Re3: Generating Longer Stories With Recursive Reprompting and Revision

Motivation

• Long-form story generation requires maintaining plot coherence, premise relevance, narration style, and factual consistency across thousands of words.These demands are qualitatively different from generating prior short stories of five sentences to one or two paragraphs.Source: §1 · p. 1

Framework

• Re3 generates 2000–2500-word stories by recursively prompting GPT3 with selectively retrieved information from an overarching plan and the story so far.Its Rewrite module reranks alternate continuations, while its Edit module makes local edits for factual consistency.Source: §1 · pp. 1–2

Contributions

• Re3’s Plan and Draft modules are fully zero-shot, allowing stories an order of magnitude longer than prior work without restriction to a particular training domain.The Plan module augments a premise with a setting, characters, and outline.Source: §1 · p. 2

Evaluation

• Up to 14% more Re3 stories were judged coherent in overarching plot and up to 20% more relevant to the initial premise than similar-length GPT3-based baselines.Evaluators also predicted up to 83% of Re3 stories to be written by humans.Source: §1 · p. 2

Automatic Story Generation

• Prior story-generation methods separately use structured plans, rerankers, or iterative editing, whereas Re3 combines planning, coherence-and-relevance reranking, and conditional editing.Plan-related approaches include outlines, latent variables, slot-filling, keywords, and control codes; Re3 additionally detects when edits are required.Source: §2 · p. 2
• Multi-thousand-word generation remains harder than typical five-sentence benchmarks because long-range plot coherence, premise relevance, and factual consistency degrade over extended horizons.Even forced long generations judged better than baselines can remain confusing and repetitive.Source: §2 · p. 2

Human-In-The-Loop Story Generation

• Unlike human-interactive systems that combine planning and revision for longer stories, Re3 focuses on fully automatic generation while remaining controllable through natural-language interaction.Both Re3’s planning and revision systems operate nearly entirely in natural language space.Source: §2 · pp. 2–3

Prompting

• Re3 treats prompting as a recursive subroutine, repeatedly recomposing prompts from contextually relevant plan and story information rather than generating the final output in one shot.This modular procedure extends beyond shorter-answer prompting methods and chain-of-thought prompting such as Kojima et al. (2022).Source: §2 · p. 3

3 Recursive Reprompting and Revision

• Re3 decomposes the writing process into Plan, Draft, Rewrite, and Edit modules.The framework repeatedly reuses generated planning components to compose prompts for subsequent story passages.Source: §3 · p. 3, §3.1 · p. 3

3.1 Plan Module

• The Plan module augments each premise with a setting, up to three characters, and a numbered story outline.GPT3-Instruct-175B generates these components, using rejection sampling for character names and resampling malformed outlines.Source: §3.1 · p. 3, Figure 2 · p. 3

3.2 Draft Module

• The Draft module generates fixed-length passages by combining relevant plan context, hierarchical story summaries, the current outline point, and the immediately preceding passage.Relevant character context is dynamically updated with entity recognition and retrieval, while GPT3-175B generates each continuation.Source: §3.2 · pp. 3–4, Figure 3 · p. 3

3.3 Rewrite Module

• The Rewrite module reranks alternative draft continuations for coherence with the previous passage and relevance to the current outline point.It uses Longformer-based coherence and relevance classifiers, plus heuristics that filter prompt repetition and first-person narration.Source: §3.3 · p. 4, Figure 4 · p. 4

3.4 Edit Module

• The Edit module detects and corrects long-range factual inconsistencies by maintaining character attribute-value dictionaries and applying controlled edits to the rewritten passage.It focuses on character attributes, extracts facts and attributes with language models, uses entailment checks, and sends editing instructions with the original continuation to GPT3 Edit API.Source: §3.4 · pp. 4–5

• RE3 generated stories of roughly 3000 tokens, or 2000–2500 words, using three-point outlines and four 256-token continuations per outline.Source: §4 · p. 5
• RE3 significantly and substantially outperformed ROLLING and ROLLING-FT in coherence and premise relevance, while producing fewer miscellaneous writing problems.Workers judged these differences significant in Table 1.Source: Table 1 · p. 6, §4 · p. 6
• Mechanical Turk workers compared RE3 with each baseline on interest, coherence, relevance, humanlikeness, and five categories of writing problems.Each story pair was judged by three workers, with separate pairwise comparisons for each baseline.Source: §4 · p. 6
• 83.3% and 80.0% of RE3 stories were judged human-written in the two baseline comparisons.Source: §4 · p. 7
• Although RE3 can miss premise or outline elements and contain confusing or contradictory passages, its planning infrastructure can self-correct toward the original high-level plot.Table 3 illustrates correction of an initially incorrect character identity and premise or outline deviations.Source: §4 · p. 7, Table 3 · p. 7
• Ablations indicate that the Plan and Rewrite modules are critical to performance, whereas the Edit module makes little difference.Source: Table 4 · p. 8

