Can language models learn to diversify their discourse-level narrative patterns over time?
This explores whether LLMs can grow more varied in their higher-level storytelling structure — character agency, chronology, plot architecture — rather than just surface wording, and whether anything in their training lets that variety improve with time.
This explores whether language models can become more varied at the level of narrative *structure* — not word choice, but the deeper choices that shape a story: who has agency, how events are ordered, how a plot is built. The corpus is fairly blunt here: the forces pushing models toward sameness at this level are strong, and the mechanisms that would let them diversify over time are mostly absent.
Start with where the uniformity lives. Work on AI fiction detection found that machine-written stories are separable from human ones at 93% accuracy using *only* discourse-level features like character agency and chronological structure — and crucially, that signal survives even when you strip out stylistic tells Can AI stories be detected without analyzing writing style?. The reason it resists 'humanization' is the same reason it's hard to diversify: these are structural commitments baked deep into generation, not surface edits you can sprinkle on. So the very thing the question asks about turns out to be the thing that's most stubbornly fixed.
Why so fixed? Two converging pressures. First, the 'Artificial Hivemind' effect: across 70+ models and 26K open-ended prompts, different models independently converge on strikingly similar outputs because they share overlapping training data and alignment procedures Do different AI models actually produce diverse outputs?. Diversity loss isn't a bug of one model — it's a basin the whole field falls into. Second, alignment training itself locks in a single communicative identity: RLHF and system prompts impose a static persona that can't switch register or trade off values the way human pragmatics demands Can language models adapt communication style to different contexts?. If a model can't even shift conversational register by context, expecting it to autonomously broaden its narrative repertoire is a tall order.
There's a deeper wrinkle worth knowing. A model isn't really 'committed' to one narrative voice in the first place — the 20-questions regeneration test shows LLMs hold a superposition of possible characters and *sample* one at generation time, so regenerating yields a different-but-consistent output each time Do large language models actually commit to a single character?. That means the raw variety is latent and available; what collapses it is the sampling distribution that training and alignment have sculpted. Diversification, then, isn't about teaching new patterns — it's about whether that distribution can be reshaped. And here the news is hard: when in-context cues conflict with strong training-time associations, the parametric priors win, and prompting alone can't override them Why do language models ignore information in their context?.
The 'over time' half of the question is where the corpus offers the one genuine opening. Models don't diversify within their fixed weights, but agents *can* learn across episodes without any weight update — Reflexion stores verbal self-critiques in episodic memory and improves trial by trial, as long as the feedback signal is unambiguous enough to prevent rationalizing Can agents learn from failure without updating their weights?. That hints at a path: narrative diversification might be achievable not by retraining but by an outer loop that remembers and reflects on its own prior structural choices. But notice the catch — Reflexion works because success/failure is binary and clear, and narrative variety has no such crisp signal. The thing that makes story structure resistant to detection-evasion is the same thing that makes 'did I diversify?' hard to score. So the honest answer the corpus points to is: the latent variety exists, the training pressures bury it, and learning-over-time is possible only where feedback is sharp — which, for discourse-level narrative, it currently isn't.
Sources 6 notes
StoryScope achieved 93.2% accuracy separating AI from human fiction using only discourse-level features like character agency and chronological structure, retaining 97% of performance while eliminating stylistic cues. These structural choices resist humanization because they require rewrites, not surface edits.
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.