In classical game theory and mechanism design, the Common Prior Assumption (often called the Harsanyi Doctrine) serves as a foundational bedrock. It posits that if rational agents disagree about the likelihood of an event, it is purely because they have observed different pieces of information. At the beginning of time, before any data was gathered, every agent shared an identical statistical model of the universe.
While mathematically elegant, this assumption introduces severe friction when applied to real-world systems. Humans and autonomous agents do not emerge from a factory pre-programmed with the same baseline probability distributions. They possess divergent backgrounds, rely on different environmental sampling methods, and harbor distinct subjective biases. When a system assumes a uniform prior that does not exist, its mathematical guarantees evaporate.
A rapidly growing body of research focuses on relaxing this assumption, rewriting the rules of game theory, information elicitation, and social choice to accommodate heterogeneous beliefs.
Why Consensus Fails to Materialize
To understand why the common prior assumption is so restrictive, consider Robert Aumann’s famous 1976 theorem showing that honest, rational Bayesian agents with a common prior cannot "agree to disagree." If their opinions on an event are common knowledge, their beliefs must collapse into an identical consensus. Yet, in the real world, experts look at the same data and disagree constantly.
Why does consensus fail? One historical critique was that calculating agreement is too computationally or communicatively taxing for real agents. However, [Scott Aaronson] in The Complexity of Agreement demonstrated that communication and computation are not the fundamental bottlenecks preventing consensus. Instead, persistent divergence of beliefs stems directly from the absence of a common prior. When agents start with different foundational assumptions, sharing information or increasing communication does not force their beliefs to converge.
To model this mathematically without the common prior crutch, researchers have had to expand the boundaries of probability theory itself. For instance, [Eric Pacuit] et al. in Common p-Belief with Plausibility Measures: Extended Abstract generalized the agreement framework to qualitative "plausibility measures," mapping out the precise boundaries of how consensus can or cannot form when agents start with heterogeneous biases.
The Breakdown of Incentive Engineering
The absence of a common prior is not just a philosophical problem; it breaks the machinery of information elicitation and peer prediction. Peer prediction mechanisms are designed to incentivize honest reporting (such as crowdsourced product reviews or expert forecasts) when there is no objective ground truth to verify the answers. They work by comparing an agent's report with reports from their peers, using a scoring rule calibrated to a shared prior distribution.
As [Aviv Zohar] et al. observed in Robust Mechanisms for Information Elicitation, traditional scoring rules fail when agents sample different environments or hold divergent beliefs. If an agent knows their prior differs from the system's assumed prior, the agent's optimal strategy shifts: they may lie about their true observation to better align with what they think the "average" peer will report.
This vulnerability extends across several domains:
- Subjective Evaluations: In Incentives for Subjective Evaluations with Private Beliefs, [Goran Radanovic] et al. highlight that heterogeneous backgrounds make standard peer-prediction mechanisms ineffective because participants operate under diverse private beliefs rather than a single shared model.
- Prediction Markets: The very existence of a prediction market often contradicts the common prior assumption. As [Vincent Conitzer] points out in Prediction Markets, Mechanism Design, and Cooperative Game Theory, the primary purpose of a prediction market is frequently to discover a baseline probability distribution; assuming a shared starting point from the outset defeats the purpose of the market itself.
- Information Aggregation: When a principal attempts to pool expert opinions, a lack of a common prior makes it impossible to reliably distinguish between an agent's genuine new information and their deep-seated initial bias, a challenge detailed by [Rafael Frongillo] et al. in Elicitation for Aggregation.
Designing for Heterogeneous Beliefs
If we cannot assume a common prior, how do we build robust multi-agent systems? Modern research tackles this by designing mechanisms that either dynamically learn agent beliefs or remain invariant to them.
One approach is to adaptively learn incentives on the fly. In Partial Truthfulness in Minimal Peer Prediction Mechanisms With Limited Knowledge, [Goran Radanovic] et al. proved that while perfect truthfulness is impossible when the mechanism lacks knowledge of the agents' subjective priors, one can achieve "partial truthfulness" by using multi-armed bandit algorithms to learn and adjust incentives dynamically.
