Suppose you want to rewrite a 3,000-word noir detective story as a fairy tale, or as hard science fiction, or as a news feature. The plot, the characters and who knows what should survive, while the voice, conventions and texture change. Doing this for $N$ genres naively means learning a rewriter for every ordered pair, which is $N(N-1)$ systems. Vision researchers met the same growth problem in image translation and solved it with a star layout: every domain connects only to one central "hub" domain, and an $A \to B$ translation passes through it. Long text adds a second difficulty. Language models drift over thousands of words: characters change attributes, earlier facts are forgotten, and the story wanders. We propose to treat a structured narrative state as the hub. This state is a plan-like record of events, character attributes and causal links, each written as a short natural-language description. A frozen language model gets two small trained modules. A reader turns a story in any genre into this state. A writer renders the state into any genre, given only a plain-language description of that genre.
The central claim is that one representation can do two jobs. Because the hub records who did what and what is true of whom, it is the object long-form coherence methods already plan with, and it is also what a cross-genre rewrite must preserve. Building translation around it gives linear scaling in the number of genres, since one reader and one writer serve all of them. It allows genres never seen in training, because each genre is specified by a description rather than by an index. It also turns the known failure of hub-based translation into something we can measure: when the hub carries too little information, the rewrite loses plot facts. The new pieces are the structured hub, description-conditioned reading and writing over a frozen backbone, and the measurement of the hub bottleneck on long text.
What prior work establishes and what it leaves open
Two pieces of prior work supply the scaling idea. ComboGAN gives each domain one encoder and one decoder, so an $A \to B$ translation runs encoder $A$ then decoder $B$. This cuts the pairwise growth catalogued in Linear Scalability for Multi-Domain Image Translation from quadratic to linear. Universal Multi-Domain Translation via Diffusion Routers goes further. A single denoising network, conditioned on source and target domain labels, is trained on only $K-1$ hub-paired datasets and still translates between pairs it never saw, reaching FID 6.12 against UniDiffuser's 11.02 on Segment→Face. That paper also names our main risk: transitivity fails when the hub lacks the mutual information that links the peripheral domains. Both works operate on static images with categorical domain labels. We move to long sequential text, and we make the hub itself the object of study rather than a fixed modality such as a segmentation map.
Narrative-coherence work supplies the hub content. Re3 plans with character portraits and outlines, drafts from distilled summaries instead of raw context, and checks consistency with an attribute dictionary. It reports a 14% absolute gain in plot coherence and 20% in premise relevance over a rolling-window GPT-3 baseline. It also reports that its rule-based edit step was weak, and that human judges credited quality mostly to reranking. Hierarchical Neural Story Generation and DiscoDVT (Ji et al.) likewise generate from a premise or discrete discourse plan. COINS represents causal links as natural-language inference rules, but only for five-sentence stories. All of these treat the plan as scaffolding for generating new text in a single domain. We use it as an interlingua: something extracted from an existing text and re-rendered elsewhere. That gives the plan a precise success criterion, since whatever is lost in the round trip is measurable.
Semantic label embedding supplies the conditioning. Work such as Chai et al. and Bapna et al. replaces class indices with natural-language descriptions so that models can reach unseen classes. StructCBM classifies by similarity to prototypical and discriminative concept descriptions, reaching 85.45% 10-shot accuracy on AGNews. We reuse the description idea in two places: genre identity, so that an unseen genre is just a new paragraph, and hub fields, so that each fact can be checked by entailment rather than string matching. We use StructCBM's prototypical concepts and deliberately avoid its pairwise discriminative concepts, which grow as $O(C^2)$ and would reintroduce the very scaling we are removing.
