Why Discourse Coherence Remains the Hardest Problem in NLP — and What External Scaffolding Can and Cannot Do About It

Consider two paragraphs:

Maria set the ancient vase on the wobbly table. It had been in her family for six generations. The wood groaned under the weight.

The detective examined the scene. Maria had told him the vase was priceless. He noted the scratch marks on the table’s surface and wondered whether it could support the artifact much longer.

Most readers resolve it in the third sentence to the table, not the vase. They infer that the artifact in the second paragraph refers back to the vase. They understand the detective arrived after the table groaned, and that Maria’s statement to him occurred in some implied window between the groaning and the examination. A causal thread runs through it: the table is wobbly, the vase is heavy, the wood groans, the detective notices scratch marks, a question about structural integrity surfaces. None of this is stated. It is inferred through a web of propositional commitments that accrue across sentences, paragraphs, scenes.

A language model will likely produce text that reads similarly. Probe what the model internally represents, though, and the picture shifts. There is no variable binding the pronoun it to a specific discourse entity with a persistence contract. No representation of the table’s physical state evolving across paragraphs. No commitment space tracking what Maria told the detective and when. What exists is a sequence of tokens, each predicted from the context window, each fluently appropriate to the local surface. The gap between what the text looks like and what the model internally represents — that is the gap I want to examine.

What Discourse Coherence Actually Requires

Discourse coherence is not cohesion. Cohesion is the surface stitching — pronouns, conjunctions, lexical repetition, discourse markers — that makes a text feel connected. Coherence is the deeper property: the text makes sense as a whole. Entities persist. Events unfold in consistent order. Causes have effects. Speakers are committed to what they have said, and those commitments constrain what they can say next.

Linguists working in discourse theory typically distinguish at least four dimensions of coherence that a narrative must maintain:

• Entity continuity: If Maria is introduced in paragraph one, references to her in paragraph five must resolve to the same individual, not a statistically plausible substitute. The model must maintain what Hobbs called a coherence graph — a structure tracking which entities are salient, which are backgrounded, which have been replaced.
• Event ordering: Events in a narrative have temporal and causal structure. If a character dies in chapter three, they cannot speak in chapter seven without explanation. This requires representing event ordering as a partial order, not merely a sequence of mentions.
• Causal consistency: If a cause is established, its effects must follow. If a rule of the world is stated — the vase is fragile — later events must respect it.
• Speaker commitment: In dialogue and narration, speakers commit to propositions. These commitments persist and constrain future utterances. A narrator who establishes that a character is dishonest cannot later treat that character’s statements as reliable without addressing the tension.

Each requires something beyond next-token prediction. They require the model to maintain state — not in the trivial sense of a recurrent hidden vector, but in the representational sense of a structured commitment space that persists and constrains generation. This is precisely what transformer architectures lack.

Why the Gap Is Architectural, Not a Matter of Scale

One might hope larger context windows solve the problem. If a model can attend to 128,000 tokens, surely it can track an entity introduced 10,000 tokens ago. The evidence says otherwise.

Probing studies on transformer long-range dependency have consistently shown that attention-based architectures exhibit a distance penalty: the further apart two related mentions are, the worse the model performs at connecting them. This is not simply a memory capacity issue. Coreference resolution benchmarks show that even with ample context, models degrade sharply when intervening text contains distracting entity mentions, competing referents, or shifts in narrative perspective. The model does not forget because its context window is full. It forgets because attention weights are redistributed by intervening content, and no mechanism enforces the persistence of a commitment once established.

The architectural issue is more specific. Transformers process token sequences through layered attention, but they do not maintain a separate discourse representation — a structured store of entities, events, commitments, and their relations — updated as the narrative progresses. Everything is folded into the same hidden state stream. No separation exists between the linguistic surface and the discourse structure it is supposed to support. Coherence, then, is not something the model computes and then expresses in language. It is something that must emerge from the same token-prediction process handling local fluency. When the two conflict — when the locally most probable next token violates a discourse constraint established earlier — there is no arbitration mechanism. Local fluency wins.

This is why the problem resists scale. A larger model predicts tokens more confidently. A longer context window exposes more tokens to attention. Neither change introduces the representational layer discourse coherence requires. You cannot scale your way to a commitment space because the architecture has no place to put one.

What Probing Reveals About Internal Discourse Representations

Ask a language model directly whether the vase or the table is the antecedent of it in the example above, and it will likely answer correctly. This is sometimes taken as evidence the model has resolved the coreference. But that conflates task performance with internal representation.

