For developers building multi-agent systems, the design of information disclosure policies is a critical variable. A preprint published on Nov. 5, 2026, examines how the precision of information provided to large language models (LLMs) affects their performance in trade and negotiation. The study suggests that increasing information does not consistently improve performance; instead, outcomes are highly dependent on the specific model and the structure of the market.
The researchers evaluated ten LLMs across three controlled environments: multi-resource trade, bilateral bargaining, and double auctions. To isolate the impact of information, they held models, incentives, and market rules constant while varying the input levels. In trade, least disclosure provides agents with their own values, coarse disclosure adds the counterparty’s ordinal ranking, and full disclosure provides exact values. In bargaining, coarse information only signals whether a mutually profitable price exists, while full disclosure reveals buyer value and seller cost. In double auctions, least disclosure gives own value or cost, coarse adds within-side rankings, and full disclosure provides exact values or costs.
The results demonstrate that the relationship between transparency and market efficiency varies significantly by model. In a two-resource trade setting, GPT-6 Sol saw its realized gains drop from 97.7% to 90.8% when moving from coarse to full disclosure. Conversely, Claude Sonnet 5 improved from 72.1% to 85.6% under the same conditions. These findings illustrate that the impact of disclosure levels is not uniform across different models.
Distinguishing between potential and completed outcomes is essential for understanding these results. Candidate quality measures potential gains from pairs formed by submitted price rankings, even if the actual bids and asks prevent those trades from clearing. Realized performance measures the outcomes that are successfully completed. In double auctions, the gap between these metrics can be revealing. For instance, Claude Opus 5.5 saw realized gains fall from 94.6% to 64.1% under full versus coarse disclosure, while its candidate quality remained above 0.99 in both conditions. The study reports this gap without establishing a general causal mechanism for why it occurs.
The researchers also observed specific behavioral patterns during the experiments. In every tested model, sellers asked for prices further above their costs under full disclosure. Similarly, under full disclosure, buyers offered prices further below their values in every model except Gemini 3.7 Flash. These observations describe the behaviors exhibited within these specific controlled settings rather than providing a universal account of how models negotiate.
Task complexity and communication constraints further influence these outcomes. In a ten-resource study, full information improved candidate quality for all ten models and realized performance for nine. The authors note that adding resources changes the available exchanges and the difficulty of finding them, which complicates any assumption that coarse information is inherently superior. Furthermore, a separate communication-length experiment focused on markets where one party held greater extraction opportunity. Extending conversations from two to twelve messages improved welfare more under least than full information for eight of ten models. For GPT-5.1, welfare increased by 30.6 percentage points under least information and 2.8 points under full information. This finding is specific to that tested subset and does not establish a general effect for all market settings.
This research shares a conceptual connection to Bayesian Persuasion, a 2011 article by Kamenica and Gentzkow. Their work provides a theoretical framework for a sender choosing a signal for a receiver whose action affects both parties’ welfare. It is important to note that this is a theoretical model, not empirical evidence regarding current LLM deployments.
For builders, the practical evaluation recipe must move beyond testing a single model under one information setting. A robust test plan should vary the model, the precision of information shared, the recipient of that information, the market structure, the specific objectives, and the communication budget. Builders should record both the opportunities agents identify and the trades they successfully complete. Furthermore, it is necessary to track how gains are divided, rather than focusing solely on the total, as these measures should not be collapsed into a single score.
The evidence provided by this study remains bounded to selected controlled markets with known utilities and explicit objectives. Results vary by model, market structure, and interaction rules, and the reported confidence intervals capture variation across markets rather than repeated model responses to identical inputs. The authors emphasize that these experiments are a reason to test information policy, not a forecast of live market behavior, as the generalization to real-world or deployed-agent scenarios remains untested.