DeepMind Study Reveals LLMs Fail at Reliable Random Sampling

Ask a large language model to pick a random number, and it will likely give you an answer. However, that answer is rarely truly random. New research from Google DeepMind suggests that the stochasticity we expect from these models is often an illusion.

The study, titled The Illusion of Stochasticity in LLMs, argues that reliable random sampling is a fundamental requirement for the next generation of AI agents. Without it, agents cannot navigate environments or make decisions that require genuine unpredictability. Current models, the researchers conclude, simply do not meet this baseline.

Collaborating with the National University of Singapore, the research team included experts such as Petar Veličković and Razvan Pascanu. Their findings, presented at the COLM 2026 conference, expose a significant gap between a model's mathematical understanding and its operational execution.

Statistical Failures in Sampling

The core issue is a disconnect between probabilistic reasoning and output generation. While LLMs can estimate probabilities and handle complex math, they struggle to produce outputs that actually follow a specific distribution. To test this, researchers asked models to sample from standard uniform and Gaussian distributions.

To verify the results, the team employed Chi-Square and Kolmogorov-Smirnov tests. These are standard statistical benchmarks used to determine if a dataset matches a theoretical pattern. The results were definitive: the models produced near-zero p-values. This indicates that the probability of the outputs following the target distribution was virtually non-existent.

According to cryptobriefing.com, the biases were not subtle. Instead of a smooth distribution, models gravitated toward specific, predictable numbers like 7 or 42. The outputs were also sensitive to the order in which they were requested, further breaking the requirements for true randomness.

Workarounds and Agentic Implications

The research does offer a silver lining for developers. While the models fail to roll their own dice, they can sample accurately if they are provided with random numbers generated by external code. This suggests that the limitation is not in the model's ability to process randomness, but in its ability to generate it internally.

This distinction is vital for the development of agentic systems. As we move toward more complex workflows, such as those supported by evertune.ai or the multi-agent systems targeted by Nvidia's Nemotron 3, the ability to call external tools becomes a necessity rather than a luxury. An agent that cannot sample reliably is an agent that cannot be trusted in any environment requiring stochastic decision-making.

This failure highlights a broader trend in artificial intelligence analysis. We are seeing a shift from simply increasing parameter counts to refining the reliability of model behavior. As models like the recently discussed GPT-6.1 Terra aim for better computer use and visual reasoning, the underlying mechanics of how they handle uncertainty must be addressed.

Contextualizing the Stochastic Gap

Historically, we have treated LLMs as black boxes of probability. We assumed that because they are trained on probabilistic next-token prediction, they inherently understand randomness. This study proves that mathematical competence does not equal operational stochasticity. A model can know the formula for a Gaussian distribution while being fundamentally incapable of executing it.

For practitioners, this means that any application requiring true randomness—from Monte Carlo simulations to game theory applications—cannot rely on the raw text output of an LLM. Developers must implement a tool-use architecture where the model requests a seed or a random float from a verified Python environment. Relying on the model to pick a number is not just a minor error; it is a systematic failure of the model's logic.

As the industry tracks hundreds of new releases via platforms like ai-tldr.dev, the focus often remains on benchmarks like ARC-AGI or reasoning capabilities. However, this research suggests that the plumbing of AI—the basic ability to handle randomness—is a critical bottleneck that could undermine the reliability of even the most advanced agentic frameworks.