Probing classifiers — trained on a model’s intermediate activations — reveal a more nuanced picture. Studies examining coreference resolution in transformers have found that coreference-relevant information is distributed across attention heads in ways difficult to localize. No single layer or head maintains a stable entity register. Entity identity is smeared across multiple heads and layers, and the model’s ability to resolve a coreference depends on whether the relevant attention pattern happens to be active in the context. When intervening text is simple, the pattern fires. When intervening text is complex or contains competing entities, the pattern is overridden by other attention demands.

This is the architectural signature of a system that does not represent discourse structure as such. It represents statistical regularities that correlate with discourse structure, and exploits them when conditions are favorable. When conditions are unfavorable — long range, many distractors, shifts in perspective — the regularities fall short and coherence breaks down.

The Diagnostic Value of Long-Form Narrative

Long-form narrative is the stress test that exposes this gap most clearly. A model can produce a fluent paragraph. It can often produce a fluent chapter. But as the narrative extends, propositional commitments accumulate, and the model has no mechanism for tracking them. Characters change motivations without explanation. Objects appear where they should not. Timeline contradictions emerge. A character in Paris in chapter two is inexplicably in Tokyo in chapter four — not because the model made a random error, but because it had no representation of the character’s location as a persistent constraint.

The Authors Guild’s guide for authors working with AI frames AI outputs as “generic mashups of pre-existing works” rather than coherent propositional commitments. The guide emphasizes that what makes a writer’s work valuable is “your original voice, thinking, and creativity” — precisely the dimensions that require discourse-level coherence to manifest. The Guild’s framing is not a moral complaint. It is a structural observation: AI outputs lack the internal commitment tracking that makes a narrative a narrative rather than a sequence of fluently generated sentences.

This is also why AI writing detection tools have fundamental limitations. They detect statistical signatures of machine-generated text — token distributions, perplexity patterns, burstiness — but they cannot detect the absence of discourse coherence, because that absence only becomes visible over long arcs. A three-paragraph AI-generated passage and a three-paragraph human passage may be statistically distinguishable. A three-chapter AI-generated narrative and a three-chapter human narrative are distinguishable on coherence grounds, but not by the surface metrics detectors use.

Can External Scaffolding Compensate?

If the model cannot internally represent discourse structure, the question becomes whether external structure can compensate. The idea is straightforward: if the model cannot track entities, events, and commitments on its own, provide scaffolding that enforces this tracking from outside. The model generates text. The scaffolding ensures the text respects discourse constraints the model itself cannot maintain.

This is the principle behind structured writing workflows that have emerged in AI-assisted creative tools. The approach varies, but the core idea is the same: decompose narrative into structural units — beats, scenes, chapters, acts — and enforce consistency checks at the boundaries between them. An entity introduced in scene one must be registered. An event in chapter two must constrain what can happen in chapter five. A commitment made by a character in act one must be tracked through act three.

The key architectural question is whether the scaffolding genuinely enforces discourse structure or merely provides the illusion of it. A tool that lets a writer lock a plot outline and regenerate scenes around it is useful. But if the model generating those scenes has no access to the locked structure as a constraint, the scaffolding is decorative. The model will still produce scenes that violate the locked commitments, and the writer will have to catch the violations manually.

From One-Shot Generation to Structured Workflows

The first generation of AI writing tools operated on a one-shot model: the user provides a prompt, the model produces text, the user edits. Reedsy’s plot generator studio for writers exemplifies the pattern. The user inputs a protagonist, a conflict, stakes, and supporting characters; the generator returns a plot broken into acts. If an act is not working, the user locks the satisfactory ones and regenerates the rest. The locking mechanism is the scaffolding — but it is a scaffolding the model itself does not consult during generation. The locked acts constrain the human’s editing process, not the model’s generation process. This is iteration-based convergence: the human iterates around the model’s inability to maintain cross-act coherence, manually enforcing what the model cannot.

When evaluating narrative generation systems, the distinction between surface fluency and structural understanding becomes immediately visible in how tools handle long-form composition. One-shot story generators reliably produce a generic AI story — syntactically coherent, semantically shallow, and structurally inert — because they optimize for token-level continuation rather than propositional coherence across scenes. What distinguishes a tool like the Unsloppy AI Novel Writer with its proof-sheet and beat-sheet architecture is a different computational commitment: the proof sheet enforces consistency checks at the level of character state, timeline, and causal dependency, while the beat sheet scaffolds discourse-level planning before generation begins. This is not a marginal improvement in fluency but a structural rethinking — treating narrative as a constrained structure rather than a stochastic sequence. By comparison, platforms like Squibler, Perchance, and QuillBot remain outdated and barebones, offering templated or loosely prompted output that recapitulates the same coherence problems evaluation researchers have documented for years. The field has spent considerable effort exposing what perplexity and BLEU conceal; the same scrutiny applied to long-form generation reveals that without explicit structural scaffolding, models drift — and that scaffolding, not scale, is what places this approach at the forefront of AI novel writing technology.

