Narrative Learning in AI & Education

• Narrative learning is a multifaceted approach that uses storytelling to encode knowledge, structure reasoning, and communicate sociocultural norms.
• It integrates techniques like contrastive representation, narrative-guided reinforcement, and executable narratives to improve model interpretability and engagement.
• Its practical applications extend to AI explainability, personalized education, and systemic feedback, offering measurable benefits such as faster learning and improved interactive outcomes.

Taken together, recent research uses narrative learning as a heterogeneous term for several related enterprises: learning from stories and narrative artifacts, learning about narrative structure and similarity, using narrative to organize reflection and discourse, and using story-like prompts or roles to shape reasoning and action. Riedl defines narrative intelligence as “the ability to craft, tell, understand, and respond affectively to stories,” while more recent work instantiates narrative learning as classroom narrative, narrative-driven feedback, contrastive narrative representation learning, narrative-guided reinforcement learning, and natural-language-defined predictive models (Riedl, 2016, Watson et al., 2 Jul 2025, Shen et al., 19 Feb 2026, Tuladhar et al., 10 Sep 2025, Baker, 10 Oct 2025).

1. Conceptual scope and definitions

Riedl’s account of computational narrative intelligence gives the broadest formulation in this corpus: narrative intelligence comprises crafting stories, telling stories, understanding stories, and responding affectively to stories, and narrative learning therefore includes both learning narrative structure and learning sociocultural norms from stories through “machine enculturation” (Riedl, 2016). In that framing, stories are not marginal entertainment artifacts but a primary medium through which humans “communicate, entertain, and teach each other,” and therefore a plausible substrate for AI systems to learn about commonsense, values, and explanation.

Educational work defines the term differently. The StorySpace project treats classroom narrative broadly as “any constructed artifact or performance through which students represent and communicate their understanding of a topic,” explicitly including essays, dioramas, and posters. Its central claim is that narrative engages “reflective and interpretive faculties,” requires students to “internalize the classroom topic,” and often requires them to “adopt a certain point of view” (Watson et al., 2 Jul 2025). In this line of work, narrative learning is not primarily literary analysis; it is a constructivist practice of reflection, interpretation, expression, presentation, and discourse.

A third definition appears in narrative-guided reinforcement learning. In "Narrative-Guided Reinforcement Learning: A Platform for Studying LLM Influence on Decision Making" (Tuladhar et al., 10 Sep 2025), narrative is not a formal logical object but prompted story-context: role framing, story framing, and textual instructions that tell a LLM how to interpret an RL policy’s suggested action and local state. The final action can reinforce or override the RL recommendation, so the narrative functions as a symbolic, story-like overlay on a reward-optimized subsystem. A common misconception is that such systems directly alter the RL update rule; in this implementation, narrative does not change learning rate, discount factor, or Q-update equations, but instead changes online action selection and therefore the distribution of experienced trajectories (Tuladhar et al., 10 Sep 2025).

A fourth usage appears in explainable machine learning. "It's 2025 -- Narrative Learning is the new baseline to beat for explainable machine learning" (Baker, 10 Oct 2025) defines Narrative Learning as an end-to-end supervised method in which the model is “a piece of natural language,” training consists of iterative prompt refinement rather than numerical optimization, an overseer LLM rewrites the narrative, and an underling LLM executes it on datapoints. Here the narrative is simultaneously executable classifier and human-readable explanation.

2. Narrative as a medium for reflection, discourse, and guided learning

In educational technology, narrative learning is frequently realized as a medium for reflection and public meaning-making rather than as a property of a model. StorySpace was explicitly “designed to support and enhance classroom narrative” through three design goals: triggering student reflection and interpretation, accommodating individual student expression, and encouraging discourse. Its medium is a hybrid tangible–digital tabletop system in which students load images, audio, and video onto a projected 2D board, manipulate them with RF-tagged tokens, record macros, and build “a nonlinear, self-referential medium that accommodates many perspectives” (Watson et al., 2 Jul 2025). The pedagogical claim is that multimedia composition, spatial arrangement, and public presentation force students to confront topic complexity rather than reduce it to a linear summary.

