Writers, studios, and institutions often hold collections of long, structured narrative text that they will not pool: unpublished manuscripts, scripts in a house format, or case narratives written to a local template. Each collection is too small to teach a model to sustain a plot over thousands of words, and each follows its own formal rules, such as stanza and meter, slugline-and-dialogue layout, or a fixed section order. This proposal asks a narrow question. When several such collections are trained together by federated learning, meaning each site trains on its own data and sends only model updates to a server that averages them, which part of a story generator should be averaged? Our bet is that the plot plan travels well across sites and the surface wording does not. The model that decides "what happens next, to whom, and how it connects to the previous scene" should be shared. The model that turns each scene into text, and the hard rules on that text, should stay at the site.
The plan in two sentences. Each site automatically converts its own documents into pairs of an abstract plot outline and text, using structure the documents already contain: segment boundaries, recurring named characters replaced by placeholder slots, connective words between segments, and measured form properties. No one labels anything, which is what "distant supervision" means here. Sites jointly train only an outline model on these placeholder-based outlines, while each site keeps a private text writer and enforces its formal rules exactly at decoding time with a local rule-checking automaton. We test whether this split gives longer-range coherence than training alone, and better readability under hard constraints than averaging the whole generator.
Core Claim
Long-form generation fails in two separate ways, and they respond to data pooling in opposite directions. Narrative drift is the first: a story loses its premise, drops causal threads, or contradicts what it said about a character. Rolling-window generation drifts badly, and explicit plan-then-draft pipelines reduce this. Re3 generates an outline, drafts against it, and reranks and edits the drafts; it reports a 14% absolute gain in stories judged to have a coherent overarching plot over rolling-window GPT-3. DiscoDVT plans discrete latent codes, trained with an auxiliary objective that predicts the discourse relation between adjacent blocks. Learning to plan well needs many stories, and plot structure (setup, complication, causal and temporal links, character recurrence) is plausibly shared across genres. That makes the plan the part that benefits from pooling.
Formal compliance is the second failure. Scaling a language model does not make it follow hard symbolic rules (Can a Computer Really Write Poetry?). Exact compliance comes from symbolic machinery at decoding time: the finite-state acceptors of Hafez, or the decision diagrams and constraint solvers of Intertwining CP and NLP. That machinery has a known cost. When the model's word distribution is far from what the rules permit, pruning leaves fluent-looking but meaningless text. Deep-speare scores above human poets on meter (4.1 vs. 3.8) and rhyme (4.4 vs. 4.1), yet readability is 2.7 vs. 4.8.
Our specific hypothesis is that averaging the text writer across sites with different forms moves each site's word distribution away from its own rules. The local automaton then has to prune more, and readability falls. Federating only the planner should avoid this while still giving small sites the coherence benefit of pooled plot data. The central contribution is a measured answer to which layer of a two-level generator should be federated when forms differ across sites.
This differs from the closest work. FedMRG federates low-rank adapters (LoRA, small trainable weight matrices added to a frozen model) over the whole report generator, pairing a shared adapter with a site-specific one. FedSUMM modulates gradients by how far local predictions deviate from global ones. Both target short outputs and split shared from personal parameters along the adapter axis, not along the plan-versus-text axis. Resource-Adaptive Federated Text Generation conditions on fixed control codes such as star ratings and works only with short reviews and abstracts. None of these handles long-horizon planning or hard formal rules.
Research Program
The program has four parts.
Building outlines from documents. Each site runs a deterministic extractor over its documents. Segments come from structure already present: paragraph or scene breaks, stanza boundaries, and script sluglines. For each segment the outline records four things:
- the characters present, replaced by consistent placeholder slots (CHAR1, PLACE2);
- a handful of content keywords, extracted in the spirit of Hafez's topical words;
- the discourse relation to the previous segment, read from explicit connectives using the relation inventory DiscoDVT trains on;
- a form specification measured by tools: line and syllable counts from the CMU pronunciation dictionary, rhyme classes, and for SongCi the tone pattern implied by the Cipai template, as in MRCG.