5.1 Ablation Study

• Ablations show that the Plan and Rewrite modules are critical for overall plot coherence and premise relevance, whereas the Edit module contributes little.Remaining continuity problems include non-character-centric inconsistencies, changing facts, and omitted outline plot points.Source: §5.1 · p. 8

5.2 Further Analysis of Edit Module

• STRUCTURED-DETECT outperforms two entailment-based baselines for detecting character-based inconsistencies.The evaluation uses 200 input pairs derived from 50 constructed setup-and-story tuples, with consistency framed as a classification task.Source: §5.2 · p. 8, Table 5 · p. 8
• The simplest ENTAILMENT baseline achieves a ROC-AUC barely above chance performance of 0.5, underscoring the need for highly precise inconsistency detection.STRUCTURED-DETECT is designed to scale to longer passages, where its advantage may widen, although all systems perform poorly in the simplified evaluation.Source: §5.2 · p. 9
• Over multiple-thousand-word stories, Edit-module errors compound because detection misses inconsistencies and correction can struggle with large changes or introduce undesired edits.The module can fix isolated details, but full stories also contain setting and scene inconsistencies and false positives involving changing character attributes.Source: §5.2 · p. 9

Contributions

• Over 2000-word stories show Re3’s strong performance on long-range coherence and premise relevance.The framework is presented as an initial attempt to address these challenges in longer stories.Source: §6 · p. 9
• Re3 emulates human long-form writing with general-purpose language models, and its natural-language modules support potential adaptation and human control.Its dynamic reinjection of contextual information could also transfer to non-story domains.Source: §6 · p. 9

Future length

• Re3 stories remain short stories despite being an order of magnitude longer than prior work, motivating hierarchical outlines toward novellas and novels.The authors describe a 7500-word story generated with multiple levels of hierarchical outline generation.Source: §6 · p. 9

Evaluation

• Human evaluation is costly and noisy, making evaluation the greatest barrier to increasing story length and limiting larger ablation studies.The authors call for metrics detecting off-topic drift, contextual nonsensicality, and long-range factual contradictions.Source: §6 · p. 9

Limitations

• Despite plot coherence and premise faithfulness, Re3 remains below beginner human writers in long-range factual continuity, theme, setting, pacing, and foreshadowing.The current Edit module is not human-level, and incorporating these broader continuity dimensions remains nontrivial.Source: §6 · p. 9

• Evaluation difficulty limits sample sizes, detailed ablations, and formal assessment of stories longer than 2000–2500 words.Re3 can generate such longer stories, but they are not formally evaluated; evaluation costs substantially exceed story-generation API costs.Source: Limitations · p. 10
• Many prompts, reranking heuristics, and hyperparameters are manually selected, leaving substantial room to improve individual modules.These choices were not selected through careful validation because evaluation is difficult.Source: Limitations · p. 10
• Re3’s custom story-generation modules, including its character attribute-value dictionary, may require manual redesign for other domains.Adapting prompts and parameters beyond stories is not currently automatic.Source: Limitations · p. 10
• The Edit module does not improve the main metrics, has low absolute ROC-AUC even when beating baselines in a controlled setting, and handles only character-attribute contradictions.The module is intended to address long-range factual continuity, but its current implementation covers only one common class of errors.Source: Limitations · p. 10
• Re3’s performance may decrease in languages lacking strong general-purpose language models such as GPT3.Source: Limitations · p. 10

Ethics Statement

• Re3 focuses on story generation to reduce misuse risks, while its Edit module targets long-range factual consistency and could incorporate a real-world knowledge base for truthful generation.The authors frame story generation as comparatively innocuous and propose adapting factuality safeguards to support truthful generation.Source: Ethics Statement · p. 10
• Re3 may inherit and amplify GPT3’s biases in stories, but its model-agnostic framework could benefit from improved base-model debiasing or controlled generation methods.The passage cites Radford et al. (2019), Brown et al. (2020), Lucy and Bamman (2021), Dathathri et al. (2019), Krause et al. (2020), and Yang and Klein (2021).Source: Ethics Statement · p. 10
• Re3’s performance depends on the quality of its base language model and may therefore suffer on non-English languages.This limitation is explicitly tied to the generator’s quality.Source: Ethics Statement · p. 10

