Beyond the Collective World Model: From Shared Language to Causally Identifiable Relational Fields
Year: 2026
Type: preprint
Abstract: The observation that contemporary language models acquire useful world-related regularities from collectively produced text is historically established, but it leaves an important scientific question unresolved: which relational properties of the processes that produced the text can be identified from a learned representation, and under what circumstances can those properties support causal claims? This paper proposes a bounded framework for investigating candidate relational Fields across successive regimes of experience, communication, text, computational representation, and situated action. A Field is specified as a typed and mission-relative structure consisting of a domain, states, couplings, transitions, observations, interventions, boundaries, and evidentiary witnesses. We distinguish statistical reproduction, relational identification, causal employment, and cross-regime preservation, and formulate observational non-identifiability, witness dependence, and composition constraints. The account is deliberately compared with established alternatives: causal abstraction, structural causal models, predictive state representations, active inference, and information-theoretic approaches to collective coordination. Taniguchi's Collective World Model and Generative Emergent Communication serve as one substantial architectural counterpart within a longer history of distributed semantics and shared learning, not as the historical origin of the general premise. Three finite constructions illustrate why text reconstruction does not guarantee unique source recovery, why mathematically consistent correspondence can remain false, and why intervention evidence changes causal conclusions. No new agent experiment, language-model benchmark, or physical Field observation is claimed. Instead, the manuscript supplies a falsifier-first research programme with information-matched baselines, independent witnesses, predeclared refusal conditions, and explicit outcomes under which the Field hypothesis should be reduced or rejected. Its contribution at this stage is a testable specification and a carefully bounded research agenda, not demonstrated superiority over conventional mathematical models.
Author: Figurelli, Rogério
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225775783
Cites:
Cited by:
Related to: