Study: LLM output diversity isn’t a design flaw, it’s a data gap

3 days ago 28

Anyone who has spent time chatting with an AI assistant has probably noticed it: ask the same question a few different ways, and you often get eerily similar answers back. That repetition has a name in AI research circles — mode collapse — and a new study looks at why it happens and, more importantly, what can be done about it. The research, authored by Kirill Skobelev and published in September 2026, digs into LLM output diversity and finds that the story is more nuanced than the field’s recent warnings suggested.

Key takeaways

Mode collapse happens when large language models produce outputs that repeat or resemble one another more than the real population of responses they’re meant to reflect.
Diversity problems can go either way — models can be under-diverse or over-dispersed, depending on the specific model and dataset.
Enough supervised fine-tuning data pulls a model’s output diversity toward the target distribution it was trained on.
Fine-tuning itself carries no built-in bias toward collapsing diversity or inflating it — the direction depends on the data available.
The gap between a model’s diversity and the real target is mathematically capped by the square root of the KL divergence between the two.

Mode Collapse and Diversity Issues in Large Language Models

Mode collapse describes a simple but consequential problem: when an LLM is asked the same type of question repeatedly, its answers start to look alike far more than they should. Prior work by Doshi and Hauser (2024), Bisbee et al. (2024), and Xie et al. (2026) flagged this as a recurring concern, arguing that model outputs tend to be under-diverse compared with the population of responses they’re supposed to represent.

Skobelev’s paper takes that concern seriously but complicates it. Rather than treating mode collapse as a fixed flaw baked into how LLMs work, the study argues that whether a model collapses toward sameness — or swings the other way into excessive spread — depends entirely on the specific model and the dataset it was trained on. There’s no universal verdict here; the outcome shifts case by case.

Definition and Metrics of Mode Collapse

To make this measurable rather than anecdotal, the research relies on a concrete metric: the probability that two responses, sampled independently from the same fixed prompt, end up matching each other. Researchers call this a “collision” probability, and it can also be expressed as the expected similarity between two outputs under a kernel function. This gives a numerical handle on something that’s usually described only in vague terms like “repetitive” or “samey.”

Variability of Diversity Across Models and Datasets

What emerges from this measurement approach is that under-dispersion isn’t the only failure mode worth watching. Some models and datasets push output diversity too low, echoing the mode-collapse worry raised in earlier research. Others do the opposite, scattering responses more widely than the real-world distribution warrants. Both are forms of miscalibration, and both matter for anyone relying on an LLM to reflect a realistic range of human-like answers — whether that’s for survey simulation, creative writing, or code generation.

Impact of Supervised Fine-Tuning on Output Diversity

Fine-tuning turns out to be the lever that determines which way diversity tips — and, encouragingly, it can correct itself with enough data. The paper’s central claim is that with sufficient supervised fine-tuning data, a model’s output diversity converges toward the diversity of the target distribution the fine-tuning data was drawn from. In plain terms: feed the model enough representative examples, and its spread of answers starts matching reality rather than drifting toward either extreme.

Effect of Fine-Tuning Data Volume

This isn’t just a theoretical nicety. Across every experiment the researchers ran, adding more target data during fine-tuning consistently moved model diversity closer to the human or synthetic benchmark it was meant to imitate. That pattern held whether the underlying task involved synthetic languages, human survey responses, or programming problems, suggesting the effect isn’t tied to one narrow use case.

Lack of Inherent Bias Toward Mode Collapse

Perhaps the most counterintuitive finding is that supervised fine-tuning carries no inherent lean toward collapsing diversity or inflating it. Finite-sample fine-tuning — meaning fine-tuning done with a limited, real-world amount of data rather than an infinite ideal dataset — can leave a model either under-dispersed or over-dispersed, and which one happens depends on the model and dataset involved rather than on some structural flaw in the fine-tuning process itself. This matters for the ongoing debate about mode collapse in language models: it reframes the issue from “fine-tuning causes collapse” to “fine-tuning’s outcome depends on how much and what kind of data you give it.”

Theoretical Framework and Validation

To explain why diversity drifts one way or the other, the study builds a mathematical framework rather than relying purely on observation — and then tests that framework against real experiments across very different domains.

Bias-Variance Decomposition of Diversity Gap

The researchers derive what they call a bias-variance decomposition of the expected gap between a model’s collision probability and the target’s collision probability. This is the mechanism behind the “no inherent bias” finding: it separates the systematic error from the random, sample-driven error, showing mathematically why finite data can push things in either direction depending on the draw.

Mathematical Bound via KL Divergence

The paper also establishes a hard limit on how wrong diversity calibration can get. The absolute gap between model and target diversity is bounded by the square root of the Kullback-Leibler divergence — a standard measure of how different one probability distribution is from another — running from the target distribution to the model. The practical upshot: a model that’s already close to optimal under standard cross-entropy training simply cannot have wildly miscalibrated diversity. Good language modeling, in other words, puts a ceiling on how bad the diversity problem can be.

Empirical Validation Across Diverse Experiments

None of this stays purely theoretical. The team tested both the decomposition and the KL-based bound across three distinct experimental settings: small transformer models trained on synthetic languages, four LLMs fine-tuned on human survey data, and the same four LLMs fine-tuned on CodeNet, a dataset built from human-written code solutions. This spread — from artificial languages to human opinion data to real programming problems — gives the findings a broader footing than a single narrow test case would.

Across all three settings, the pattern held: more target data during fine-tuning moved model diversity closer to the human or synthetic benchmark, exactly as the theory predicted. That consistency is what turns this from an interesting mathematical curiosity into something with practical weight for anyone building or fine-tuning language models.

Why This Matters for AI Development

The practical implication is straightforward: diversity miscalibration in LLMs isn’t a permanent design flaw — it’s largely a symptom of finite-sample error that shrinks as fine-tuning better approximates the real target distribution. That reframes model calibration in AI as a data problem more than an architectural one. Teams worried about repetitive or overly erratic model outputs may find that the fix lies less in redesigning the model and more in expanding or refining the fine-tuning dataset itself.

This also tempers some of the alarm raised in earlier research about LLMs being categorically under-diverse. Skobelev’s findings suggest the reality is more case-by-case: some models collapse toward sameness, others scatter too widely, and the direction depends on the specific combination of model and dataset — not on some universal tendency baked into how large language models work.