This mirrors how WikiAsp, Summary Cloze, and BioGen mine headers, citations, and infoboxes for supervision. Those projects found that such alignments are noisy and that models under-use high-level guidance. We take the second finding as a warning: the outline must be consumed by a dedicated planner, not simply concatenated onto the input.
Some extracted keywords will not actually be supported by the segment text. Following Partially-Aligned Data-to-Text Generation, each site computes token-level supportiveness scores locally and down-weights the writer's loss on unsupported tokens. Because this happens on-site, it needs no central ground truth.
The shared planner. The planner maps a premise plus a genre tag to a sequence of placeholder outlines. It is a LoRA adapter on an open 8-billion-parameter base model, the scale Attribute-Guided Genre Expansion used. It is trained by standard federated averaging. Because outlines contain slots and keywords rather than prose, what the planner sees is already abstracted. We will measure leakage rather than assume it away.
The local writer and the rule automaton. Each site trains its own writer adapter to expand one outline segment into text, conditioned on the full outline and a running summary of earlier segments, as in Re3's drafting stage. The site's formal rules are compiled into an automaton that masks disallowed tokens during decoding. The aggregated parameters never see the automaton, which avoids mixing hard pruning into averaged gradients.
A diagnostic for mismatch between writer and rules. We define constraint pressure for a generated text $x$ of length $T$ as
$$\pi(x) = -\frac{1}{T}\sum_{t=1}^{T} \log \sum_{w \in A_t} p(w \mid x_{<t}),$$
where $A_t$ is the set of tokens the automaton allows at step $t$ and $p$ is the writer's next-token distribution. A large $\pi$ means the writer keeps wanting tokens the rules forbid. The hypothesis predicts that a fully federated writer will show higher $\pi$ than a local one, and that $\pi$ will correlate negatively with human readability ratings.
Baselines
Base model only. The naive reference is the frozen base model prompted with the premise and a description of the form, decoded with the same local automaton. It shows what exact rule enforcement gives with no training at all. A second weak reference is local-only training: each site trains its own planner and writer on its own data with the same adapter budget. This is the "no federation" point and should be weakest for the smallest sites.
Whole-generator federation with personalization. The strongest existing federated competitor is a FedMRG-style dual adapter applied to the whole generator: one averaged generic adapter and one site-specific adapter, coupled by distillation. It is the strongest because it already lets each site keep private style while sharing everything else, so it is the direct alternative to our plan/text split. FedMRG reports BLEU-4 12.68 and clinical-efficacy F1 33.9 on MIMIC-CXR. That is a different task, so the number only calibrates that the method works in its home setting. We also run FedSUMM's discrepancy-modulated averaging, which reports ROUGE-1 1.9% above FedAvg on CSL.
Coherence and form reference points. On coherence, we run a Re3-style plan, draft, and rerank pipeline on the base model, with no training, as the reference. Re3 reports +14% overarching-plot coherence and +20% premise relevance over rolling-window GPT-3. On form, PoetryDiffusion reports 95.37% rhyme and 91.64% tone accuracy on SongCi, against ChatGPT's 51.55% and 70.23% (PoetryDiffusion). Our automaton guarantees compliance, so the contest on form is readability under compliance, not compliance itself.
Upper bound. The ceiling is centralized training on the pooled data with the same architecture.
Ablations. Two ablations isolate the key ingredients:
- federating the writer while keeping the planner local, which reverses our split;
- replacing distantly extracted outlines with no outlines, which is plain federated fine-tuning.
Fairness and the bar to beat. All systems use the same base model, adapter rank, number of communication rounds, bytes transmitted, local steps, and a matched hyperparameter search of equal size. There is no reported result on Lacuna for federated long-form generation under hard form constraints, so every number in this setting is new.
Beating the whole-generator baseline means two things together. Pairwise coherence must be equal or better. Readability under constraints must be higher by a margin that holds over three seeds. The plan-only split must also close at least half the coherence gap between local-only training and the centralized ceiling for the smallest sites.
