Matched Story-Continuation Paradigm

• The matched story-continuation paradigm is a controlled framework that holds a fixed narrative context while comparing alternative continuations.
• It spans multiple instantiations including forced-choice textual tests, dialogue response retrieval, and visual story generation to evaluate context-sensitive coherence.
• Researchers use this paradigm to measure narrative consistency, participant-state tracking, and causal plausibility, thereby isolating post-training impacts and stylistic artifacts.

Matched story-continuation paradigm denotes a family of narrative evaluation and generation setups in which a fixed story prefix, dialogue history, paired story context, or visual prefix is held constant while one or more continuations are compared, selected, or generated relative to that shared context. The term is used explicitly for a controlled checkpoint comparison in “Narrative Flattening” (Li et al., 27 May 2026), where the same human-written story is cut at controlled points and the held-out suffix becomes a matched human continuation, but closely related formulations appear throughout story ending selection, dialogue response retrieval, procedural plausibility judgment, and multimodal story continuation (Srinivasan et al., 2018, Si et al., 2021, Jiang et al., 2023, Maharana et al., 2022). In all of these forms, the central object is not free continuation in isolation; it is the relation between a continuation and a preceding narrative state.

1. Core definition and design space

In its explicit 2026 formulation, the paradigm is a controlled continuation experiment over the same underlying human story. “Narrative Flattening” reveals a prefix at cut point $c \in \{40\%,60\%,80\%,90\%\}$, treats the remaining original text as the matched human continuation, gives the same prefix to multiple checkpoints from the same model lineage, and computes all measurements only on the continuation (Li et al., 27 May 2026). Prefixes are computed over sentence sequences, so each prefix ends at a sentence boundary, the target continuation length is set to the word count of the held-out human suffix, and five model continuations are sampled for each story–cut–model tuple (Li et al., 27 May 2026). This establishes matching at the story-prefix/continuation-task level rather than sentence-to-sentence alignment.

Earlier work instantiates the same logic under different names. In the Story Cloze Test, the match is between a four-sentence prompt and two candidate fifth sentences, one labeled the “right” ending and one the “wrong” ending (Srinivasan et al., 2018). In CRD3 dialogue continuation, the match is between a long narrative dialogue context and a candidate next utterance, evaluated with Hits@1/10 and Hits@5/10 over ten candidates (Si et al., 2021). In TRIP and LEAP, the match is between paired plausible and implausible procedural stories, with the model required not only to choose the plausible alternative but also to localize conflicting sentences and participant-state violations (Jiang et al., 2023). In visual story continuation, the match is between future frames and a source frame or conditioning video, so that generation remains coherent with previously established characters and scenes (Maharana et al., 2022, Bugliarello et al., 2023).

Instantiation	Shared context	Matching target
Story Cloze	Four-sentence prompt	Correct fifth sentence (Srinivasan et al., 2018)
CRD3 dialogue continuation	Summary history, relation labels, recent turns	Gold next response among 10 candidates (Si et al., 2021)
TRIP / LEAP	Paired plausible and implausible stories	Plausible story plus conflict explanation (Jiang et al., 2023)
Narrative Flattening	Same human prefix at $40\%,60\%,80\%,90\%$	Human suffix and model continuations (Li et al., 27 May 2026)
StoryBench continuation	First 0.5s of ground-truth video plus future prompts	Future video segments (Bugliarello et al., 2023)
Source-frame visual continuation	First frame plus storyline	Remaining coherent frames (Maharana et al., 2022, Tao et al., 2024)

This suggests that “matched” has at least three recurrent meanings: shared candidate context, shared human prefix, and source-grounded continuation. A plausible implication is that the paradigm is best treated as a controlled comparison framework rather than a single benchmark format.

2. Canonical textual instantiations

The canonical textual form is forced-choice ending selection. “A Simple and Effective Approach to the Story Cloze Test” studies the task in which a system is presented with a four-sentence story prompt and must determine which of two potential fifth sentences is the “right” ending (Srinivasan et al., 2018). That paper is important less because it strengthens the paradigm than because it diagnoses its fragility: training on the validation set outperforms training on the official training set, a no-context model is already strong, and using only the last sentence of the prompt yields higher accuracy than encoding the whole prompt, with the best model reaching 76.5% test accuracy (Srinivasan et al., 2018). The benchmark therefore measures matched continuation under significant stylistic and distributional artifacts.

