Abstract
This paper introduces CoCre-Sam (Collective-Creature Sampling), a framework that explains the mysterious coherence of Ouija boards (Kokkuri-san) as a decentralized sampling process. By modeling participants as agents with internal language models exerting physical forces, we demonstrate that their collective interaction performs Langevin MCMC sampling from a fused probability distribution. This approach provides a mathematical explanation for how structured meaning emerges from noisy, decentralized micro-actions, effectively acting as a real-world "Product of Experts."
Introduction
The "ideomotor effect"—unconscious muscle movements—has long been used to explain how a Ouija board pointer moves. However, this psychological explanation fails to account for the computational mystery: how do multiple people, each with different thoughts and linguistic biases, collectively "negotiate" to spell out coherent, grammatically correct sentences that no single participant intended? This paper addresses the gap between individual implicit knowledge and emergent collective behavior by proposing a mathematical framework that describes how decentralized, noisy micro-actions aggregate through a shared physical interface to produce structured meaning.
The CoCre-Sam Framework
The authors propose a hierarchical control framework called CoCre-Sam to model the collective interaction of participants as a dynamical system.
Agent Representation and Dynamics
Each participant (agent) is endowed with an internal character-based language model (LM). This LM defines an energy landscape over the letter board, where lower energy corresponds to more probable next letters. The movement of the pointer (planchette) is modeled as a dynamical system where each agent exerts a "micro-force" based on the local gradient of their own energy landscape, perturbed by stochastic noise representing hand tremors or uncertainty.
Theoretical Equivalence
A core contribution of this work is the proof that the sum of these individual forces is equivalent to the Unadjusted Langevin Algorithm (ULA). This means the planchette’s trajectory is actually a process of sampling from a fused energy landscape (the sum of individual landscapes). This physical fusion approximates a Product of Experts (PoE) model, where the resulting distribution favors outcomes that are acceptable to all participants rather than just one.
Experimental Setup and Results
To validate the mechanism, the authors used controlled, synthetic datasets consisting of flower names and biased agents trained on character-level 6-gram models.
Key Findings
- Meaningful Emergence: In simulations with agents biased toward different flower types (e.g., "colorful" vs. "subdued"), CoCre-Sam successfully generated linguistically plausible words that balanced the preferences of both agents.
- Perplexity Scores: Quantitative analysis showed that sequences generated by CoCre-Sam had lower perplexity when evaluated against the fused model (1.919) than when evaluated against individual agent models, confirming a successful negotiation of knowledge.
- The Role of Stochasticity: Noise in human movement is not just interference; it provides the "thermal fluctuations" necessary for Langevin dynamics to explore the energy landscape. "Middle Noise" conditions produced the best balance of reliability and diversity.
Discussion and Conclusion
CoCre-Sam suggests that the "spirit" (Kokkuri-san) is a mathematical fusion—an emergent collective creature represented by the fused probability distribution. Because the resulting trajectory isn't traceable to any single person, participants lose their "sense of agency" and naturally attribute the movement to an external entity. This work demonstrates that language model fusion can happen through embodied physical interaction in the real world. Future work will involve integrating state-of-the-art LLMs and conducting experiments with human participants to verify if actual physical movements follow the predicted Langevin dynamics.
Summary updated