Source-linked AI summary
Semantic Bayesian World Models
Tommaso Soru
TL;DR
The paper addresses the mismatch between probabilistic foundation-model reasoning and crisp knowledge-graph assertions. It proposes Semantic Bayesian World Models, in which ontologies constrain priors, observations update graph beliefs, and actions intervene on the world, enabling semantically grounded and inspectable agent reasoning.
Problem
Foundation models and autonomous agents reason in probabilities, while knowledge graphs represent reality as crisp assertions, leaving their integration as a data-feeding pipeline rather than unified reasoning architecture.
Method
The paper defines Semantic Bayesian World Models as probabilistic distributions over RDF graphs with ontology-based priors, Bayesian observation updates, transition kernels for actions, and belief annotations in RDF 1.2.
Results
The proposed model supports separately sourced and updateable beliefs, ontology-constrained hypotheses, computed interventions, and uncertainty propagation through semantic entailment and planning.
Takeaways & Limitations
SBWMs provide a shared, dereferenceable, and machine-actionable fabric of beliefs for agents to exchange, inspect, update, and act upon under uncertainty.
Takeaways & Limitations
The proposal remains constrained by probabilistic identity, #P-hard inference, combinatorial conditional tensors, and poorly calibrated logit-derived confidences.
Abstract
from arXiv · showhide
Knowledge graphs describe reality in crisp assertions, while the systems now consuming them, foundation models and autonomous agents, reason natively in probabilities. We argue that this mismatch is why the integration of language models and knowledge graphs remains a data-feeding pipeline rather than a unified reasoning architecture. We envision Semantic Bayesian World Models (SBWMs): a Web that describes the world not as a database of facts but as a shared, evolving fabric of beliefs over knowledge graphs, where ontological axioms constrain priors, observations update beliefs by Bayesian conditioning, and actions intervene upon the world. We work through what an agent gains from such a model: a home-security agent deciding whether the figure at the gate is a courier or a burglar, an actuarial estimate aggregated by entailment rather than by string frequency, a planning task that language models reliably fail, and the estimation of quantities that no document has ever stated. We then set out what the community must build to make them possible: belief annotation over RDF~1.2, probabilistic entailment regimes, semantic calibration layers, and protocols by which agents that have never met can exchange, and disagree over, calibrated beliefs.
1 Introduction
The paper identifies a representational mismatch between crisp knowledge graphs and probabilistic foundation models, then proposes Semantic Bayesian World Models as a unified architecture for evolving, actionable beliefs.
- Knowledge graphs encode crisp assertions and entailments but cannot represent consumers’ graded epistemic states.Their open-world assumption models what is unsaid, not how strongly stated claims are believed.
- Foundation models reason in probability distributions, yet their integration with knowledge graphs remains a retrieval-and-prompting pipeline.The paper characterizes the obstacle as representational rather than merely technical.
- Semantic Bayesian World Models represent the Web as shared, evolving beliefs over knowledge graphs.Ontological axioms constrain priors, observations update beliefs by Bayesian conditioning, and actions intervene on the world.
- The paper argues that knowledge should be organized semantically and probabilistically as stable propositions with coherent degrees of belief.It presents this as a deliberately strong conjecture about scaling language models to superhuman intelligence.
- The paper develops the proposal through its model, agent capabilities, and the infrastructure required to build it.The three movements cover what an SBWM is, what agents gain from it, and the community’s construction agenda.
2 Semantic Bayesian World Models
SBWMs combine Bayesian uncertainty, knowledge-graph semantics, and foundation-model learning in a graph-based world model whose beliefs can be updated, queried, completed, and acted upon.
- Motivation and positioning: SBWMs occupy the previously unoccupied intersection of probabilistic, semantic, and learned representations.Bayesian networks, knowledge graphs, and foundation models each provide complementary capabilities, while their three-way combination remains open.