“Incorporating Structured Commonsense Knowledge in Story Completion” keeps the same Story Cloze setting but replaces simple matching with a three-signal continuation scorer over narrative sequence, sentiment evolution, and ConceptNet-derived commonsense knowledge (Chen et al., 2018). The full model reaches 87.6% accuracy, outperforming FTLM at 86.5%, and removing the knowledge component drops performance to 85.6% (Chen et al., 2018). The underlying point is methodological: in matched continuation, both endings can be locally fluent, so the model must compare them along event progression, affective trajectory, and real-world plausibility, not only lexical overlap.

A dialogue-specialized version appears in “Telling Stories through Multi-User Dialogue by Modeling Character Relations” (Si et al., 2021). There the continuation unit is the next character response in CRD3, the context includes a summary of the story so far, pairwise relationship cluster labels, and the last $n$-turns of character interactions, and the model ranks the gold response among ten candidates with a bi-encoder architecture (Si et al., 2021). The best configuration, Base+Summary+Relations with multitask next-character prediction, reaches Hits@1/10 = 21.3 and Hits@5/10 = 74.6 (Si et al., 2021). This formulation makes the matched continuation problem explicitly social: the right continuation is the one that best fits plot, speaker identity, and relation structure.

These textual variants share a common constraint: the continuation is evaluated relative to a fixed prompt rather than judged in isolation. They differ in what counts as a match—ending plausibility, next-turn appropriateness, or goal-conditioned fit—but all treat continuation as context-sensitive discrimination rather than unrestricted generation.

3. Procedural, goal-directed, and controllable matching

A major extension of the paradigm replaces surface coherence with explicit latent-state tracking. “Transferring Procedural Knowledge across Commonsense Tasks” presents LEAP, whose source task TRIP contains pairs $(P,P')$ of plausible and implausible stories with three supervision layers: participant-level preconditions and effects for attributes across time, a pair of conflicting sentences in the implausible story, and a story plausibility label $P_{plau} \in \{0,1\}$ (Jiang et al., 2023). Its joint story model aligns common participants across the two stories and predicts plausibility from the pair, which makes the setup especially close to matched continuation-style evaluation (Jiang et al., 2023). The explanation chain is built into the model: participant extraction $\rightarrow$ stepwise attribute prediction $\rightarrow$ conflict detection $\rightarrow$ story plausibility (Jiang et al., 2023).

LEAP also demonstrates that matched continuation can be made explanatory rather than merely contrastive. On zero-shot transfer tasks, it evaluates on ROCStories, CODAH, PIQA, aNLI, and RICA; with augmentation it reaches 90.6 on ROCStories, 68.7 on CODAH, 68.6 on PIQA, 71.8 on aNLI, and 57.5 on RICA (Jiang et al., 2023). The paper’s most specific design lesson is that participant-state tracking transfers better to continuation selection than conflict detection alone: removing entity attribute loss causes a large drop on ROCStories, whereas removing conflict sentence loss barely hurts ROCStories and slightly improves PIQA and aNLI (Jiang et al., 2023). This suggests that matched continuations are often “less commonsensical but not necessarily conflicting.”

A different route appears in “Goal-Directed Story Generation” (Alabdulkarim et al., 2021). Its KG-DQN method samples 25 candidate next sentences from GPT2-sci-fi and selects one using a DQN over a knowledge-graph state extracted from the story so far (Alabdulkarim et al., 2021). Although the paper is framed as goal-directed generation toward a target VerbNet class, KG-DQN is structurally a continuation matcher: it learns $Q_\theta(G,a)$, a score over story state $G$ and candidate continuation $40\%,60\%,80\%,90\%$0 (Alabdulkarim et al., 2021). The result is strongly goal-sensitive candidate selection, with goal achievement for admire-31.2 rising from 16.74% for GPT2-Sci-Fi to 91.74% for KG-DQN (Alabdulkarim et al., 2021).

