This paper explores a fundamental question in the philosophy of science through the lens of computational modeling: Can scientific progress happen if agents only hold one theory at a time?
Drawing on Paul Feyerabend’s concept of "theoretical pluralism," the authors introduce a model where agents can maintain multiple, incompatible theories simultaneously. Their findings suggest that for a community to successfully navigate a paradigm shift, agents must be "pluralist"—capable of not just updating their data, but also tracking and communicating their confidence in the theories themselves.
TL;DR
The authors develop a multi-agent model where "pluralist" agents maintain and communicate support for multiple rival theories. They demonstrate that sharing which theory one favors is more critical for spreading paradigm shifts than simply sharing data or predictions, providing a blueprint for how AI systems might better support scientific discovery.
What Problem Does This Paper Address?
Most existing models of how scientific communities share information (network epistemology) assume agents are "monists." A monist agent folds all new observations into a single, internal model of the world. However, philosophers like Feyerabend argued that science works precisely because people dare to hold onto "incompatible rival theories." Without these rivals, the failures of a dominant theory might remain hidden.
As Large Language Models (LLMs) are increasingly used in research, a critical question arises: If an AI only maintains one "best" representation of the world, will it lead us down locally valid paths that are fundamentally wrong? The authors seek to understand what representation of the world AI systems need to avoid this trap and how different types of communication influence collective belief.
Methods
The authors use a multi-agent variational inference framework. Unlike standard models, their agents maintain two distinct levels of belief:
- Within-theory beliefs ($\rho$): A probability distribution over specific quantities (e.g., "If Theory A is true, I expect the mass to increase by X").
- Between-theory support ($q$): A weight representing how much the agent favors one theory over another.
As shown in Figure 1, agents receive private observations from a "world state" and evaluate them against their candidate theories. They then communicate with neighbors via two channels:
- Within-theory sharing: Exchanging what a theory predicts, assuming it is correct.
- Model-support sharing: Exchanging how much they actually trust the theory itself.
The model also includes "reference anchoring," a retention parameter that simulates a scientist’s commitment to a background theory, preventing them from abandoning a model the moment a single noisy data point appears.
Datasets
The paper doesn't use a traditional static dataset. Instead, it uses a simulated environment based on a classic case from the history of science: the Chemical Revolution.
- The Rival Theories: The 18th-century "Phlogiston" theory (which predicted substances lose mass during combustion) vs. the "Oxygen" theory (which predicted a mass gain).
- The Variables: The environment generates noisy observations of three variables: combustion extent, net gas uptake, and solid mass change.
Key Insights and Ideas
- The "Crisis" as Entropy: When the environment changes (e.g., mass starts increasing instead of decreasing), pluralist agents go through a period of high "individual model entropy." This spike in uncertainty is a mathematical analog to a period of "crisis" in science, where agents are genuinely torn between two explanations.
- Verdicts vs. Estimates: The most striking insight is the difference in communication. Sharing estimates (what a theory predicts) helps agents make better predictions but doesn't necessarily make them change their minds about which theory is right. Sharing verdicts (model support) is what actually triggers a community-wide paradigm shift.
- Selective Observation: If one group of agents only looks at gas uptake and another only looks at mass change, they can converge on different theories even if they talk to each other—unless they specifically share their "model support" weights.
Key Results
- Delayed Revision: The model successfully replicates the "stickiness" of scientific theories. Revision is delayed because new evidence must first overcome the accumulated support for the old theory.
- Spreading the Revolution: In experiments with a few "fully observing" agents and many "restricted" agents (who see weak evidence), the restricted agents only adopted the superior Oxygen theory when they received "model support" messages.
- Accuracy vs. Adoption: Interestingly, sharing within-theory beliefs sometimes increased prediction error for restricted agents initially, as it pulled them toward the wrong predictions of a theory they hadn't yet abandoned.
Limitations and Future Work
The authors acknowledge several limitations:
- Fixed Theories: Agents in this model don't "invent" new theories; they are given a set of candidates at the start.
- Passive Observation: Agents don't choose which experiments to run. A natural next step is to allow for "active inference," where agents choose observations that best disambiguate between their rival theories.
- LLM Integration: While the paper motivates the study through the lens of AI, it doesn't test actual LLMs. The authors suggest that future work should investigate whether LLMs behave more like monist or pluralist agents and how that affects their utility in scientific workflows.
Summary updated