- Model definition: An SBWM is a tuple M = (Σ, P0, T, O) containing an ontology, graph prior, action transition kernel, and observation model.Its belief state is a distribution over graphs updated by Bayes’ rule, with the ontology functioning as the prior.
- Ontology-constrained inference: Entailment coherence constrains probabilistic beliefs so that refinement cannot increase mass beyond the more general proposition.Subclass and subproperty relations transfer these bounds across shared URIs and languages.
- Ontology-constrained inference: Ontologies can supply priors for unseen subclasses, while semantic calibration projects neural scores onto the axiom-consistent feasible region.The paper illustrates repair when a predicted cup probability exceeds its container probability.
- Causal querying: SPARQL can be adapted for causal inference: WHERE performs conditioning, while UPDATE implements interventions through graph surgery.Belief annotations attach to RDF statements, and conditional structure is represented in lifted networks over triple patterns.
- Representation and learning: Sparse semantic tensors represent graph priors and conditionals, enabling completion of quantities absent from every source.The proposed completion process uses semantically adjacent cells to produce posterior estimates with provenance.
- Building an SBWM: Construction can combine probabilistic extraction, semantic aggregation, tensor completion, transition kernels, and observation models using existing components.The paper’s central implementation gap is preserving extracted probabilities instead of thresholding them away.
3 SBWMs at Work
SBWMs let agents combine separately sourced, semantically identified beliefs into inspectable posteriors, then update and act through ontology-constrained reasoning. The examples span security inference, actuarial aggregation, and planning under uncertainty, illustrating capabilities unavailable to crisp knowledge graphs or string-based language models alone.
- 3.1 A Camera in the Garden: Bayesian decomposition combines beliefs about a visitor’s goal from different publishers under a shared vocabulary and explicit provenance.The beliefs are attached to stable triple identifiers, allowing the agent to merge and inspect them without jointly trained models.
- 3.1 A Camera in the Garden: Ontology constraints make competing explanations share probability mass and guarantee P(Visitor) ≥P(Courier) when Courier and Burglar are disjoint subclasses of Visitor.The constraint holds regardless of the neural scorer’s behavior.
- 3.1 A Camera in the Garden: Interventions let the agent evaluate actions such as switching on a floodlight by updating the world graph and comparing predicted outcomes under each hypothesis.The action is computed through graph surgery and a transition kernel rather than selected from a prompt alone.
- 3.2 An Actuary’s Query: Semantic aggregation estimates actuarial risk over entailed vehicle classes and instances instead of counting near-identical strings.The semantic likelihood attaches to concepts, so equivalent claims in different languages update the same belief.
- 3.3 The Car Wash Test: Formal backward chaining shows that washing the car requires locating it at the car wash, so the plan must move the car rather than walk.Domain and subproperty axioms connect the washing goal to the required location relation.
- 3.4 Why Symbols Matter: Across the vignettes, SBWMs isolate, source, update, and aggregate beliefs over entailments, while string associations attach credence to phrasings rather than propositions.The paper argues that forecasting, planning, and science require these semantic and probabilistic operations.
4 What Must Be Built
Making Semantic Bayesian World Models practical requires infrastructure that carries calibrated beliefs, provenance, semantic constraints, and disagreement across agents. The agenda also faces open challenges in probabilistic identity, tractable inference, and confidence calibration.
- A W3C belief-annotation vocabulary should transmit probabilities, calibration methods, and provenance together.
- Probabilistic entailment should treat classical entailment as the probability-one case and subsumption as a monotonicity constraint.
- Semantic calibration layers should connect neural scorers and triple stores, while probabilistic SHACL represents shapes as soft constraints with violation costs.
- Federated belief exchange should let unfamiliar agents publish, merge, and revise priors under explicit provenance despite numerical disagreement.
- Open challenges include probabilistic identity, #P-hard weighted model counting, combinatorial conditional tensors, and model-dependent confidence calibration.
- The vision should be judged by calibration and coherence, not merely accuracy, while existing probabilistic components remain fragmented by discarded confidences.