Emotional Arc Systems

The field is shifting from affect recognition to affect orchestration.

The interesting frontier is not asking a model whether a scene is happy or sad. It is asking whether the scene changes the audience's prediction, alters their allegiance, deepens a character contradiction, and earns the next emotional beat.

A fuller map: emotional arcs sit at the intersection of eight active research fronts.

The frontier is not one paper family. It is an emerging stack that combines reader psychology, controllable generation, narrative planning, interactive agents, reward modeling, diversity metrics, multimodal affect, and mechanistic interpretability.

The useful literature is a braid: psychology, narratology, planning AI, LLMs, games.

Emotion classification is only one small branch. Emotional arc systems need older theories of story experience, symbolic narrative control, modern LLM generation, and product feedback loops.

The clearest way to think about this is as three coupled tracks.

A strong emotional arc model tracks audience state, character state, and plot constraints separately. The affect often comes from their misalignment.

Why “just ask the LLM to write a moving story” underperforms.

Modern models can produce emotionally fluent paragraphs. The failure mode is not bad sentences; it is bad affective causality.

The practical conclusion: emotional arc generation needs explicit state, not because humans write from spreadsheets, but because machines otherwise lose the difference between beautiful prose and earned feeling.

The central design problem is a three-way tension: authorial shape, player agency, and generative freedom.

A good CYOA/roleplay system should not choose between railroaded plot and incoherent sandbox. It needs a runtime bargain: the creator locks the dramatic spine, the player controls values and tactics, and the LLM improvises local texture inside auditable boundaries.

The useful system mental model: do not maximize freedom; preserve meaningful freedom inside a dramatic corridor.

The report now treats agency/coherence as an operating system problem. These visual models show where products fail and what a creator tool should make visible before generation happens.

How an LLM can write emotional arcs in 2026.

The strongest architecture is a hybrid writer's room: narratology supplies control variables, LLMs supply semantic imagination, search supplies alternatives, reward/evaluation supplies pressure, and audience data supplies calibration.

CYOA, roleplay, and TTRPGs need emotional state machines, not just branching prose.

Linear stories optimize one intended sequence. CYOA and roleplay optimize a possibility space: many paths should feel agentic, emotionally legible, mechanically valid, and still converge toward satisfying dramatic pressure.

Given the current scene, propose 3 player choices.
For each choice return:
- surface action
- hidden value tested: loyalty / curiosity / power / mercy / self-preservation / truth
- expected player emotion
- NPC appraisal changes
- world-state changes
- future callback seed
- convergence strategy: unique branch / soft rejoin / hard bottleneck
- risk of railroading or false choice
Do not write prose yet. Design the choice.

You are the narrator, not the rules engine.
Input: player action, dice result, canonical world state, NPC appraisal ledger, target emotional beat.
Output:
1. vivid narration grounded in the dice result
2. one emotional consequence
3. one world-state consequence
4. one NPC belief update
5. 2-3 possible next affordances
Never alter inventory, HP, DC, or facts unless provided by the rules engine.

Prompting should externalize the hidden affect machine.

Test-time prompting is useful when it acts like a control protocol, not when it merely asks for “more emotion.”

Define the emotional contract for this story before writing plot.
Return:
- target audience affect at beginning / midpoint / ending
- primary affect dimensions: valence, arousal, suspense, empathy, curiosity, relief
- forbidden shortcuts: coincidence, sudden confession, unearned rescue, melodrama
- central wound and desire for each major character
- audience knowledge gap: what the audience knows that characters do not
- 8-12 beat arc with affect deltas, not prose

For beat N, produce a scene plan with:
1. Before-state: world, character belief, audience belief
2. Event: what changes externally
3. Appraisal: why each character emotionally changes
4. Audience delta: what emotion should rise/fall
5. Suspense variable: what possible escape route narrows
6. New question created
7. Irreversible consequence
8. Exact affect vector: valence, arousal, tension, intimacy, agency

Evaluate the scene as an emotional arc component, not as prose.
Score 1-7 with evidence:
- target affect delta achieved
- character appraisal is psychologically plausible
- audience knowledge gap is clear
- causal consequence is irreversible
- tension is earned rather than asserted
- novelty: does this avoid common LLM plot echoes?
- aftertaste: what emotional residue remains for the next beat?
Recommend one rewrite operation only.

Rewrite only the beat mechanics, then prose.
Operation: increase suspense by reducing perceived escape routes while preserving character agency.
Constraints:
- no new villain reveal
- no coincidence
- protagonist makes a rational choice that worsens the situation
- audience realizes the danger 2 paragraphs before protagonist
- end with arousal high, valence negative, curiosity unresolved

EmotionPrompt-style emotional stimuli can change LLM behavior, but for story work the bigger win is structural prompting: make the model reason about stakes, appraisal, information asymmetry, and affect deltas. Emotional prompting says “this matters.” Arc prompting says “here is the machine that makes it matter.” EmotionPrompt

Measurement should ask: did the arc change the reader, not did the text contain emotion words?