Why not just…?
Why not just prompt a strong open model locally with Re3-style planning and constrained decoding, and skip training? This is the most serious threat, and it is a baseline here, not a straw man. The evidence against it is partial. Unconstrained LLMs miss formal rules badly, with ChatGPT at 51.55% rhyme accuracy on SongCi. Adding an automaton fixes compliance but invites Deep-speare-style word salad when the model's preferences fight the rules, and our constraint-pressure measure tests exactly that. If the training-free pipeline matches federated planning on coherence and readability, we will report that federation is unnecessary for this purpose.
Why not federate the whole generator with personalized adapters, as FedMRG does? That is the strongest existing method and the comparison the proposal stands or falls on. The writer-federated ablation and the dual-adapter baseline will show whether averaging surface parameters raises constraint pressure and lowers readability. If the dual adapter already shields form, the plan/text split adds nothing beyond fewer communicated parameters, and we would say so.
Why not have sites release differentially private synthetic stories and train centrally? Differential privacy adds calibrated noise so individual training records cannot be recovered. Resource-Adaptive Federated Text Generation shows this works for short reviews and abstracts with fixed control codes. It has not been shown to preserve multi-thousand-word causal structure, and private noise tends to erase rare long-range dependencies first. We will run one variant: private synthetic outlines plus text pooled centrally, compared at a matched privacy level.
Why not train each site alone? For large sites this may be enough. The prediction is that gains concentrate in small sites, so results are reported per site, sorted by data size.
Proposed Experiments
Sites and data. We simulate sites with non-uniform data by genre. The 13 genres of the Multi-Genre Collection (stage play, movie script, rap, podcast, and others) become 13 sites. We add a Wikiplots site and a WritingPrompts site for long prose, an English sonnet site using the SONNET and PoetryDiffusion corpora, and SongCi split across sites by Cipai family. Site sizes are deliberately skewed by subsampling so that some sites are small.
Evaluation. Coherence is judged by human pairwise preference on overarching plot and premise relevance, following Re3's protocol. Automatic coherence checks are repetition rates (rep-8 and rep-16) and character-slot consistency, measured against the planner's outlines. Form is measured with deterministic checkers. Readability and emotion are rated on Deep-speare's 1–5 expert scale, and genre fidelity on WritingBench. Leakage is measured by copied 5-grams from each site's training text, with Hafez's 1.2 per sonnet as a reference, and by a membership-inference test on the shared planner.
Decisive checks. There are three.
- Does the plan-only split match or beat the whole-generator baseline on coherence while scoring higher on readability under constraints?
- Is that readability gain mediated by lower constraint pressure?
- Do small sites recover a substantial share of the centralized coherence ceiling?
A negative answer to the first check, with the training-free Re3-plus-automaton pipeline close behind, would be a useful result in its own right. It would mean the right place for shared learning in this setting is not where the field has been putting it.
References
- Re3: Generating Longer Stories With Recursive Reprompting and Revision
- DiscoDVT: Generating Long Text with Discourse-Aware Discrete Variational Transformer
- Can a Computer Really Write Poetry?
- Generating Topical Poetry
- Intertwining CP and NLP: The Generation of Unreasonably Constrained Sentences
- Deep-speare: A joint neural model of poetic language, meter and rhyme
- LLM-driven Medical Report Generation via Communication-efficient Heterogeneous Federated Learning
- Personalized Federated Learning via Gradient Modulation for Heterogeneous Text Summarization
- Resource-Adaptive Federated Text Generation with Differential Privacy
- Generating Chinese Ci with Designated Metrical Structure
- WikiAsp: A Dataset for Multi-domain Aspect-based Summarization
- Summary Cloze: A New Task for Content Selection in Topic-Focused Summarization
- BioGen: Generating Biography Summary under Table Guidance on Wikipedia
- Partially-Aligned Data-to-Text Generation with Distant Supervision
- Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion
- PoetryDiffusion: Towards Joint Semantic and Metrical Manipulation in Poetry Generation