Recent research published on arXiv (2607.18476v1) reveals a phenomenon that shakes the foundations of how we understand language models: when asked to respond in a structured format, like JSON, the diversity of their answers collapses. In an experiment replicating the 'One-Word Census' across 44 models, it was observed that the simple instruction 'Reply with JSON only' is enough to make responses much more homogeneous. The modal answer rises from 41% to 64% of the pool, distinct answers fall from 52 to 36, and mean answer-choice surprisal drops from 1.80 to 1.58 bits. This effect is not marginal: six models individually shift toward the mode, while the most conformist models remain immobile. It is a sharpener, not a re-indexer: the plain-chat modal answer survives in 28 out of 31 categories.
The study further reveals that the collapse does not come from the decoder that enforces a schema (like response_format), but from the model's response to the linguistic register itself. That is, models learn to behave differently depending on the output format they are asked to use. When requested JSON, their behavior becomes more predictable and less diverse, as if they had been trained to speak a more standardized dialect. This has immense practical implications for companies integrating language models into their systems: the software consumption surface (structured output) is significantly more homogeneous than the chat surface where models are evaluated and compared.
For organizations developing custom software with artificial intelligence, this finding means they must carefully consider which output format they use. If they seek creative or diverse responses, JSON may be suppressing exactly the variety they need. Conversely, if consistency and standardization are desired, JSON is the best ally. At Q2BSTUDIO, as a company specialized in software development and technology, we know that choosing the response format is a crucial technical decision that affects the performance of AI systems. That is why, when designing AI solutions for our clients, we evaluate the usage context: from conversational agents to virtual assistants that require structured responses to feed automated processes.
The research also shows a register gradient: compression is significant and specific to the answer-delivery formats models are trained to speak (JSON -0.22 bits, XML -0.19 bits), absent for YAML and CSV, and reversed for an arbitrary bracket wrapper (+0.13 bits). This suggests the phenomenon is linked to tool-use post-training. When a company integrates language models into its cloud infrastructure, whether on cloud AWS/Azure, it must consider that structured output is not only a technical requirement but also modifies model behavior. If diversity is needed for idea generation or brainstorming, perhaps it is better to use a freer format and then structure the response later.
Another relevant aspect is security. In the field of cybersecurity, where language models are often used to detect anomalous patterns or generate reports, the homogeneity induced by JSON could reduce the ability to detect outliers or unexpected responses. A model that always responds similarly might miss rare warning signs. Therefore, at Q2BSTUDIO we design systems that combine artificial intelligence with specific business rules, ensuring that the model's natural diversity is not lost when it is needed.
The research also has implications for Business Intelligence tools. When using models to generate SQL queries or reports in BI/Power BI, structured output is essential. However, the study shows that enforcing a schema (response_format) does not compress further than the simple request (-0.03 bits), indicating that the collapse lies in the model's response to the register, not the decoder. This suggests that companies can obtain consistent responses simply by asking for JSON, without needing to impose rigid schemas, which simplifies integration.
In the context of AI agents, where the ability to choose among multiple valid answers is crucial, this phenomenon can be a double-edged sword. On one hand, homogeneity can make agents take more predictable and safer decisions. On the other hand, it can limit their ability to explore novel options. At Q2BSTUDIO we develop AI agents that balance exploration and exploitation, using techniques like temperature sampling and prompt diversification to counteract the homogenizing effect of JSON when necessary.
The original arXiv article highlights that the effect is progressive: the most distinctive models move toward the mode, while the conformist floor remains immobile. This recalls the 'diversity paradox' in machine learning systems: the more interfaces are standardized, the more model richness is lost. For companies seeking custom software with AI, the lesson is clear: not all output formats are equivalent. The choice of format must be made with full awareness, considering both technical requirements and the impact on model behavior.
In conclusion, the research on diversity collapse with structured output is a reminder that language models are sensitive to the interaction context. The software consumption surface is not neutral; it shapes model responses. For companies integrating AI into their processes, understanding this phenomenon is key to designing robust, versatile systems tailored to their needs. At Q2BSTUDIO, as a technology partner, we help our clients navigate these complexities, offering solutions ranging from cloud implementation to creating intelligent agents, always with a business-value-focused approach.