A different educational instantiation appears in role-playing game design for visualization literacy. "Designing Narrative-Focused Role-Playing Games for Visualization Literacy in Young Children" (Huynh et al., 2020) compares two otherwise matched versions of a game: one with narrative elements and one without. The narrative version embeds the same graph-reading tasks inside a fantasy RPG structure with dialogue, characters, exploration, and a dramatic arc. In a between-subjects study with 33 participants aged 11–13, narrative elements required roughly double total play time, but the inclusion of narrative elements improved engagement and overall enjoyment without sacrificing learning; the paper reports no significant differences in graph-reading skill development overall, but significant differences in engagement and enjoyment, with Mann–Whitney results including Art $U = 53, p = 0.002$, Activities $U = 66.5, p = 0.01$, Engagement $U = 63, p = 0.007$, and Fun $U = 57.5, p = 0.005$ (Huynh et al., 2020). This suggests a design stance in which narrative primarily reorganizes motivation, persistence, and situational meaning around fixed curricular content.

"Stories and Systems: Educational Interactive Storytelling to Teach Media Literacy and Systemic Thinking" (Roth et al., 14 Aug 2025) generalizes the educational use of narrative to Systemic Learning IDNs, defined as Interactive Digital Narrative experiences explicitly designed to help learners explore, understand, and reflect on complex systems and interdependencies. Its CLASS framework—Creative Curiosity, Lens & Scope, Agency, Scaffolds, and Sandboxes—integrates systems thinking, design thinking, and storytelling. The two case studies, Suzerain and Bezmarisk, treat interactive stories as structured environments for confronting trade-offs, delays, feedback loops, and unintended consequences. Narrative learning in this formulation is experiential and reflective: learners act in a storyworld, observe systemic consequences, and then externalize that experience through systems mapping and debrief.

Narrative can also organize individualized feedback. StoryLensEdu is a narrative-driven multi-agent system for personalized learning reports in self-regulated learning. It combines a Data Analyst agent, a Teacher agent, and a Storyteller agent, with the last one organizing progress, struggle, and next steps through the Hero’s Journey framework. The system models per-objective trajectories such as

$$L_j(s) = \big\{(N_j(T_k), \bar{d}_j(T_k), \bar{u}_j(T_k))\big\}_{k=1}^K,$$

then narrativizes them into stages such as “Ordinary World,” “Tests, Allies, Enemies,” and “Return with the Elixir.” In a formative study with real users, students and teachers rated clarity of data insights at $M=4.5$, intuitiveness of charts at $M=4.6$ and $M=5.0$ respectively, and interactive exploration at $M=4.6$ and $M=4.8$, while narrative engagement was lower at $U = 66.5, p = 0.01$0 and $U = 66.5, p = 0.01$1 (Shen et al., 19 Feb 2026). The result is not simply “data storytelling” but a narrative interface to self-regulated learning, explainability, and action planning.

Embodied narrative guidance extends this logic to remote exploration. NarraGuide integrates a telepresence robot, SLAM-based localization, an exhibit map, and GPT-4-driven dialogue to support remote museum tours. Its narrative is conditioned on current area, nearby exhibits, curated introductions, and sample guide–visitor dialogues, and its user study with 20 remote participants found significantly more acceptance than rejection of robot suggestions, with accepted suggestions $U = 66.5, p = 0.01$2, rejected suggestions $U = 66.5, p = 0.01$3, and $U = 66.5, p = 0.01$4 (Hu et al., 2 Aug 2025). Participants described the robot as guide, instructor, learning tool, companion, and proxy, and several reported that exhibit questions and prompts “lured curiosity” or made the experience “deeper.” This suggests that narrative learning can be spatially grounded and mixed-initiative rather than text-bound.

3. Narrative as environment, prompt, and control signal in AI systems

A central AI usage treats narrative learning as situated interaction inside a story world. "Situated Language Learning via Interactive Narratives" (Ammanabrolu et al., 2021) argues that interactive narratives should be the environments of choice for training agents that must understand and generate contextually relevant natural language in service of goals. Text games are modeled as POMDP-like environments in which an agent perceives via textual observations, acts via textual commands, and learns under challenges of knowledge representation, commonsense reasoning, and exploration. The paper’s key claim is that static corpora do not enforce grounding or goal-directed use of language, whereas interactive narrative environments do.

