tolerated window of sustained immersion decline is bounded at 0.082. For entropy and insight, the framework applies Fisher-Rao elastic registration to phase-align trajectories, extracting continuous mean curves alongside positionwise amplitude envelopes (value bounds) and slope envelopes (bounding adjacent differences). This dual calibration simultaneously accommodates story-specific emotional arcs while enforcing canonical global envelopes for suspense and reflective payoff.
3. Online Reader-in-the-Loop Revision: Closed-Loop Hierarchical Steering via Targeted Feedback Regenerating entire multi-thousand-word narratives upon failure incurs prohibitive inference costs and risks narrative drift, while unstructured textual critiques often fail to provide actionable direction. ReaderScope introduces a paragraph-level hierarchical controller spanning 60 units (5 acts, 10 half-acts, 6 paragraphs per half-act). As the Story Composer completes each 6-paragraph unit under the active half-act goal, a panel of 5 simulated readers reads the text and updates memory states. The Feedback Controller scans for multi-dimensional violations: unbroken emotional monotonicity, persistent immersion drops, or envelope boundary violations in entropy and insight. When violations occur, the controller pinpoints the exact paragraph range and translates numeric deviations into qualitative rewriting suggestions (e.g., dampening excessive uncertainty to clarify causality, or introducing character hesitation to restore suspense). The Composer rewrites only that 6-paragraph unit within a budget of \(B=1\), ensuring fast convergence and structural stability before proceeding down the narrative tree.
Loss & Training¶
ReaderScope is an inference-time agentic framework requiring no parameter fine-tuning. All generators, simulated readers, and evaluators operate via prompting on LLM backbones (standardized on gpt-4o-mini for fair comparison). Online rewriting adopts a budget \(B=1\) per half-act to maintain computational tractability.
Key Experimental Results¶
Main Results¶
Evaluations are conducted on 20 story requirements from the StoryBox benchmark against four competitive baselines: vanilla sequential prompting (Rolling), outline-guided planning (DOC), multi-agent sandbox simulation (StoryBox), and critique-based iterative refinement (Critics). Round-robin pairwise comparisons are gathered under automated evaluation (DeepSeek-V4-Flash) and human evaluation (10 annotators, 200 pairwise decisions, evaluated across five dimensions). Statistical significance uses cluster bootstrapping over story requirements with Holm multiplicity correction.
The table below presents human preference scores (\(\text{Pref} \in [0, 1]\), where \(0.5\) denotes parity) alongside paired baseline differences:
| Evaluation Dimension | ReaderScope (Ours) | 95% Clustered CI | Diff vs Rolling | Diff vs DOC | Diff vs Critics | Diff vs StoryBox |
|---|---|---|---|---|---|---|
| Narrative Craft | 0.681 | [0.550, 0.800] | +0.094 | +0.394* | +0.344* | +0.075 |
| Emotion | 0.725 | [0.606, 0.838] | +0.281* | +0.381* | +0.331* | +0.131 |
| Suspense | 0.719 | [0.581, 0.844] | +0.206 | +0.394* | +0.369* | +0.125 |
| Immersion | 0.663 | [0.519, 0.800] | +0.188 | +0.281 | +0.331 | +0.013 |
| Insight | 0.706 | [0.606, 0.800] | +0.188 | +0.406* | +0.300* | +0.138 |
*Note: Differences marked with * remain statistically significant after panel-wide Holm correction (\(p_{\text{Holm}} < 0.05\)). Automated and human evaluations show strong system-level alignment (Pearson \(r=0.877\), Spearman \(\rho=0.846\)).
Ablation Study¶
To verify the contribution of each reader-response dimension and determine trajectory enforcement fidelity, the authors conduct component-wise ablations alongside control fidelity metrics:
| Dimension / Configuration | Criterion Type | Window Success (%) | Story Success (%) | In-Band Points (%) | Norm. Band Dev. | Score Gain vs Ablation |
|---|---|---|---|---|---|---|
| Full Model | - | - | - | - | - | Baseline |
| Emotion Control | Monotonic Reversal Window | 100.0 | 100.0 | - | - | +0.64 (Largest Drop) |
| Entropy Control | Statistical Amplitude & Slope | 97.5 | 75.0 | 86.5 | 0.0061 | +0.40 |
| Immersion Control | Anti-Decline Window | 100.0 | 100.0 | - | - | +0.28 |
| Insight Control | Statistical Amplitude & Slope | 100.0 | 100.0 | 99.9 | 0.0000 | +0.08 |
Key Findings¶
- Affective regulation is the primary engine of story quality: Ablating emotion feedback leads to the sharpest performance deterioration across all metrics (Gain = +0.64), proving that modulating valence shifts and emotional pacing is central to sustaining narrative drive.
- Suspense entropy is the most challenging target to enforce: While entropy achieves a 97.5% window-level success rate, its full-story success rate drops to 75.0%, indicating that narrative suspense is acutely sensitive to timing, where a single misplaced resolution disrupts whole-story tension.
- Competitiveness with sandbox generation: ReaderScope consistently achieves higher mean preference scores than StoryBox across all metrics (\(\Delta\) between +0.013 and +0.138), but clustered confidence intervals span zero, establishing ReaderScope within the leading preference tier without proving statistical superiority over StoryBox.
Highlights & Insights¶
- From Post-Hoc Evaluation to In-Loop Generation Control: ReaderScope moves literary reader metrics into generation-time steering, converting abstract aesthetic ideals into bounded trajectory targets.
- Dual-Track Calibration Respects Literary Diversity: Recognizing that emotional arcs vary widely by plot while suspense and insight adhere to broad structural bounds prevents formulaic, over-standardized narrative outputs.
- Targeting Bounded Envelopes over Monotonic Maximization: The framework demonstrates that literary pacing thrives on bounded moderation—excessive suspense creates causal chaos, while insufficient uncertainty induces reader boredom.
Limitations & Future Work¶
- Reliance on LLM Reader Proxies: Simulated readers may exhibit inherent model biases and lack the cultural, psychological, and demographic diversity of real human reading cohorts.
- Limited Reference Corpus Diversity: Trajectory envelopes are calibrated on 20 classical English novels, potentially underrepresenting non-Western pacing traditions, modern genre fiction, and non-traditional narrative formats.
- Inference Overhead from Multi-Agent Feedback: Maintaining memory states across 5 reader agents and executing online half-act revisions significantly multiplies API call volume and generation latency compared to single-pass methods.
Related Work & Insights¶
- vs DOC / Re3 (Plan-and-Write Systems): While outline-guided models preserve macro-level plot consistency, their narratives frequently suffer from monotonous pacing; ReaderScope introduces temporal affective dynamics on top of hierarchical outlines.
- vs Critics (Text-Criticism Frameworks): Unstructured natural-language critique models often generate vague, qualitative suggestions; ReaderScope grounds revision in concrete mathematical trajectory deviations.
- vs StoryBox (Emergent Multi-Agent Sandbox): StoryBox generates rich emergent interactions from character agents but lacks global trajectory steering; ReaderScope provides external reader-experience governance that could naturally complement sandbox generation.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ [Pioneering formalization of reader-response theory into actionable, multidimensional trajectory controls for story generation]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive human and automated pairwise evaluations, clustered bootstrap significance testing, trajectory fidelity metrics, and sensitivity audits]
- Writing Quality: ⭐⭐⭐⭐⭐ [Rigorous conceptual foundations, elegant mathematical definitions, and lucid case-study analysis]
- Value: ⭐⭐⭐⭐⭐ [Provides a crucial paradigm shift from author-centric factual consistency to audience-centric experiential control in creative NLP]