A Character Name Generation

• Character names are generated with GPT3-Instruct-175B from the premise, setting, and prior character descriptions, with each generated name and description added before the next name.The prompt design allows names already present in the premise to be copied directly.Source: A Character Name Generation · p. 13
• Table 6 illustrates the prompt used to generate the first name in the Plan module.The example prompt accompanies the premise, setting, and character-description context used by the name-generation procedure.Source: Table 6 · p. 13, A Character Name Generation · p. 13
• For each character, the system samples 10 names and filters problematic strings, punctuation, repeated novel strings, and names lacking two-word full-name structure.The heuristics exclude role or attribute terms such as “protagonist,” “age,” and “gender,” while preferring names with two words.Source: A Character Name Generation · p. 13
• The name-generation heuristics are sufficient for this work but leave room to improve occasional edge cases and fairness, including potential bias from GPT3.The passage identifies both generated-name quality and the generation system’s bias as limitations.Source: A Character Name Generation · p. 13

B Details on Additional Reranking Heuristics

• Rewrite first filters empty outputs and then reduces repetition by detecting repeated five-word sequences or sentences with sufficiently small edit distance.These checks target repetition both within a generation and between the generation and its prompt.Source: B Details on Additional Reranking… · p. 14
• Rewrite filters stylistically jarring generations containing commentary-like strings, excessive potentially ordinary-story markers, or paragraph colons near the beginning.Examples include “\nComment,” “copyright,” and likely analysis headers.Source: B Details on Additional Reranking… · p. 14
• Rewrite maintains consistent third-person narration by filtering continuations containing first- or second-person markers such as “I,” “we,” and “you” outside quotations.The heuristic detects these markers to avoid shifts in narrative perspective.Source: B Details on Additional Reranking… · p. 14

Attribute-value detection

• The detection system extracts character facts from passages, identifies attribute keys, regenerates their values, and filters results using agreement and entailment checks.It retains facts supported by at least 2 of 3 GPT3-Instruct-175B outputs, extracts keys with GPT3-Instruct-13B, discards hallucinated values, and recomputes values from the original facts.Source: C Details on Editing System… · pp. 14–15

Conflict resolution

• Conflicting attributes are compared with an entailment model: entailed updates replace old values, neutral relations preserve them, and contradictions are flagged for editing.Attribute-value pairs are converted into simple sentences before comparison.Source: C Details on Editing System… · p. 16

Attribute completion

• The system completes relationships involving other characters by inferring reciprocal attributes and matching detected names to existing characters.For example, a fact that Lucy is Karen’s friend supports reciprocal friend relations in the structured dictionary.Source: C Details on Editing System… · p. 16

Controlled evaluation

• For controlled evaluation, the system outputs continuous contradiction probabilities by taking the entailment model’s maximum contradiction probability across attribute-key conflicts.This modification enables ROC-AUC scoring instead of discrete contradiction decisions.Source: C Details on Editing System… · p. 16

Usage sources

• Filtering and reranking typically generate 10 outputs per call, producing substantial token usage from GPT3-175B and GPT3-Instruct-175B.Source: D Data on API Usage · p. 16
• The Edit module is responsible for most GPT3-Instruct-13B usage and some GPT3-Instruct-175B usage, and is the sole user of the Edit API, which involves rejection sampling when it makes no change or returns an overly lengthy response.Source: D Data on API Usage · p. 16

Cost and usage

• A single RE3 story costs a few dollars under the reported settings, while baselines and ablations require fewer API calls.These estimates use the same settings as the main experiments and average API usage across five runs.Source: D Data on API Usage · p. 16

• WritingPrompts (Fan et al., 2018) is the only preexisting story dataset, supporting training for the relevance and coherence rerankers and the ROLLING-FT generator.Source: E Dataset Usage · p. 16
• GPT3 produces WritingPrompts passage summaries for relevance-reranker training and generates contradictory setups and story beginnings for Edit-module analysis without using a preexisting dataset.Source: E Dataset Usage · p. 16
• All data used or generated for the paper, along with documentation, is available through the authors’ codebase.Source: E Dataset Usage · p. 16

