Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe

• URL: http://arxiv.org/abs/2607.25292v1
• Date: Tue, 28 Jul 2026 04:58:46 GMT
• Title: Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
• Abstract summary: Silicon sampling uses language models as proxies for human survey respondents.<n>We show this draw does not exist: instruction-tuned models do not sample from distributions, they collapse to a single output.<n>We propose Prompt-Perturbed Argyle, which reduces the same error by 21% at no added cost.
• Score: 2.6646205080230536
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Silicon sampling uses language models as proxies for human survey respondents, treating each model call as an independent draw from the persona's response distribution. We show this draw does not exist: instruction-tuned models do not sample from distributions, they collapse to a single output. The same persona on the same question returns the same answer on more than half of items in a public-opinion benchmark. The collapse is sharp: the model's internal probabilities concentrate on a single option, and the failure is substantially amplified by instruction tuning: across three model families with materially different post-training pipelines, every instruction-tuned model fails on every task we test, while base models fail far less often. Strikingly, the same model that cannot sample from a distribution can describe it accurately in a single call. We call this gap the KNOWS/DOES split, and trace it to a degenerate sampling primitive visible in the logits and induced by alignment training. Exploiting this split, asking the model to describe the response distribution in one call more than halves the error against human survey data compared to persona aggregation. For applications that require per-persona outputs, we propose Prompt-Perturbed Argyle (PPA), which reduces the same error by 21% at no added cost.

Related papers

• UnpredictaBench: A Benchmark for Evaluating Distributional Randomness in LLMs [21.05820098186702]
  We introduce UnpredictaBench, an evaluation that tests the ability of large language models to capture true underlying distributions.<n>We find that even simple distributional simulation remains challenging, making it a necessary first step toward using LLMs as stand-ins for complex systems.
  arXiv  Detail & Related papers  (2026-06-04T18:18:02Z)
• Large Language Models Are Bad Dice Players: LLMs Struggle to Generate Random Numbers from Statistical Distributions [50.1404916337174]
  We present the first large-scale, statistically powered audit of native probabilistic sampling in large language models (LLMs)<n>We show that batch generation achieves only modest statistical validity, with a 13% median pass rate, while independent requests collapse almost entirely.<n>We conclude that current LLMs lack a functional internal sampler, necessitating the use of external tools for applications requiring statistical guarantees.
  arXiv  Detail & Related papers  (2026-01-08T22:33:12Z)
• Rate of Model Collapse in Recursive Training [13.722324504719282]
  We ask how fast model collapse occurs for some well-studied distribution families under maximum likelihood (ML or near ML) estimation.<n>Surprisingly, even for fundamental distributions such as discrete and Gaussian distributions, the exact rate of model collapse is unknown.<n>Our results show that for discrete distributions, the time to forget a word is approximately linearly dependent on the number of times it occurred in the original corpus.
  arXiv  Detail & Related papers  (2024-12-23T15:21:50Z)
• Model Equality Testing: Which Model Is This API Serving? [59.005869726179455]
  API providers may quantize, watermark, or finetune the underlying model, changing the output distribution.<n>We formalize detecting such distortions by Model Equality Testing, a two-sample testing problem.<n>A test built on a simple string kernel achieves a median of 77.4% power against a range of distortions.
  arXiv  Detail & Related papers  (2024-10-26T18:34:53Z)
• Large Language Monkeys: Scaling Inference Compute with Repeated Sampling [81.34900892130929]
  We explore inference compute as another axis for scaling, using the simple technique of repeatedly sampling candidate solutions from a model.<n>Across multiple tasks and models, we observe that coverage scales with the number of samples over four orders of magnitude.<n>In domains like coding and formal proofs, where answers can be automatically verified, these increases in coverage directly translate into improved performance.
  arXiv  Detail & Related papers  (2024-07-31T17:57:25Z)
• Ablation Based Counterfactuals [7.481286710933861]
  Ablation Based Counterfactuals (ABC) is a method of performing counterfactual analysis that relies on model ablation rather than model retraining. We demonstrate how we can construct a model like this using an ensemble of diffusion models. We then use this model to study the limits of training data attribution by enumerating full counterfactual landscapes.
  arXiv  Detail & Related papers  (2024-06-12T06:22:51Z)
• IBADR: an Iterative Bias-Aware Dataset Refinement Framework for Debiasing NLU models [52.03761198830643]
  We propose IBADR, an Iterative Bias-Aware dataset Refinement framework. We first train a shallow model to quantify the bias degree of samples in the pool. Then, we pair each sample with a bias indicator representing its bias degree, and use these extended samples to train a sample generator. In this way, this generator can effectively learn the correspondence relationship between bias indicators and samples.
  arXiv  Detail & Related papers  (2023-11-01T04:50:38Z)
• Contextual Dropout: An Efficient Sample-Dependent Dropout Module [60.63525456640462]
  Dropout has been demonstrated as a simple and effective module to regularize the training process of deep neural networks. We propose contextual dropout with an efficient structural design as a simple and scalable sample-dependent dropout module. Our experimental results show that the proposed method outperforms baseline methods in terms of both accuracy and quality of uncertainty estimation.
  arXiv  Detail & Related papers  (2021-03-06T19:30:32Z)
• One for More: Selecting Generalizable Samples for Generalizable ReID Model [92.40951770273972]
  This paper proposes a one-for-more training objective that takes the generalization ability of selected samples as a loss function. Our proposed one-for-more based sampler can be seamlessly integrated into the ReID training framework.
  arXiv  Detail & Related papers  (2020-12-10T06:37:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.