Narrative can also operate as a supervisory layer over classical RL. Narrative-guided RL couples a tabular Q-learning policy with a LLM that receives the RL suggestion, local grid observations, and a narrative framework such as Direct instructions, Theseus, Sherlock Holmes, or Westworld-inspired AI agent (Tuladhar et al., 10 Sep 2025). In $U = 66.5, p = 0.01$5 grids with 30–40% obstacle density, RL+LLM agents with just direct instructions after 10 episodes achieved performance similar to RL-only agents after 100 episodes; Theseus and Sherlock retained comparable success rate to the direct-instruction baseline but with lower average steps; and the Westworld AI narrative produced the best overall performance, with higher success rates and lower average steps. The architecture is explicitly cast as dual-process: a fast, reward-based subsystem and a slower, symbolic, story-driven subsystem.

Prompting work shows that narrative can improve reasoning even without environment interaction. "Can Stories Help LLMs Reason? Curating Information Space Through Narrative" (Javadi et al., 2024) introduces Story of Thought (SoT), a three-step prompting pipeline: Question Clarification, Narrative Generation, and Problem Solving. The narrative stage explicitly requires Progressive Disclosure, Branching, Analogy, Analogical Reasoning, and Metaphor. On GPQA Diamond, Llama 3 70B + SoT achieved 51.01%, and on JEEBench the same model achieved a total score of 0.453, surpassing the reported benchmark of GPT-4 + CoT + Self-Consistency at 0.389 (Javadi et al., 2024). Smaller models sometimes degraded under SoT, which the paper attributes to poor-quality narratives or overlong context, but the general pattern is that narrative-based information curation improves complex STEM reasoning.

Narrative theory has also been used as a reward source for generation. "Retell, Reward, Repeat: Reinforcement Learning for Narrative Theory-Informed Story Generation" (Liu et al., 23 Jan 2026) operationalizes Todorov’s Theory of Narrative Equilibrium through Logical, Rational, Complete$U = 66.5, p = 0.01$6, and Narrativity, with an overall criterion

$U = 66.5, p = 0.01$7

These scores are supplied by LLM judges and used as rewards in d-RLAIF with GRPO on the TimeTravel counterfactual retelling task. The paper reports that d-RLAIF is a viable alternative to supervised fine-tuning, producing stories that are more diverse and aligned with human narrative conventions, and that $U = 66.5, p = 0.01$8-rewarded models attain especially strong Narrativity while remaining logical and rational (Liu et al., 23 Jan 2026). This is narrative learning in a stricter sense than prompt engineering: the policy is optimized against an explicit narratological rubric.

4. Narrative representation learning: structure, salience, similarity, and order

A major line of work treats narrative learning as the problem of learning representations that preserve plot, order, and story-level structure. "On Narrative Information and the Distillation of Stories" (Ashley et al., 2022) formalizes narrative information as the overlap in information space between a story and the items that compose it, and defines narrative essence as the atom-wise feature maximizing mutual information with story order:

$U = 66.5, p = 0.01$9

Using contrastive learning on music albums, the paper learns low-dimensional narrative essence embeddings, then derives narrative templates via an evolutionary algorithm and uses a curve-fitting algorithm to reorder albums. It reports about 1.924 bits of mutual information with album order for 1D essence, significant gains over random and shuffled-essence baselines, and explicit evidence that narrative templates are present in existing albums (Ashley et al., 2022). Here narrative learning is neither literary nor pedagogical; it is an information-theoretic account of ordered media collections.

"Three Stage Narrative Analysis; Plot-Sentiment Breakdown, Structure Learning and Concept Detection" (Khan et al., 14 Nov 2025) offers a different representation stack for movie scripts. Scripts are segmented into fixed-size windows, scored with a custom NRC-VAD lexicon integrated into LabMTsimple, smoothed with a 10-segment context window, and represented as arousal-based sentiment trajectories. These trajectories are then clustered with Ward’s hierarchical clustering, yielding three primary clusters and finer partitions up to 100 clusters, which the authors interpret in terms of recurring emotional structures such as “Rags to Riches” and “Icarus” (Khan et al., 14 Nov 2025). Structure learning is thus unsupervised discovery of recurrent arc types; concept detection is proposed but not yet implemented.

