AI learns to balance internal knowledge and tool use, improving efficiency

Researchers from UC San Diego and Tsinghua University have developed a method that improves artificial intelligence’s ability to understand when to use external tools versus relying on built-in knowledge, similar to how human experts approach problem-solving. Using a relatively small language model with 8 billion parameters, the team achieved a 28% improvement in answer accuracy and nearly 14% increase in tool usage precision across test datasets, outperforming larger models in specific domains. The two-step training process, called “Adapting While Learning,” first teaches the AI to learn from tool-generated solutions, then trains it to categorize problems as “easy” or “hard” and decide whether to use tools accordingly. Source: VentureBeat