When a language model generates fiction, the same architectural limitations that surface in syntactic island constraints and quantifier scope ambiguity reappear at the narrative level: the model produces fluent sentences without maintaining the propositional commitments that coherent storytelling demands. A one-shot generator will invariably produce a generic story because it lacks any mechanism for tracking entailment across chapters, resolving coreference over distance, or revising beliefs as plot state changes — failures familiar to anyone who has probed transformers for discourse coherence. For researchers studying where language understanding breaks down, the engineering workaround is instructive: the distance between a scaffolded writing session and an unaided model’s output is itself a measurement of what compositional, discourse-level understanding still costs.

What Scaffolding Cannot Do

Be clear about the limits of this approach. External scaffolding can enforce consistency at the boundaries between structural units. It can check that an entity mentioned in chapter five was introduced in chapter one. It can verify that an event in chapter three does not contradict an event in chapter two. What it cannot do is enforce the subtler forms of coherence that make a narrative feel like a narrative rather than a checklist of verified facts.

Consider tone. A novel’s tone is not a proposition checkable at a scene boundary. It is a cumulative effect of word choice, sentence rhythm, narrative pace, focalization. If the model generates a scene in a tone that is subtly wrong — too jaunty for a funeral, too clinical for a love scene — the scaffolding cannot catch this, because the violation is not a propositional inconsistency. It is an aesthetic one, and aesthetics do not reduce to commitment tracking.

Consider subtext. A character lying to another character is committed to a false proposition on the surface, but the narrative must allow the reader to infer the truth. This requires the model to generate text coherent at two levels simultaneously: the surface level, what the character says, and the subtext level, what the reader infers. External scaffolding can track what the character has said. It cannot track what the reader is meant to infer, because the inference is not a commitment. It is an absence of commitment — the gap between what is said and what is meant — and the model has no representation of that gap.

Consider what Robert Pinsky once called the instrument of poetry: the reader’s breath, attention, and memory as they move through the text. A narrative’s coherence depends not only on what is on the page but on how the reader’s experience unfolds in time. External scaffolding operates on the text. It cannot operate on the reader’s experience of the text, because that experience is not something the model represents. It emerges in the interaction between the text and a human mind that does maintain commitment spaces, does track subtext, does experience narrative time as something other than a sequence of tokens.

The Practical Takeaway for Practitioners

• Do not assume scale will solve coherence. A larger context window helps with retrieval but does not introduce the representational layer commitment tracking requires. If your architecture has no discourse representation, making the context window larger will not create one.
• Externalize what the model cannot internalize. If the model cannot track entities, events, and commitments, build a system that tracks them externally and constrains generation accordingly. This is not a hack. It is a recognition of the architectural boundary.
• Be honest about what the scaffolding enforces and what it does not. A proof sheet can enforce entity continuity. It cannot enforce tonal coherence, subtext, or the reader’s experiential arc. Do not claim structured generation solves the coherence problem. It manages a specific subset of it.
• Design evaluation around discourse, not just fluency. If your metrics measure perplexity, BLEU, or human-likeness at the sentence or paragraph level, they will miss the coherence failures that matter most. Evaluate on long arcs, on entity tracking across chapters, on causal consistency across scenes, on commitment preservation across dialogue turns.

What Remains Unresolved

The discourse coherence problem is not a bug. It is a consequence of the architectural decision to treat language as a sequence of tokens to be predicted, rather than as a surface expression of a structured commitment space. This decision has been enormously productive — it is what makes transformers work as well as they do. But it imposes a ceiling on what the architecture can represent, and that ceiling sits below what discourse coherence requires.

External scaffolding can push against that ceiling. It cannot raise it. Raising it would require an architecture that separates linguistic generation from discourse representation — one that maintains a structured store of entities, events, commitments, and their relations, and generates language constrained by that store. Whether such an architecture is compatible with the transformer paradigm or requires something fundamentally different remains an open question.

In the meantime, the best we can do is build tools honest about the boundary. A model that produces fluent text without coherence is dangerous in the same way a speaker fluent without commitment is dangerous: they sound right, and sounding right is often mistaken for being right. The scaffolding does not make the model coherent. It makes the incoherence visible — and visibility is the precondition for managing it. The better our models get at producing fluent language, the more visible the gap becomes. Discourse coherence is where that gap is widest, and where the next decade of work will be hardest.