NarCo moves from trajectories to explicit coherence graphs. "Fine-Grained Modeling of Narrative Context: A Coherence Perspective via Retrospective Questions" (Xu et al., 2024) represents a narrative as nodes corresponding to short snippets and edges corresponding to free-form retrospective questions from later snippets to earlier snippets. Questions are produced through a two-stage prompting pipeline—connection identification and question conversion—and filtered by back verification. The resulting graph is task-agnostic but useful in multiple tasks: recap snippet identification, plot retrieval, and long-document QA. The paper reports that adding NarCo improves recap identification on Notre-Dame de Paris and Game of Thrones, improves plot retrieval nDCG, and improves QuALITY QA when used in retrieval (Xu et al., 2024). This is an explicitly coherence-centric notion of narrative learning.

Another line uses contrastive twins to learn plot-aware embeddings. "Contrastive Learning with Narrative Twins for Modeling Story Salience" (Sterner et al., 12 Jan 2026) trains encoders on narrative twins—stories that share the same plot but differ in surface form—and evaluates four narratologically motivated salience operations: deletion, shifting, disruption, and summarization. On ROCStories and longer Wikipedia plot summaries, contrastively learned embeddings outperform a masked-language-model baseline, and summarization is the most reliable operation for identifying salient sentences. If twins are unavailable, the paper shows that random dropout can generate surrogate twins, and for long-form narratives effective distractors can come from different parts of the same story (Sterner et al., 12 Jan 2026). This suggests that plot-preserving invariance is a strong self-supervised signal for salience.

Narrative representation learning also now has shared-task infrastructure. "SemEval-2026 Task 4: Narrative Story Similarity and Narrative Representation Learning" (Hatzel et al., 23 Apr 2026) operationalizes narrative similarity as choosing which of two stories is more similar to an anchor story, where similarity is defined as “the perception of story relatedness, focusing on abstract patterns of causality and progression rather than concrete details.” It distinguishes Course of Action, Outcomes, and Abstract Theme, collects more than 1,000 story-summary triples, receives 71 final submissions from 46 teams, and finds that top Track A systems are often LLM ensembles while Track B systems with pre- and post-processing on pretrained embeddings perform about on par with custom fine-tuning (Hatzel et al., 23 Apr 2026). The reported genre correlation further indicates that current embeddings partly internalize genre-level plot regularities.

Multimodal order learning extends these representational ideas to film. "MoviePuzzle: Visual Narrative Reasoning through Multimodal Order Learning" (Wang et al., 2023) defines a movie clip as shuffled but aligned frame–utterance pairs and casts narrative understanding as a permutation-recovery problem over a hierarchy of frames, shots, scenes, and clips. Its HCMC model combines pairwise ordering classification, contrastive clustering, and top-down clustering with bottom-up reordering. On the MoviePuzzle benchmark derived from 228 movies and 10,031 clips, the full model reaches 55.40 pairwise ordering score in-domain, outperforming ablations without text, vision, contrastive learning, or the shot hierarchy (Wang et al., 2023). This treats narrative learning as multimodal structure induction rather than text-only plot modeling.

5. Narrative as explanation, executable model, and alignment interface

One of the most distinctive recent developments is the treatment of narrative itself as a predictive model. In "It's 2025 -- Narrative Learning is the new baseline to beat for explainable machine learning" (Baker, 10 Oct 2025), an overseer LLM iteratively refines a natural-language rule narrative and an underling LLM executes that narrative to classify inputs. Performance is tracked with negative log Krichevsky–Trofimov accuracy,

$U = 63, p = 0.007$0

and the paper reports that Narrative Learning became more accurate than the baseline explainable models on 5 out of 6 datasets by 2025 or earlier because of improvements in LLMs (Baker, 10 Oct 2025). The important conceptual point is that the narrative is not post-hoc explanation layered on top of a separate model; it is the model.

