Small Models Can Also Have Great Wisdom: How Agentic Workflows Compensate for the Disadvantages of Parameter Scale

In the field of large language models, the scale race of "more parameters equal better performance" has led to high costs and deployment barriers. The open-source project "workflows-over-weights" proposes a hypothesis: using agentic workflows (web search + self-criticism loops) to enable small models with 7B parameters to challenge large models in expert-level benchmark tests, exploring the feasibility of small models compensating for their parameter disadvantages.

Currently, there is a model parameter scale race in the AI field. Top models have hundreds of billions or even trillions of parameters; although they perform well, they bring heavy computational burdens, high deployment costs, and environmental pressures. Small and medium-sized enterprises and individual developers can hardly afford the cost of using large models, raising a key question: Is a huge model really necessary to solve practical problems?

Agentic workflows consist of three core components:

1. Tool usage: Proactively call web search to expand knowledge boundaries;
2. Self-criticism and reflection: After generating an initial answer, check its accuracy, completeness, and logic, and correct any issues;
3. Multi-round iterative optimization: Combine tool usage and self-criticism to gradually approach the optimal solution.

The project selects HLE-Verified as the test benchmark, covering expert-level fields such as scientific reasoning, mathematical proof, code generation, and knowledge Q&A. The test model is Qwen2.5-7B, which has advantages like low deployment cost, fast inference speed, energy efficiency, open-source control, etc.

The evaluation pipeline includes:

1. Baseline test: Performance of the pure model without agentic enhancement;
2. Workflow-enhanced test: Analyze the problem → Call search → Generate initial answer → Self-criticism → Revise → Output final answer;
3. Comparative analysis: Compare the baseline and enhanced modes, and compare small models + workflows with large models.

Enlightenments from the technical route:

1. Knowledge retrieval is better than parameter memory; models should learn to retrieve and use knowledge efficiently;
2. Iterative optimization is the key to intelligence, simulating the human process of repeated deliberation;
3. Small models have broad commercial prospects; local deployment can reduce costs and protect privacy.

The method has limitations: increased latency, cost of search API calls, and risk of error accumulation. Future directions include optimizing iterative strategies, intelligent tool selection mechanisms, and collaborative optimization of workflows and model fine-tuning.

This project represents a paradigm shift from pursuing large models to pursuing smart systems. Intelligence is the embodiment of problem-solving strategies and metacognitive abilities. Small models can exert great value through workflow design, promoting the democratization of AI technology. We look forward to subsequent experimental data and scenario applications.