Length comparison

• RE3-SHORT and RE3-LONG generated two or six 256-token passages per outline section, respectively, instead of the main setup’s four.All variants used the same three-section outlines; results were reported in Table 13.Source: F Length vs. Story Quality Analysis · p. 17
• With only 50 stories, the comparison was insufficient for clear quantitative conclusions, although annotators judged longer stories more interesting.Longer stories may also encounter writing problems because they contain more text.Source: F Length vs. Story Quality Analysis · p. 17

Length tradeoffs

• Over longer horizons, generation may become repetitive or lose the plot, whereas stopping too early may produce stories that seem truncated before reaching major plot points.The authors suggest dynamically determining passage length as future work to balance these risks.Source: F Length vs. Story Quality Analysis · p. 17

Baseline comparisons

• Aggregate writing-problem differences become significant even though differences in individual metrics are largely not significant, with RE3 never significantly worse.This pattern applies to comparisons with the ROLLING and ROLLING-FT baselines.Source: G Full Metrics for Miscellaneous… · p. 17, Table 14
• Table 14 reports the fraction of stories marked with each individual writing problem when comparing RE3 with ROLLING and ROLLING-FT.Bolded differences indicate statistical significance at p < 0.05.Source: Table 14

Ablation comparisons

• Table 15 reports individual writing-problem rates for RE3 compared with ablations removing the Plan, Rewrite, and Edit modules.Bolded differences indicate statistical significance at p < 0.05, while individual differences are largely not significant.Source: Table 15 · p. 18

H Mechanical Turk Evaluation Details

• Annotators compared two stories in a Mechanical Turk survey and were paid $1.50 per hit.The survey asked workers questions comparing the stories.Source: H Mechanical Turk Evaluation Details · p. 18
• Figure 6 illustrates the Mechanical Turk survey format, with most story text omitted from the example.Source: Figure 6 · p. 19

I Annotator Agreement

• Fleiss’ kappa agreement for individual metric annotations in pairwise comparisons between RE3 and baselines was relatively poor.Source: Table 16 · p. 20
• Fleiss’ kappa was usually positive but low for most comparisons between RE3 and ablations.Source: Table 17 · p. 20, I Annotator Agreement · p. 20
• The evaluation asked annotators to quickly read or skim two fairly lengthy passages so they could evaluate more stories.This procedure meant that many details could be missed.Source: I Annotator Agreement · p. 20
• Because many metrics were subjective, individual annotations were expected to be noisy and agreement was poor.The authors expected expert annotators to improve agreement but noted that this would substantially increase costs.Source: I Annotator Agreement · p. 20

J Example Stories

• RE3’s stories generally maintain a coherent overarching plot relevant to the initial premise, whereas ROLLING and ROLLING-FT often produce rambling, mutually irrelevant short passages.RE3 may deviate from the detailed outline generated by its Plan module, and some confusing passages remain.Source: J Example Stories · p. 20
• Some RE3 problems may result from evaluation-cost and methodological limitations, while a 7500-word example generated without those limitations may better represent the framework’s capabilities.The longer example is presented as an indication of the framework’s potential ceiling.Source: J Example Stories · p. 20

J.1 Examples for Premise 1

• For Premise 1, RE3 follows the premise of Diane learning that Mark is cheating, but its revenge execution is incomplete and the ending contains confusion, factual inconsistency, and lowercased names.The initial outline specified that Diane would have an affair of her own, which the generated story does not show.Source: Table 20
• ROLLING’s Premise 1 story deviates completely from the premise, reconciles the couple unexpectedly, ends oddly, and leaves Jonathon’s relationship unclear.The supplied description identifies complete premise deviation, rapid reconciliation, an odd ending, and confusion about Jonathon’s relation to the couple.Source: Table 21
• ROLLING-FT’s Premise 1 story is confusing about the protagonist’s infidelity, abandons the husband and Erin, shifts to bakery interactions, and inconsistently identifies Rory.The story description notes that the husband and Erin disappear after the opening and that Rory alternates between the main character and another character.Source: Table 22 · p. 31

