Deterministic Event-Graph Substrates as World Models for Counterfactual Reasoning

2026-05-18 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

A new class of world models, called "event-graph substrates," represents state as an append-only log of typed RDF triples and answers counterfactual queries by forking the log under structured interventions. This approach offers inspectability at the triple level, supports exact counterfactuals, and transfers across domains without learned components. The substrate, implemented with a 1,400-line CLEVRER-DSL interpreter, significantly outperforms the symbolic-oracle baseline NS-DR on the CLEVRER video causal-reasoning benchmark across descriptive (9.89%), explanatory (20.26%), and counterfactual (17.65%) per-question metrics. It also exceeds the parametric ALOE baseline on descriptive and explanatory tasks. Furthermore, a new "twin-EventLog" benchmark demonstrates the substrate's superior agent memory consistency, surpassing Llama-3.1-8B by 18.80 percentage points and a Park/Concordia-style LLM simulator by 65 percentage points on joint accuracy.

Key takeaway

For research scientists developing agentic systems requiring transparent and precise causal reasoning, consider implementing event-graph substrates. This approach offers formal guarantees on inspectability and replay consistency, outperforming large language models and symbolic baselines on specific causal-reasoning tasks. You should prioritize this method when exact intervention semantics and auditable predictions are critical, especially in closed-event reasoning scenarios.

Key insights

Deterministic replay over typed event deltas enables auditable, exact counterfactual world modeling without learned latent simulation.

Principles

• Counterfactual queries reduce to graph-theoretic operations.
• Closed-event reasoning favors deterministic replay.
• Structured execution pathways are load-bearing for reasoning.

Method

Event-graph substrates use an append-only log of typed RDF triples, a deterministic replay function, and an intervention vocabulary to answer counterfactual queries by forking the log at a specific tick.

In practice

• Use event-graph substrates for auditable causal reasoning.
• Apply ancestor duality for explanatory queries.
• Augment with heuristics for emergent interactions.

Topics

• Event-Graph Substrates
• Counterfactual Reasoning
• Deterministic Replay
• Causal-Ancestor Duality
• CLEVRER Benchmark

Code references

• oxigraph/oxigraph

Best for: Research Scientist, AI Scientist, Robotics Engineer

Related on AIssential

• Deterministic Event-Graph Substrates as World Models for Counterfactual Reasoning — Event-graph substrates offer inspectable, domain-agnostic world models for exact counterfactual reasoning via RDF triple logs.
• Knowledge-Based Zero-Replay Debugging of Multi-Agent LLM Traces — Zero-replay prediction using event knowledge graphs significantly reduces the cost of debugging multi-agent LLM traces.
• Post-Deterministic Distributed Systems: A New Foundation for Trustworthy Autonomous Infrastructure — Autonomous agents challenge deterministic distributed systems, requiring a new model for coordinating diverse, semantically equivalent outcomes.
• AILS-NTUA at SemEval-2026 Task 12: Graph-Based Retrieval and Reflective Prompting for Abductive Event Reasoning — AILS-NTUA's three-stage system won SemEval 2026 Task 12 for Abductive Event Reasoning, achieving a first-place rank with an accuracy score of 0.95.
• Project Ariadne: A Structural Causal Framework for Auditing Faithfulness in LLM Agents — LLM agent reasoning traces often act as "Reasoning Theater," not faithful generative drivers, due to Causal Decoupling.
• From Chaos to Causality: Debugging Multi-Agent Systems — Causal inference, using episode graphs and counterfactual replay, can debug multi-agent systems by identifying root causes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.