Agent Reasoning with Graph World Models
summary
Claims that existing Neo4j graphs convert directly into lightweight world models and that storing two simple statistics per transition edge enables rollouts that guide action selection. Demonstrations use a toy RL rollout evaluator over a 2D grid, a scene with object interactions, and a clinical-trial planning pipeline; toggling the model off vs on shows failures vs reliable task completion. Code runs against Neo4j + Cypher in a dockerized Neo4j instance, using transition-edge stats (mean, visit count) to drive predictions.
AI-written summary · about
chapters (8)
- 0:00Introduction to talk
- 1:33World models and graphs
- 5:25Spatial environment demo
- 8:17Scene-based scenario demo
- 9:42Generalized environment demo
- 13:13Building world models
- 16:23Conclusions and references
- 18:56Q&A session
related talks
17mContext Graphs for Explainable, Decision-Aware AI Agents
14mUsing Graphs to Fine-Tune Agents
18mConnecting the Dots with Context Graphs
01h 26mThinking in Graphs: The Cognitive Foundations of Agentic AI
35mGraphRag in Action: Revolutionizing Fraud Detection Techniques
30mPerformant GraphRAG Improving Neo4j to Drive Autonomous AI ...
28mGrounded Intelligence Boosting LLMs with Real-Time Graph Context
24mUnlocking AI's Hidden Connections With Graphs