J.2 Examples for Premise 2

• For Premise 2, RE3 produces an outline whose characters and plot make sense, then opens with Shannon meeting Hannah while discussing her relationship with Marcus and plans to move to a farm.The premise is described as somewhat exotic and potentially challenging to follow; the generated opening establishes Shannon, Hannah, Marcus, their work, and Shannon’s planned marriage and farm life.Source: Table 24 · p. 31, §J.2 · p. 31
• RE3’s Premise 2 story eventually becomes generally well-written and coherent, portraying Shannon and Jaxon’s love and Shannon’s return to her own time.The story’s second half centers on Shannon’s conflict over leaving Jaxon and ends with her entering the time machine and vanishing from the 19th century.Source: §J.2 · p. 34, Table 25
• RE3’s story has substantial setup problems: it never explicitly explains Shannon’s return to the 19th century and introduces Jaxon confusingly.The supplied story later clarifies that Jaxon is Shannon’s lover and helps her return to her own time.Source: §J.2 · pp. 33–34, Table 25
• ROLLING’s Premise 2 story is reasonably coherent overall but completely off-topic, following a boy who flies, encounters a dragon, and searches for a golden egg instead of Shannon.The story contains no mention of Shannon, who is female in the premise.Source: §J.2 · pp. 34–35, Table 26
• ROLLING-FT’s Premise 2 story mentions Shannon but otherwise fails to follow the premise and lacks a clear overarching plot, instead presenting disjointed scenes.The generated scenes follow Shannon through campus, class, and a group dinner without a clear plot connecting them.Source: §J.2 · pp. 37–39, Table 27 · p. 41

J.3 Examples for Premise 3

• RE3’s Premise 3 outline makes sense for the premise, and its story follows Payton, the sole survivor of a fatal accident, returning for one day to help Landon and others.Payton feels guilty about surviving while her friends died, asks God for permission to return, and encounters Dylan while beginning her planned good deeds.Source: Table 29 · p. 41, §J.3 · pp. 41–42
• RE3’s Premise 3 story is largely faithful at first but confusingly introduces the initial deal with God as a flashback before becoming disconnected when Payton meets Dylan and apparently returns to heaven.The second half implies that Payton and Dylan go back to heaven together.Source: Table 30
• ROLLING’s story shifts to Payton’s suicide attempt and Billie’s efforts to help her recover, making it largely irrelevant to the premise and narratively jarring around the asterisk markers.The continuation focuses on Payton’s trauma, Billie’s support, and their move into a new home.Source: §J.3 · pp. 46–47, Table 31
• ROLLING-FT places Payton in Heaven after her death, where she confronts her past, questions her identity, and awaits a possible cure while navigating an unfamiliar city.The excerpt centers on The Golden Dragon in Bilgravia and Payton’s internal conflict over being alive again and leaving her current body.Source: §J.3 · pp. 47–48
• ROLLING-FT’s Premise 3 story is extremely confusing, leaving Payton’s status between heaven and Earth unclear and centering on internal dialogue with The Voice before Ivy appears.The story’s confusion stems from unclear setting and shifting conversations.Source: Table 32 · p. 51

J.4 Examples for Premise 4

• RE3’s Premise 4 setup makes sense for the premise, introducing Shannon as a high-school senior caring for younger siblings after their mother’s death.The outline identifies characters and an overarching situation consistent with the relatively mundane premise.Source: Table 34 · p. 51, §J.4 · p. 51
• RE3’s Premise 4 story is initially coherent and premise-relevant, following Shannon’s responsibilities and her meeting a boy named Luke.The narrative develops Shannon’s family obligations before introducing Luke at school.Source: §J.4 · pp. 51–53, Table 35 · p. 54
• The Premise 4 story is cut off prematurely: its latter half remains broadly on topic but becomes a conversation among Shannon and her friends, while Luke is never mentioned again.The story concludes with discussion of Shannon’s siblings and a friend calling her “such a good sister.”Source: §J.4 · pp. 53–54, Table 35 · p. 54
• ROLLING shifts from the mother’s death to Shannon’s relationship with her father and ultimately reverses the mother’s apparent death.The generated story includes extensive domestic and school scenes involving Shannon and her father before presenting her mother as alive again.Source: §J.4 · pp. 54, 55, 57, Table 36
• Both ROLLING and ROLLING-FT are unfaithful to Premise 4, with ROLLING even ending by implying that Shannon’s mother is alive again.ROLLING begins with the mother’s death but becomes primarily about Shannon and her father; ROLLING-FT similarly departs from the premise throughout.Source: Table 37 · p. 61, Table 36
• ROLLING-FT changes Shannon into a mother, inconsistently portrays Brian as her boyfriend and son, and focuses on their movie-theater and Pokémon outing.The story follows Shannon and Brian through a theater visit, including conversations about movies and Pokémon games.Source: §J.4 · pp. 57, 59, Table 37 · p. 61