Work on representation rigidity acts as a constraint. The Representational Bottleneck of Frozen Foundation Models argues that lightweight modules on a frozen model can only reweight features the model already has. Shakeri et al. found that prompt learning fails to beat a linear probe when the backbone lacks the relevant grounding. This predicts that description-conditioning will work for genres the backbone knows well and fail for styles it never saw. ABBA-Adapters, which obtain rank up to $r^2$ from a Hadamard product of two low-rank updates and beat LoRA on GSM8K (66.26% vs 61.94% on Mistral-7B), give a method for probing whether a more expressive update moves that ceiling. Res-Tuning shows that detached bypass modules let many tasks share one backbone forward pass. We use this only as a serving argument, not as a core ingredient.
Core Claim
A frozen language model with one description-conditioned reader and one description-conditioned writer, joined by a structured narrative-state hub, should satisfy three predictions. First, it should rewrite long stories between any pair of genres at quality comparable to pair-specific systems. Second, it should handle held-out genres given only a description. Third, it should preserve plot and character facts over thousands of words better than direct chunk-by-chunk rewriting, because the writer is anchored to an explicit state rather than to a drifting rolling context.
The hub bottleneck can be stated as a quantity. Let $F(x)$ be the set of atomic facts in source story $x$, extracted by a separate annotator model that the system never sees. Let $y$ be the rewrite. Then fact retention is
$$R(x,y) = \frac{1}{|F(x)|}\sum_{f \in F(x)} \mathbb{1}[\,y \text{ entails } f\,],$$
with entailment judged by a natural-language inference model. Under star routing, $R$ is capped by how much of $F(x)$ the hub records. Varying hub richness therefore traces a retention curve. That curve is the textual analogue of the transitivity failure noted for diffusion routers, and to our knowledge nobody has measured it.
Research Program
The backbone is a frozen open model of about 7B parameters, a size DeFine already shows is viable for long-form articles. The hub is a typed record with three parts: a scene-level event sequence, a character table whose attributes are short descriptions ("distrusts the police since chapter one"), and causal links between events written in a COINS-style rule format. We do not need parallel genre corpora, which is the main practical advantage of the star layout. For each genre we collect monolingual stories and produce silver hub states with a strong annotator model. These (story, state) pairs are the $K-1$ hub-connected datasets. The reader is trained to map story to state, with the source-genre description in its input. The writer is trained to map state plus genre description to story. Generation follows Re3: scene by scene, conditioned on the current state slice and a short summary, never on the full prior text.
The genre description is a paragraph covering conventions, voice and typical devices. It is embedded by the frozen backbone and fed to the modules as a soft prefix. It also appears as text, so the modules see both a vector and readable instructions. Both modules are low-rank adapters. A matched-budget ABBA variant tests whether rank is the limit on rare genres.
The analysis has three threads. The first measures the retention curve by training hubs of increasing richness. The second compares indirect routing (read then write) with a direct variant distilled from it, in the spirit of direct diffusion routing, to test whether the explicit hub can be skipped at inference without losing retention. The third separates genres by how familiar the backbone is with them. We construct some deliberately unfamiliar target styles, such as invented formal constraints and a made-up narrative register defined only by description and a few dozen examples. This tests where description-conditioning on a frozen model breaks, as the rigidity literature predicts it will.
Baselines
The naive reference is the frozen backbone prompted to rewrite the story one scene at a time into the target genre. It receives the genre description and a rolling window of prior output. It is the cheapest system, and it is what a practitioner would try first. The second baseline is the same backbone with the full source story in its long context, rewriting in one pass. This tests whether context length alone solves drift, which Beyond a Million Tokens and Re3 suggest it does not.
The strongest competitor is per-target-genre LoRA. One adapter per target genre is fine-tuned to map (source text, chunk-by-chunk) to target-genre text, using the same silver data rendered as pseudo-parallel pairs through the annotator. It is the strongest competitor because it already scales linearly with genre count and has a dedicated set of parameters per genre. The comparison is fair only if every system uses the same backbone, the same per-genre training text, the same total adapter parameter count, and the same hyperparameter search budget.