“Generating Diverse Story Continuations with Controllable Semantics” generalizes the paradigm from selecting one matched continuation to generating matched alternative sets (Tu et al., 2019). The model conditions on a control attribute $40\%,60\%,80\%,90\%$1 and learns $40\%,60\%,80\%,90\%$2, where $40\%,60\%,80\%,90\%$3 is the story context and $40\%,60\%,80\%,90\%$4 the next sentence (Tu et al., 2019). By varying $40\%,60\%,80\%,90\%$5, the same context can be paired with multiple continuations matched on the prompt but contrastive in sentiment, length, predicates, semantic frames, or induced clusters (Tu et al., 2019). The paper finds that semantic frames provide a stronger combination of diversity and quality than other control variables, which makes them especially useful for constructing contrastive matched continuations (Tu et al., 2019).

4. Metrics and diagnostic criteria

The earliest matched continuation setups are dominated by accuracy and retrieval metrics, but later work measures whether the continuation is matched for the right reasons. LEAP’s in-domain evaluation on TRIP includes Accuracy, Consistency, and Verifiability; Verifiability requires a correct story label, correct conflict identification, and correct conflicting participant attributes in the conflict sentences (Jiang et al., 2023). This turns continuation selection into explanation-sensitive evaluation.

“Imagined versus Remembered Stories” introduces sequentiality as a measure of narrative flow, defined by comparing the probability of a sentence with only its topic summary versus the topic plus preceding context:

$40\%,60\%,80\%,90\%$6

(Sap et al., 2022). The study uses matched remembered and imagined stories built from Hippocorpus topic summaries and shows that imagined stories have higher sequentiality than autobiographical stories, while retold autobiographical stories become more sequential over time (Sap et al., 2022). In a matched continuation perspective, sequentiality measures how strongly the next sentence is licensed by the preceding discourse rather than by topic alone.

“LLMs Preserve Semantic Isotopies in Story Continuations” contributes a semantic continuity audit rather than a contrastive benchmark (Cavazza, 6 Oct 2025). It generates continuations for 10,000 ROCStories prompts with five LLMs, validates GPT-4o as an isotopy extractor to 69.4% on a linguistic benchmark, and keeps one main isotopy per text, removing 2.58% of isotopies that fail to cover the primer (Cavazza, 6 Oct 2025). It then measures coverage, density, spread, and coverage balance, defined as the ratio of post-completion to pre-completion coverage; values close to $40\%,60\%,80\%,90\%$7 indicate balanced continuation of the isotopy across the prompt/continuation pivot (Cavazza, 6 Oct 2025). This is effectively an implicit matched continuation test without an explicit mismatched condition.

At the most controlled end, “Narrative Flattening” measures continuation behavior against matched human suffixes along thematic motion, affective prevalence, and linguistic diversity (Li et al., 27 May 2026). Across those dimensions, post-training compresses variation: thematic transitions become more uniform, high-intensity emotions give way to neutrality, and stylistic diversity across stories shrinks (Li et al., 27 May 2026). Because architecture, tokenizer, scale, and pretraining are held fixed across Base, SFT, DPO, and RLVR, the matched continuation design isolates post-training as the changing variable (Li et al., 27 May 2026).

5. Multimodal and visual continuation

In visual storytelling, the paradigm shifts from choosing the right sentence to generating the right next frame or video segment under matched visual context. “StoryDALL-E” explicitly proposes story continuation as a task distinct from standard story visualization: the first image $40\%,60\%,80\%,90\%$8 is given as a source frame, and the model generates the remaining frames conditioned on both captions and that source image (Maharana et al., 2022). Retro-fitted cross-attention layers use the source-image embeddings as keys and values, allowing the model to copy relevant visual elements from the initial frame and improve continuity in the generated visual story (Maharana et al., 2022).

