Dynamic Reasoning Agent: An Adaptive Selection System for Nine Reasoning Algorithms

The ASU research team has developed a general-purpose reasoning agent system whose core capability lies in dynamically selecting optimal combinations of reasoning algorithms by analyzing task characteristics through meta-reasoning. Under the constraint of using only a single SOL LLM API, it significantly improves the ability to solve complex problems. The system integrates nine reasoning algorithms covering multiple technical routes such as basic, enhanced, verification optimization, and search strategies, and achieves adaptive reasoning through task analysis, algorithm selection decision-maker, and execution feedback loop. This article will discuss aspects including background, design, implementation, experiments, applications, and limitations.

Large language models perform well in various tasks, but when facing complex reasoning problems, a single reasoning strategy is difficult to achieve optimal results. Different reasoning algorithms have their own areas of expertise: Chain of Thought is suitable for step-by-step derivation, Self-Consistency improves answer stability, and Tree Search excels at multi-path exploration. The key problem is how to select the most appropriate algorithm for a specific task. The ASU team's research directly addresses this challenge by building a self-decision-making general-purpose reasoning agent system.

Dynamic Algorithm Selection

The core of this agent is meta-reasoning ability—analyzing problem characteristics before solving the problem, then deciding which reasoning algorithms to call, and flexibly configuring strategies based on problem complexity, type, and precision requirements.

Nine Reasoning Algorithms Library

The system integrates nine algorithms covering multiple technical routes:

• Basic Reasoning Methods: Direct generation, zero-shot Chain of Thought, few-shot prompting (suitable for simple to medium complexity problems);
• Enhanced Reasoning Techniques: Tree of Thought, Graph of Thought (handling complex scenarios with multi-step decisions);
• Verification and Optimization Methods: Self-Consistency voting, Verification Chain, Reflection Mechanism (improving answer quality and reliability);
• Search Strategies: Beam Search, Best-First Search (systematic exploration of the solution space).

Task Analysis Module

The agent first extracts features from the input problem. Evaluation dimensions include problem type (mathematics, logic, common sense, etc.), complexity (number of steps, knowledge domain), and answer format (numerical, choice, open-ended, etc.). The feature vector serves as the basic input for algorithm selection.

Algorithm Selection Decision-Maker

Select the optimal combination from the algorithm library based on task characteristics. The selection logic can use a rule engine, lightweight classifier, or LLM meta-reasoning judgment, supporting single algorithm execution and multi-algorithm combination (improving performance through result fusion).

Execution and Feedback Loop

Execute after selecting the algorithm, monitor indicators such as the number of generated tokens, confidence score, and consistency check results; if the initial selection effect is not good, the system can reselect and switch strategies.

Optimization Under Single API Constraint

The research was conducted under strict resource constraints—only using the SOL LLM API provided by ASU, without relying on multi-model integration or external tool calls. All intelligence comes from the refined use of a single model, proving the possibility of improving performance through strategy optimization rather than resource accumulation.

From Course Project to Research Prototype

As a final project for the CSE476 Natural Language Processing course, this work demonstrates the potential of transforming academic course projects into meaningful research. The code implementation focuses on modularity and scalability, providing a basic framework for subsequent research.

Reasoning Efficiency Optimization

Dynamic algorithm selection can balance performance and cost: simple problems are solved quickly with lightweight methods, while complex problems use computationally intensive advanced technologies, avoiding one-size-fits-all resource waste.

Research on Algorithm Combination Strategies

It provides an experimental platform for reasoning algorithm combination strategies. By recording algorithm selection decisions and effect feedback for different tasks, more accurate meta-decision models can be trained.

Educational Significance

As a teaching project, it shows how to integrate what is learned in the course (prompt engineering, reasoning technology, agent design) into a complete system, which is a model case of combining NLP education and practice.

The limitations of the current implementation include: the overhead of algorithm selection itself, the accuracy of feature extraction, and the generalization ability when facing completely new problem types. Future work can explore meta-learning based on historical data to allow the system to continuously improve selection strategies from past decisions.