Dual Engines of Thoughts (DEoT): A Framework for Multi-Agent, Multi-Level Reasoning

Dual Engines of Thoughts (DEoT) is a multi-layered, multi-agent analysis framework designed specifically for handling open-ended problems. Inspired by the complexity of human reasoning processes, DEoT introduces a dual-engine model that emphasizes both breadth and depth in reasoning. Unlike traditional single-path generation models, DEoT enables flexible switching between exploratory and deductive thinking, making it particularly effective for advanced tasks such as analytical question answering, decision generation, and strategic recommendations.
‍
By dynamically balancing exploration and deduction, DEoT outperforms models like GPT-4o in areas such as depth of analysis, logical coherence, and innovation, achieving an overall success rate of 86% in benchmark evaluations. This demonstrates DEoT's leading capabilities in complex problem-solving and reasoning.

Technical Overview

Core Technology: Intelligent Mode Switching

The system autonomously modulates its reasoning engine and exploration strategies in response to input signals, task context, and the stage of reasoning progression.

Initially, it rapidly maps the global problem landscape during early-stage analysis.

Subsequently, it transitions to concentrated, in-depth reasoning on critical elements.

Dual Reasoning Engines: Breadth Exploration × Depth Tracing

The Breadth Engine conducts wide-angle analysis across multiple dimensions, uncovering latent hypotheses, key actors, and causal triggers.

The Depth Engine constructs extended logical chains, simulates complex causal dynamics and multi-stage evolutions, and forecasts potential paths and outcomes.

Engine Controller and Entropy-Guided Switching

A central controller manages dynamic transitions between breadth and depth modes based on information entropy, exploration coverage, and reasoning progress.

The framework introduces the Effective Reasoning Information Rate (ERIR) as a quantifiable metric to evaluate and optimize the system’s overall reasoning efficiency.