AR-LDM extends this logic to latent diffusion by conditioning each new frame on history captions and generated images, with story continuation defined as the case where the first frame is given as a source frame (Pan et al., 2022). Its auto-regressive factorization,

$40\%,60\%,80\%,90\%$9

makes continuation the native operating mode rather than an auxiliary feature (Pan et al., 2022). The emphasis is on carrying forward character identity, scene layout, and underspecified narrative details that captions alone do not restate (Pan et al., 2022).

“StoryBench” turns this into a benchmark for continuous story visualization (Bugliarello et al., 2023). In its story continuation task, the model receives the first 0.5s of ground-truth video and a sequence of future action prompts with durations, and must generate the future segments so that they remain coherent with the matched prefix (Bugliarello et al., 2023). Because the future prompts, durations, and target frames come from the same annotated source video, the benchmark supports direct generated-to-ground-truth frame comparison through SIM, alongside FID, FVD, VTM, PQA, and human judgments of entity consistency, background consistency, and action realism (Bugliarello et al., 2023).

Recent work generalizes the visual paradigm beyond left-to-right continuation. StoryImager treats visualization, continuation, infilling, and backtracking as masked storyboard prediction, with a Target Frame Masking Strategy that adds noise only to masked target panels while preserving observed frames (Tao et al., 2024). ContextualStory conditions continuation on the first frame and the full storyline, using Spatially-Enhanced Temporal Attention, a Storyline Contextualizer, and a StoryFlow Adapter to maintain character and scene consistency across generated future frames (Zheng et al., 2024). AVC makes the matching step explicit: it retrieves the most semantically aligned prior frame with CLIP and adaptively restricts visual conditioning when the match is weak, rather than always conditioning on the last frame (Mousavi et al., 15 Oct 2025). Across these systems, “matched continuation” means source-grounded future generation rather than prompt-only synthesis.

6. Limitations, artifacts, and methodological tensions

A recurring issue is that matched continuation can be solved for the wrong reasons. Story Cloze is the clearest case: validation and test endings differ stylistically from training data, a no-context model is already strong, and last-sentence-only conditioning can outperform full-context conditioning (Srinivasan et al., 2018). The task is therefore a matched continuation benchmark whose alternatives are not cleanly controlled for annotation artifacts.

A second issue is underspecified or noisy context structure. In CRD3, pairwise character relations are automatically extracted from wiki descriptions, clustered into positive, negative, and neutral labels, and held constant over the entire story, even though true relations may evolve (Si et al., 2021). In the isotopy study, semantic matching is inferred from GPT-4o extraction without explicit human judgments on the main corpus, and there is no deliberately mismatched continuation baseline (Cavazza, 6 Oct 2025). In “Narrative Flattening,” Base uses raw completion while instruction-tuned checkpoints use a chat template, so the Base $n$0 SFT step remains partly an interface-plus-post-training comparison (Li et al., 27 May 2026).

A third tension concerns what exactly should be controlled. LEAP shows how to control procedural plausibility by participant-state constraints, but it does not construct a benchmark with explicit matching on topic, length, style, or lexical overlap between alternatives (Jiang et al., 2023). This suggests that continuation research separates into two partially independent problems: artifact control and causal or procedural control. The former reduces superficial stylistic shortcuts; the latter makes the negative option wrong for an interpretable reason.

Human-in-the-loop settings introduce another complication. “Conveying the Predicted Future to Users” presents a matched-context writing assistance study in which the same 20-sentence story snippet is paired with four alternative future plot descriptions, but all four hints are shown simultaneously rather than in isolated single-system conditions (Huang et al., 2023). The design is matched by context, yet it is not a strict matched continuation experiment because the causal effect of any single hint is not isolated (Huang et al., 2023).

Taken together, these studies indicate that the matched story-continuation paradigm is not a single benchmark recipe but a methodological family. Its strongest forms pair shared narrative context with controlled alternatives, explicit state or semantic criteria, and diagnostics that go beyond final-choice accuracy. Its most informative recent variants compare model continuations against matched human suffixes, explanation chains, or source-grounded visual futures, thereby turning continuation from a generic generation problem into a controlled probe of narrative consistency (Li et al., 27 May 2026, Jiang et al., 2023, Bugliarello et al., 2023).