Another angle is to design mechanisms that remain robust even under strategic collusion or complex correlation structures. [Grant Schoenebeck] et al. have systematically analyzed these boundaries. In Putting Peer Prediction Under the Micro(economic)scope and Making Truth-telling Focal and False Consensus, Information Theory, and Prediction Markets, the authors explore how sequential public announcements can still lead to full information aggregation rather than a false consensus, despite the lack of symmetric priors. Furthermore, in Strong Equilibria in Bayesian Games with Bounded Group Size, the authors provide mathematical thresholds for collusion robustness when relaxing the common prior assumption.
Similarly, in recommendation systems and platform design, [Anand Kalvit] et al. demonstrate in Incentivized Exploration via Filtered Posterior Sampling how platforms can use filtered information maps to guide selfish users toward exploration, even when those users hold divergent or inaccurate initial beliefs that would normally undermine standard messaging policies.
New Solution Concepts for Game Theory
When players hold fundamentally different models of reality, traditional solution concepts like the Nash Equilibrium or standard Bayesian Equilibrium lose their predictive power. This has forced a rewriting of equilibrium theory.
[Tamer Başar] in An Equilibrium Theory for Multi-Person Decision Making with Multiple Probabilistic Models extended stochastic team theory and game theory to environments where decision-makers operate under entirely different subjective probability models. This work revealed that relaxing the common prior can fundamentally alter the mathematical structure of equilibria, such as turning linear Gaussian Stackelberg solutions nonlinear.
To capture even more complex interactions—such as when agents have beliefs about other agents' incorrect beliefs—[Jack Foxabbott] et al. introduced Incomplete Information Multi-Agent Influence Diagrams (II-MAIDs) in Higher-Order Belief in Incomplete Information MAIDs. By abandoning the common prior, they provide a recursive best-response framework capable of modeling realistic human-AI safety scenarios where agents hold conflicting or incorrect models of reality.
Finally, [Joseph Halpern] et al. have systematically re-evaluated the epistemic foundations of games. In Extensive Games with Possibly Unaware Players and Characterizing Solution Concepts in Terms of Common Knowledge of Rationality, the authors demonstrate how strategic equilibria can be characterized or extended to settings where players maintain divergent beliefs about external events or are entirely unaware of certain moves, bypassing the restrictive constraints of traditional common knowledge.
Knowledge Fusion and State Compatibility
Beyond strategic games, the challenge of heterogeneous priors manifests in data fusion and quantum mechanics. When merging information from multiple sources, assuming a uniform yardstick is often fatal.
In the realm of knowledge representation, [Guilin Qi] et al. in Merging Stratified Knowledge Bases under Constraints developed operators to resolve conflicts between prioritized knowledge bases without requiring a common priority scale. In collective decision-making, [Nahla Ben Amor] et al. in Egalitarian Collective Decision Making under Qualitative Possibilistic Uncertainty: Principles and Characterization showed how groups can reach fair consensus under qualitative uncertainty when agents possess non-homogeneous knowledge bases.
Even in quantum information theory, [Min-Hsiu Hsieh] et al. in Compatibility of state assignments and pooling of information highlight the open challenge of modeling consensus and information pooling when different observers hold conflicting initial biases rather than starting from a uniform, maximally mixed state.
Across all these fields, moving beyond the common prior assumption represents a shift toward realism—acknowledging that diversity in belief is not merely an informational asymmetry to be resolved, but a fundamental property of multi-agent systems.
Key Foundational Work
The Complexity of Agreement (2004) — contributed 2 concept elements to this direction.
Robust Mechanisms for Information Elicitation (2006) — contributed 1 concept elements to this direction.
Merging Stratified Knowledge Bases under Constraints (2006) — contributed 1 concept elements to this direction.