J.5 Examples for Premise 5

• The fifth example group is introduced by a relatively mundane premise.No generated story or further finding for Premise 5 is supplied in this section part.Source: Table 38
• RE3’s initial Premise 5 setup presents characters and an outline that make sense for the given premise.The outline centers on Sarah inheriting wealth, confronting corruption in her father’s company, and considering how to use her fortune.Source: Table 39 · p. 61, §J.5 · pp. 61, 63
• The RE3 example develops Sarah’s inheritance through family revelations, corporate fraud, emotional conflict, and possible charitable uses of the money.Sarah learns she controls her father’s estate, considers confronting corruption, worries about responsibility and guilt, and imagines helping animals or starting a charity.Source: §J.5 · pp. 61–64
• RE3’s Premise 5 story follows a coherent overarching plot that clearly matches the premise, despite inconsistencies about William’s death and the source of Sarah’s money.The story initially frames Sarah’s wealth as an inheritance, but later describes it as a lottery win; William also appears dead and alive at different points.Source: Table 40 · p. 64
• ROLLING’s Premise 5 story is relatively relevant because its unnamed protagonist finds money on her doorstep and struggles with what to do with it.However, it does not describe how she spends the money, does not show meaningful guilt, and ends with an odd change in narration.Source: Table 41
• The Premise 5 example is wholly irrelevant to its premise, following Sarah’s unclear captivity before suggesting that her captors are benevolent.The story centers on Sarah being held or kidnapped by an unknown woman and later portrays the woman and her group as seemingly benevolent.Source: Table 42 · p. 70

K Example Story Generation Steps

• The example begins with an initial setup containing the story’s setting and character descriptions, then a structured attribute-detection system infers attributes for Lila, Katarina, and Oliver from that setup.The inferred attribute-value dictionaries appear in Table 44, and Table 43 is labeled as the initial setup used for prompting examples.Source: K Example Story Generation Steps · p. 70, Table 44 · p. 70, Table 43 · p. 70
• The first passage repeats the setup, prompts the model with Chapter 1, generates ten continuations, and selects the best continuation after reranking.Because no previous story text exists, the initial prompt differs from later prompts; the selected continuation begins with the friends moving into apartments and pursuing an environmentally friendly mission.Source: K Example Story Generation Steps · pp. 70–71, Table 45 · p. 71, Table 46 · p. 71
• Later prompts detect new entities, update existing character attributes, rerank ten continuations, and mark non-character entities without building attribute dictionaries for them.The final prompt concatenates previous outline parts and explicitly indicates that the upcoming passage is the story’s end.Source: K Example Story Generation Steps · pp. 71–73, Table 47 · p. 72, Table 49 · p. 74

Example Evaluation Data

• The example evaluation data includes setups s and s′ paired with corresponding stories t and t′, differing in Beth’s relationship to Julie.Tables 50–53 present the original and altered setups and their corresponding stories.Source: L Example Data for Editing System… · p. 74, Table 50 · p. 75, Table 51 · p. 75

Longer generation example

• The adapted system generated an outline as long as needed and advanced between outline points when reranking judged the next continuation sufficiently worse than the previous one.These changes prioritize high-quality stories rather than stories of a predetermined length.Source: M Longer Story Example · p. 76
• A two-level hierarchical outline generated passages for lower-level points, while completed groups were collapsed into their high-level points in subsequent prompts.GPT3-Instruct-175B listed minor plot points under each initial outline point.Source: M Longer Story Example · p. 76
• The recursive hierarchy could enable exponentially longer stories without substantially increasing prompt length; the example reached roughly 7500 words with a 1024-token generator context.The authors still observed a gap from what an experienced human writer might produce.Source: M Longer Story Example · p. 76
• A single uncherry-picked story closely followed its premise and outline and maintained a highly coherent overarching plot with a clear beginning, middle, and end.The authors judged it higher quality than the shorter Appendix J examples, while noting a few odd details and inconsistencies that did not significantly reduce understandability.Source: M Longer Story Example · p. 76

• The story follows Mark and Sophie as they survive mutant attacks, leave town, and reach a safe haven while maintaining the outlined plot’s major transitions.The narrative moves from planning their departure, through encounters with mutants and survivors, to shelter in a hotel where they rest together.Source: M Longer Story Example · pp. 83, 85, 86, Table 56
• The example is judged relatively high quality, with clear plot structure, strong premise relevance, clear transitions, and generally well-maintained local consistency.The table notes that several problems found in shorter generations do not appear in this longer example.Source: Table 56
• The story still contains inconsistent character identities and personalities, vague and drifting motivations, setting drift, and inconsistent mutant strength.For example, Sophie appears as both a main character and a stranger, flying creatures disappear from the setting, and mutants shift from nearly invincible to normally killable.Source: Table 56