There is no reported result on Lacuna for cross-genre long-story rewriting. The nearest reported figures are Re3's +14% absolute plot coherence over a rolling-window baseline for generation, and the router's FID 6.12 vs 11.02 for multi-domain image translation. Beating the per-genre adapters means three things together. On seen genres, our system must match them in human preference within a few points while showing higher fact retention $R$. On held-out genres, where per-genre adapters cannot be trained at all, it must clearly beat prompting.
The key ablation replaces the structured hub with a free-text summary of equal token length. This isolates whether structure, and not merely compression, drives retention. A second ablation replaces genre descriptions with learned genre-index embeddings. It removes zero-shot ability by construction and shows how much of seen-genre quality comes from the descriptions themselves.
Why not just…?
Why not just prompt a strong frozen model with "rewrite this in genre B"? This is the most serious threat. For genres heavily represented in pretraining, a large prompted model may match or beat us on style. Our bet is that it loses on long-horizon fact retention, because rolling-window rewriting is the regime where Re3 found plot drift. The prompting baseline is run at matched backbone size, and $R$ is plotted against story position. If prompting holds $R$ flat out to 3,000 words, the coherence half of our claim fails, and the contribution shrinks to scaling and zero-shot conditioning.
Why not just train one LoRA per target genre? This is already linear in $N$, so the scaling argument alone does not favor us. Our advantages are narrower and testable: held-out genres need no training, and data is pooled across genres. If per-genre adapters win clearly on seen genres, we will say so and report the system as the right choice for open-ended or low-resource genre sets, not as a uniformly better rewriter.
Why not just summarize the source and regenerate from the summary? This is a hub with no structure, and it is exactly the summary ablation above. If an equal-length summary retains as many facts as the typed state, then the structured hub is unnecessary and the simpler pipeline should be used.
Why not just use a long-context model on the whole story? Re3 found that what the model remembers matters more than how much, and Beyond a Million Tokens reports that long-context models still lose track of characters. We include the full-context baseline rather than assume it fails.
Proposed Experiments
The data comes from two sources. WritingPrompts stories (303,358 prompt-story pairs, average 734 words) are sorted into roughly a dozen genres by a description-based classifier, and multi-scene stories are built up to 3,000 words with Re3-style generation where human text is too short. DeFine articles serve as a non-fiction domain. Four genres are held out entirely. Evaluation uses three measures. Fact retention $R$ is computed against annotator-extracted facts, with a human audit of a sample of entailment judgments. Genre adherence is judged by a prototype-concept classifier in the style of StructCBM, trained on held-out data, plus human ratings. Human pairwise judgments follow Re3's protocol for coherence and relevance.
Three checks decide the claims. On seen genres, the hub system must reach per-genre-LoRA preference parity while showing higher $R$ beyond the first thousand words. On held-out genres, it must beat prompting on both $R$ and adherence. The retention-versus-hub-richness curve must rise and then saturate, and structured hubs must dominate equal-length summaries. On the constructed unfamiliar styles, we expect the adherence gap between our system and the per-genre adapters to widen. This would be direct evidence of the rigidity ceiling for description-conditioning. Whether ABBA's higher rank closes that gap tells us whether the limit lies in the update's rank or in the frozen backbone's features.
References
- ComboGAN: Unrestricted Scalability for Image Domain Translation
- StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation
- Universal Multi-Domain Translation via Diffusion Routers
- Re3: Generating Longer Stories With Recursive Reprompting and Revision
- Hierarchical Neural Story Generation
- DiscoDVT: Generating Long Text with Discourse-Aware Discrete Variational Transformer
- COINS: Dynamically Generating COntextualized Inference Rules for Narrative Story Completion
- Adaptive Concept Discovery for Interpretable Few-Shot Text Classification
- Learning Federated Visual Prompt in Null Space for MRI Reconstruction
- Few-Shot Adaptation of Medical Vision-Language Models
- ABBA-Adapters: Efficient and Expressive Fine-Tuning of Foundation Models
- Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone
- DeFine: A Decomposed and Fine-Grained Annotated Dataset for Long-form Article Generation
- Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs