Chakraborty, Pinaki; Caulfield, Tristan; Pym, David; (2026) Local Causal Reasoning in Multiagent Systems. In: Proceedings of The 2nd International Workshop on Causality, Agents and Large Models (CALM-25). Springer (In press).

Text Local_Causal_Reasoning_CALM2025.pdf - Accepted Version Access restricted to UCL open access staff until 14 June 2026. Download (617kB)

Abstract

Causal reasoning is essential for the design, audit, and interpretation of decision-making in multi-agent systems. Recent developments have brought this need to the fore in multi-agent LLM systems, notably in retrieval-augmented generation (RAG), where techniques from information retrieval are used to augment model inference within modular workflows. We propose a behaviour-centric model of system configurations and a unified language for reasoning about such systems. Our framework introduces an intervention operator that captures the notion of mechanism change, reflecting interventionist views of causation, while a separation-logic-style conjunction supports local reasoning via explicit system interfaces, consistent with mechanistic accounts that explain phenomena through organized and modifiable parts. Agent policy changes are treated as interventions on the components they control, enabling counterfactual analysis and attribution of responsibility within the same logic. We define actual causation directly in this language and show, via time-unfolding of finite system runs, that our notion aligns with the Halpern-Pearl account of actual causation in the acyclic structural model induced by the run. We establish van Benthem-Hennessy-Milnerstyle correspondence results: a bisimulation that respects both transitions and interventions characterizes logical equivalence under finiteness assumptions. Thus, we integrate system evolution with modular decomposition within a single language: its modalities refer directly to configuration transitions, interventions on mechanisms, and interface-indexed decompositions. We apply the framework to a retrieval-augmented generation (RAG) workflow for LLM-based systems to specify explicit interfaces, model mechanism changes as interventions, and answer design-time causal queries such as, whether some admissible mechanism change guarantees a stated safety constraint while preserving invariants modularly across an interface.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item