STRUCTURED TOOL-ONLY EXECUTION TO REDUCE HALLUCINATIONS IN AUTONOMOUS DESKTOP ASSISTANTS

Name of 1st Author (Abdul Wahid), Department of Computer Science and Engineering from Shivalik College of Engineering, Dehradun.

Name of 2nd Author (Abhinav Shrivastav), Department of Computer Science and Engineering from Shivalik College of Engineering, Dehradun.

Name of 3rd Author (Ayush Panwar), Department of Computer Science and Engineering from Shivalik College of Engineering, Dehradun.

Name of 4th Author (Deepak kumar), Department of Computer Science and Engineering from Shivalik College of Engineering, Dehradun.

Name of 5th Author (Kumar Rethik), Department of Computer Science and Engineering from Shivalik College of Engineering, Dehradun.

Designation of 1st Author (Student), Designation of 2nd Author (Student), Designation of 3rd Author (Student) Designation of 4thAuthor (Student), Designation of 5th Author (Guide)

Name of organization (Department of Computer Science and Engineering from Shivalik College of Engineering, Dehradun, India).

Abstract: The rapid growth of Large Language Models (LLMs) has enabled the development of intelligent assistants capable of understanding natural language and performing complex tasks. However, the tendency of these models to generate inaccurate or unsupported information remains a major challenge, particularly in systems that interact directly with operating systems and user resources. This paper presents RAYA v3.1 (Reasoning Assistant for Your Automation), an autonomous desktop assistant designed to improve reliability through a structured Tool-Only Execution Framework. The proposed approach restricts task execution to predefined and validated tools, reducing the likelihood of unsupported outputs and unsafe actions. The architecture incorporates a planner–executor model, schema-based validation, contextual memory, recovery mechanisms, and user confirmation for sensitive operations. Experimental observations indicate that structured tool enforcement improves execution consistency and enhances overall system trustworthiness. The study demonstrates that reliability can be strengthened through architectural design choices rather than relying solely on prompt engineering techniques.

Keywords: Autonomous Agents, Desktop Automation, Large Language Models, Tool-Based Reasoning, Hallucination Reduction, Intelligent Assistants.