Thinking Fast and Laterally: Multi-Agentic Approach for Reasoning about Uncertain Emerging Events

TL;DR

The paper addresses the challenge of reasoning under uncertainty for anticipatory events in streaming data. It introduces SaLT, a multi-agent framework that implements System-2-like reasoning through lateral thinking via dynamic topology and belief propagation. The authors contribute a systematic framework for generating lateral-thinking queries and datasets, the SaLT architecture with dynamic inter-agent topology and belief management, and empirical evidence showing improvements in hypothesis generation and retrieval over single-agent baselines. The work highlights the potential of long-range inter-agent information flow and memory-conscious processing for real-time risk monitoring and decision support.

Abstract

This paper introduces lateral thinking to implement System-2 reasoning capabilities in AI systems, focusing on anticipatory and causal reasoning under uncertainty. We present a framework for systematic generation and modeling of lateral thinking queries and evaluation datasets. We introduce Streaming Agentic Lateral Thinking (SALT), a multi-agent framework designed to process complex, low-specificity queries in streaming data environments. SALT implements lateral thinking-inspired System-2 reasoning through a dynamic communication structure between specialized agents. Our key insight is that lateral information flow across long-distance agent interactions, combined with fine-grained belief management, yields richer information contexts and enhanced reasoning. Preliminary quantitative and qualitative evaluations indicate SALT's potential to outperform single-agent systems in handling complex lateral reasoning tasks in a streaming environment.

Figures (6)

• Figure 1: Conceptual overview of the $\textsc{SaLT}$ framework. This diagram illustrates the multi-agent structure and dynamic communication flow in processing lateral reasoning queries, such as "monitor the risk to the supply chain of American semiconductor companies". The right side shows a timeline of events that gradually unfold in a datastream over time. The left side depicts the specialized agent network, representing different topics in the problem space. This network structure is dynamically generated to address the specific lateral reasoning query. Connecting lines between agents and events indicate the reasoning criteria employed by each agent to interpret the temporal context.
• Figure 2: (a) Shows normalized lateral measure, time lag complexity, and uncertainty spread across various anticipatory queries. (b) Shows an evaluate of $\textsc{SaLT}$ over a dataset with high time-lag complexity and profiles performance as a function of query complexity.
• Figure 3: Illustration of a scenario where a drought in one part of the world can affect the energy security in distant places.
• Figure 4: Illustration of a scenario where a climate change induced event can lead to upheaval in an industry.
• Figure 5: Example lateral thinking queries.