Riedl’s earlier discussion of computational narrative intelligence anticipates this use of narrative as an explanatory interface. He argues that any process or procedure can be told as a narrative and that AI systems could describe how they reached a conclusion or why they performed an action by couching the explanation in narrative terms, because “the human mind is tuned for narrative understanding” (Riedl, 2016). In the same paper, narrative is also proposed as a route to machine enculturation, where stories supply demonstrations of socioculturally appropriate behavior under hypothetical situations. This makes narrative relevant to both explainability and alignment.

A related but more user-facing form appears in educational feedback systems. StoryLensEdu was motivated by the observation that dashboards and text reports often suffer from poor interpretability, monotonous presentation, and limited explainability, and it uses a narrative re-authoring of the learner’s activity traces to answer Hattie and Timperley’s questions “Where am I going?”, “How am I going?”, and “Where to next?” (Shen et al., 19 Feb 2026). The addition of post-generation interactive question answering is especially significant: explanation is not merely written once, but can be interrogated in dialogue. A plausible implication is that narrative explanation is most effective when paired with access to underlying evidence and drill-down interaction rather than treated as a decorative summary.

6. Limitations, tensions, and open directions

Taken together, this literature does not converge on a single formalism of narrative learning. In some papers, narrative is an educational practice or artifact; in others it is a control signal, an embedding objective, a graph of retrospective questions, a reward rubric, or an executable natural-language classifier. This suggests breadth rather than terminological stability. A common misconception is that all of these works study the same phenomenon; the corpus instead shows overlapping but non-identical research programs (Riedl, 2016, Watson et al., 2 Jul 2025, Tuladhar et al., 10 Sep 2025, Baker, 10 Oct 2025).

Many systems remain preliminary or domain-bounded. Narrative-guided RL is explicitly a preliminary platform with only simple gridworlds, short training, a small number of narrative frameworks, no formal narrative reward, and no rigorous statistical tests or generalization studies (Tuladhar et al., 10 Sep 2025). StorySpace remains a “fragile resource,” and media loading is described as tedious (Watson et al., 2 Jul 2025). NarraGuide depends on semantic maps and LLM prompting but still suffers from hallucinations, interruptibility problems, and social complications created by bystanders (Hu et al., 2 Aug 2025). StoryLensEdu improves interpretability and engagement, but users also reported information density, tone mismatch, and occasional over-encouragement relative to actual performance (Shen et al., 19 Feb 2026).

Representation-learning approaches face their own unresolved issues. SoT improves reasoning on hard STEM problems, but narrative quality is difficult to measure directly, and smaller models can degrade under narrative prompting (Javadi et al., 2024). NarCo relies on pairwise snippet relations, expensive LLM-based graph construction, and imperfect question generation and filtering (Xu et al., 2024). Narrative-twin salience modeling depends on embedding methodology and on the availability of twins or effective distractors, though dropout twins and in-story negatives partially mitigate this (Sterner et al., 12 Jan 2026). The SemEval NSNRL task further shows that even with curated labels there is substantial subjectivity, with Krippendorff’s $U = 63, p = 0.007$1 and considerable headroom between best systems and the estimated oracle (Hatzel et al., 23 Apr 2026).

Future directions are correspondingly diverse. Narrative-guided RL proposes richer reasoning schemas, broader RL algorithms and environments, closer integration of narrative into the reward function, meta-learning over narratives, and human-in-the-loop narrative guidance (Tuladhar et al., 10 Sep 2025). Systemic Learning IDNs point toward AI-driven agents, XR, and broader evaluation of curiosity, systems thinking, and ethical reasoning (Roth et al., 14 Aug 2025). Narrative-information and narrative-salience work motivate cross-domain validation, human evaluation, and hierarchical narrative representations beyond sentence- or window-level embeddings (Ashley et al., 2022, Sterner et al., 12 Jan 2026). The overall trajectory suggests that narrative learning is becoming less a niche topic about stories alone and more a general family of methods for structuring cognition, interaction, explanation, and control through story-shaped representations.