4 min read
When you ask a language model to generate ten different solutions to a problem, you often get ten variations on the same answer. This phenomenon, commonly called diversity collapse, is well-documented but poorly understood at a mechanistic level. Most prior work has focused on measuring the problem or patching it with better sampling heuristics. A new paper, Sampling More, Getting Less: Calibration is the Diversity Bottleneck in LLMs (Banayeeanzade et al., 2026), takes a different approach: tracing diversity collapse back to the token-level probability distributions that LLMs produce during decoding, and showing that the root cause is miscalibration rather than any particular sampling strategy.
This matters because if diversity collapse were simply a product of greedy-ish sampling procedures like top-k or nucleus sampling, the fix would be straightforward: use better heuristics or higher temperature. The paper's central argument is that the problem runs deeper. Even with oracle-level sampling, the underlying distributions are structured in ways that make diverse, valid output statistically difficult to recover.
The core contribution is a formal framework that decomposes the diversity problem into two distinct failure modes, both framed as calibration failures in the probability mass the model assigns across valid and invalid continuations at each decoding step.
The first is order calibration: the degree to which valid tokens are ranked above invalid tokens in the model's probability distribution. If valid and invalid tokens are interleaved in the ranking, any rank-based cutoff rule (top-k, nucleus sampling, etc.) faces an unavoidable tradeoff. Cutting off too aggressively excludes valid continuations; cutting too permissively admits invalid ones. Neither option is good for diversity.
The second is shape calibration: even when valid tokens are correctly ranked above invalid ones, the probability mass may be highly concentrated on just a few of them, with the remaining valid tokens receiving negligible weight. This produces a long tail of mixed valid and invalid tokens that is practically inaccessible to sampling without also pulling in a lot of noise. High validity and high diversity become mutually exclusive under this distribution shape.
What makes the framework analytically interesting is the compounding argument. Local miscalibration at individual decoding steps does not stay local. Because sequences are generated autoregressively, small per-step failures in order or shape calibration multiply across the length of a generation, producing strong sequence-level diversity losses even when any single step looks only mildly miscalibrated. This is a cleaner mechanistic account than most prior work offers.
To test these ideas empirically, the authors construct diagnostic tasks where the set of valid continuations is exactly known. This is a smart design choice. Most naturalistic NLP tasks have fuzzy validity boundaries, which makes it impossible to cleanly separate calibration failures from genuine ambiguity. By working with controlled settings where ground truth is available, they can define oracle cutoff baselines and measure how far real model distributions fall from them.
The evaluation spans 14 language models across multiple families and scales. This breadth is important: it means the findings are not artifacts of a single model architecture or training recipe. The consistent pattern across model families strengthens the claim that miscalibration is a structural property of how current LLMs are trained, not an idiosyncratic quirk.
Key empirical findings include:
The practical implication is somewhat uncomfortable for the field. A large body of work on improving LLM output diversity has focused on sampling strategies: temperature scaling, top-p, top-k, minimum-p, and various combinations thereof. This paper suggests that all of these approaches are fighting a losing battle against the underlying distribution. If the model's probability mass is concentrated on a narrow set of valid tokens, no sampling procedure operating on that distribution can recover diversity without also degrading validity.
This shifts the research question from "how should we sample?" to "how should we train?". If calibration is the bottleneck, the fix presumably lies in training objectives or post-training procedures that explicitly encourage better-calibrated distributions over valid continuations. The paper does not propose such a fix, which is a reasonable scope decision for a diagnostic paper, but it does frame the problem in a way that points clearly toward where solutions should be sought.
There is also an implication for evaluation. Benchmarks that measure diversity at the output level, without controlling for validity, may conflate two separate problems. A model that generates highly diverse but frequently invalid outputs looks diverse by surface metrics but is not actually solving the calibration problem the authors identify. The validity-diversity framework offers a cleaner decomposition for future benchmark design.
The controlled diagnostic tasks that make the empirical analysis clean also limit its immediate generalisability. Real-world generation tasks rarely have exactly known valid sets, and it is not obvious how order and shape calibration should be operationalised when validity is fuzzy or domain-dependent. Extending the framework to naturalistic tasks will require careful thought about how to define and measure validity without ground truth.
The paper also leaves open the question of why miscalibration arises. The training dynamics that produce concentrated probability mass on a narrow subset of valid continuations are not analysed here. Understanding whether this is a consequence of RLHF, instruction tuning, the base pre-training objective, or some interaction among them would be valuable for designing targeted interventions.
Finally, the compounding argument, while theoretically compelling, is demonstrated empirically rather than derived with tight bounds. A more formal characterisation of how per-step miscalibration accumulates across sequence length would strengthen the theoretical contribution.
These are reasonable limitations for a paper that is primarily making a diagnostic contribution. The validity-diversity framework is a genuine conceptual advance, and the empirical breadth across 14 models gives the core findings credibility. For anyone working on creative generation, scientific hypothesis generation, or any application where output diversity matters, this paper provides a clearer picture of where the problem actually lives. The full paper is available at arXiv:2605.11128.