The design of information disclosure policies has emerged as a critical architectural variable for developers building multi-agent systems, as a new study indicates that the precision of data provided to large language models (LLMs) does not lead to linear improvements in performance. A preprint paper published on November 5, 2026, titled "Information Disclosure in LLM-Based Markets," systematically examines how the granularity of input data—ranging from coarse signals to full disclosure—affects LLM decision-making in high-stakes environments such as multi-resource trading, bilateral bargaining, and double auctions. The findings suggest that there is no "one-size-fits-all" configuration for information policy; instead, developers must account for a complex interplay between model architecture, market complexity, and communication constraints.
The Experimental Framework: Testing Intelligence in Markets
To isolate the impact of information disclosure, researchers tested ten distinct LLMs across three controlled economic environments. The study kept models, incentives, and market rules constant while strictly varying the level of input data provided to the agents.
In the multi-resource trade environment, researchers utilized three levels of disclosure: "least disclosure," where agents are provided only with their own resource values; "coarse disclosure," which includes the counterparty’s ordinal ranking; and "full disclosure," which grants agents access to the exact valuation of all parties. Similar frameworks were applied to bilateral bargaining and double auctions, ensuring a rigorous comparative baseline.
The research was conducted against a backdrop of increasing interest in autonomous agent deployment within financial and logistical systems. As developers shift from building single-purpose chatbots to complex agentic workflows capable of negotiation and resource allocation, understanding the "information bottleneck" has become a priority. The experiment highlights that adding more information can often overwhelm a model’s reasoning capabilities, leading to counter-intuitive performance drops.
Discrepancies in Model Performance: A Comparative Analysis
The data reveals that the relationship between transparency and market efficiency is highly non-uniform. For instance, in a two-resource trade setting, the performance of GPT-6 Sol declined significantly as disclosure increased. Realized gains for this model dropped from 97.7% under coarse disclosure to 90.8% under full disclosure. Conversely, Claude Sonnet 5 demonstrated an inverse reaction, with realized gains rising from 72.1% to 85.6% when shifted from coarse to full disclosure conditions.
This variance underscores a fundamental challenge for developers: the optimal information policy for one model may actively degrade the performance of another. The study suggests that internal weights and training methodologies—which dictate how a model prioritizes and processes contextual information—are as important as the market rules themselves.
A notable distinction in the research was the gap between "candidate quality" and "realized performance." Candidate quality measures the potential gains from trade pairs formed by submitted price rankings, while realized performance tracks successful trades. In double auctions, some models displayed high candidate quality (consistently above 0.99) while suffering a substantial decline in realized gains. For example, Claude Opus 5.5 saw its realized gains fall from 94.6% to 64.1% when moving from coarse to full disclosure, even though its ability to identify profitable trade opportunities remained near-perfect. This suggests that while some models can "see" the optimal move, they struggle to execute it under conditions of information saturation.
Behavioral Anomalies and Communication Constraints
The researchers observed consistent behavioral patterns that shed light on how LLMs navigate negotiation. Under conditions of full disclosure, sellers in every tested model raised their asking prices further above their actual costs, while buyers in nine out of ten models—with Gemini 3.7 Flash being the sole exception—lowered their offers further below their intrinsic value. These findings indicate that transparency, rather than fostering cooperation, can trigger defensive or adversarial behaviors in current-generation LLMs.
The study also addressed the impact of task complexity. When the experiments were scaled to a ten-resource environment, the results shifted dramatically. Full information improved candidate quality for all ten models and boosted realized performance for nine of them. This suggests that in more complex scenarios, the difficulty of identifying potential exchanges outweighs the potential "noise" or over-processing issues introduced by full disclosure.
Furthermore, communication length was identified as a major factor in market welfare. In a series of experiments where one party held a greater extraction opportunity, extending the allowed conversation length from two to twelve messages disproportionately favored "least disclosure" policies. For the GPT-5.1 model, welfare increased by 30.6 percentage points under least information, compared to only 2.8 points under full information. This implies that when agents are permitted to deliberate, they can infer missing information more effectively than they can process an excess of provided data.
Historical Context and Theoretical Foundations
The research draws conceptual parallels to the "Bayesian Persuasion" framework established by Kamenica and Gentzkow in 2011. That seminal work explored how a sender might strategically choose a signal to influence a receiver’s actions, a theory that has now become relevant in the context of human-AI and AI-AI interaction. While the 2011 theory provided a mathematical foundation for information design, the 2026 study serves as one of the first empirical attempts to apply these concepts to the latent space of large language models.
Industry analysts note that this research arrives at a pivotal time. As firms look to deploy agents for supply chain management, algorithmic trading, and automated procurement, the "information policy" of these systems is increasingly viewed as a form of intellectual property. If a company can optimize the information flow to its agents, it may gain a decisive edge in market negotiations. However, the study serves as a warning against premature optimization.
Implications for Future Development
For developers and systems architects, the study offers a clear roadmap for future testing. Relying on a single model or a static information setting is no longer sufficient for production-grade deployments. A robust testing methodology, according to the researchers, must include:
- Varied Information Precision: Testing performance at multiple granularities to identify the "sweet spot" for specific task sets.
- Dual-Metric Evaluation: Measuring both the identification of opportunities (candidate quality) and the successful execution of outcomes (realized performance).
- Communication Budgeting: Adjusting the allowed interaction length to reflect the complexity of the trade environment.
- Distributional Tracking: Analyzing not just the total welfare gains, but how those gains are divided between parties, ensuring the model is not unintentionally biased toward one side of a transaction.
Conclusion and Limitations
The study concludes that these results are bounded by the specific, controlled environments of the experiment and should be viewed as a guide for internal testing rather than a universal forecast for real-world market outcomes. The researchers emphasized that the models were tested against known utilities and fixed objectives, which may not mirror the chaotic, non-stationary nature of live financial markets.
As the industry moves toward more sophisticated multi-agent systems, the "Information Disclosure" variable will likely become a primary focus of competitive engineering. The evidence presented suggests that the next generation of AI-driven commerce will be defined not just by the intelligence of the agents themselves, but by the strategic design of the information environments in which they operate. For builders, the message is clear: data is not always better. In the world of high-speed, autonomous negotiation, the precision of information is a double-edged sword that requires careful, empirical calibration to master.