The best evaluation stack combines theory-based metrics, model-assisted annotation, human panels, and product telemetry.

Product opportunity: emotional arcs become an interface.

The strongest products will not say “AI writes stories.” They will let creators and systems sculpt audience experience with visible affect controls.

An ideal creator interface is an emotional control room for possibility spaces.

For CYOA, roleplay, and TTRPGs, creators need to see branches, state, emotional intensity, NPC appraisals, and convergence pressure at the same time. The best UI is not a chat box; it is a story graph plus an affect debugger plus a DM copilot.

The fascinating intersections are not obvious “emotion AI.”

Where the frontier still feels genuinely open.

The most interesting work is not another emotion classifier. It is a set of missing primitives for affective narrative systems: benchmarks, state representations, controllable decoders, reader simulators, and safety-aware product loops.

Selected research corpus.

Primary and high-signal sources used to build this dossier. Links were title-checked after the earlier incorrect-reference issue; source labels now avoid unsupported paper names.

• Emotional arcsReagan et al. — The emotional arcs of stories are dominated by six basic shapes
• LLM narrative benchmarkTian et al. — Are Large Language Models Capable of Generating Human-Level Narratives?
• Suspense generationXie & Riedl — Creating Suspenseful Stories: Iterative Planning with LLMs
• Narrative theory surveyLiu, Joshi & Dawson — Narrative Theory-Driven LLM Methods
• Theory-informed RLLiu et al. — Retell, Reward, Repeat
• Story rewardXia et al. — StoryAlign / StoryReward
• Plot diversityXu et al. — Echoes in AI: Quantifying Lack of Plot Diversity
• Psychological depthHarel-Canada et al. — Measuring Psychological Depth in Language Models
• Story evaluationChhun et al. — Do Language Models Enjoy Their Own Stories?
• Creative writingIsmayilzada et al. — Evaluating Creative Short Story Generation in Humans and LLMs
• Narrative transportationGreen & Brock — The role of transportation in public narratives
• Narrative planningRiedl & Young — Narrative Planning: Balancing Plot and Character
• Drama managementRoberts & Isbell — Drama management survey
• Foundation affectSchuller et al. — Affective computing has changed
• Emotion promptingLi et al. — Large Language Models Understand and Can Be Enhanced by Emotional Stimuli
• Mechanistic interpretabilityAnthropic — Emotion concepts and their function in a large language model
• GamesYannakakis & Melhart — Affective Game Computing: A Survey
• Arc UITaleBrush — Sketching Stories with Generative Pretrained Language Models
• Possibility-space UIElsewise — Authoring AI-Based Interactive Narrative with Possibility Space Visualization
• Long-form memoryDynamic Hierarchical Outlining with Memory-Enhancement
• Long-form reasoningLearning to Reason for Long-Form Story Generation
• Visual explorationNarrative Studio — Visual Narrative Exploration using LLMs and MCTS
• StoryletsDrama Llama / Dramamancer — LLM-Powered Storylets
• Agent societiesBookWorld — From Novels to Interactive Agent Societies
• Interactive narrativeRiedl & Bulitko — Interactive Narrative: An Intelligent Systems Approach
• Meaningful choiceForeseeing Meaningful Choices
• Experience managementAdversarial Strong Story Experience Management
• State visualizationState Space Visualization for Strong Story Experience Management Design
• LLM branchingGENEVA — Generating and Visualizing Branching Narratives Using LLMs
• Converging narrativesGenerating Converging Narratives for Games with Large Language Models
• Emergent game narrativePlayer-Driven Emergence in LLM-Driven Game Narrative
• Dramamancer designDesign Techniques for LLM-Powered Interactive Storytelling
• AI game masterStatic Vs. Agentic Game Master AI for Solo Role-Playing
• Dungeon master LLMExploring the Potential of ChatGPT as a Dungeon Master in Dungeons & Dragons
• Roleplay agentsRoleLLM / RoleBench — Benchmarking Role-Playing Abilities
• Roleplay evaluationCharacterEval — Role-Playing Conversational Agent Evaluation
• Emotional memoryEmotional RAG — Enhancing Role-Playing Agents through Emotional Retrieval
• TTRPG toolsComputational Tools for Table-Top Role-Playing Games: A Scoping Review
• AI + TTRPG cultureHow Do You Want to View This? Generative AI, Creative Ethos, and TTRPG Hobbyists
• TTRPG therapyScoping Review of TTRPG as Psychological Intervention
• D&D wellbeingCan Playing Dungeons and Dragons Be Good for You?
• D&D self-conceptEfficacy of Dungeons & Dragons for Improving Mental Health and Self-Concepts
• TTRPG affectBardo — Emotion-Based Music Recommendation for Tabletop Role-Playing Games
• Story generation surveyA Survey on LLMs for Story Generation — Findings EMNLP 2025