AdaHome: An Adaptive Smart Home Assistant using Loc...

AdaHome: An Adaptive Smart Home Assistant using Loc... | AI Research

Key Takeaways

AdaHome is an adaptive smart home assistant designed to run locally on small language models (SLMs).
Smart home assistants interpret a wide range of user commands, from explicit device control to underspecified and preference dependent requests.
In addition, existing approaches provide limited support for stable long-term personalization.
To address these issues, we present AdaHome, an adaptive smart home assistant designed for locally deployed small language models in smart home environments.
Rather than applying complex reasoning uniformly, AdaHome introduces an intent-aware planning framework that dynamically routes commands either to straightforward prompt-based or lightweight reasoning-based components.

Paper AbstractExpand

Smart home assistants interpret a wide range of user commands, from explicit device control to underspecified and preference dependent requests. While recent systems based on Large Language Models (LLMs) improve this capability, they often rely on heavyweight reasoning pipelines and cloud-based deployment, limiting their efficiency and suitability for resource-constrained environments, and raising privacy concerns. In addition, existing approaches provide limited support for stable long-term personalization. To address these issues, we present AdaHome, an adaptive smart home assistant designed for locally deployed small language models in smart home environments. Rather than applying complex reasoning uniformly, AdaHome introduces an intent-aware planning framework that dynamically routes commands either to straightforward prompt-based or lightweight reasoning-based components. For commands requiring interpretation, we adopt a Chain-of-Draft strategy to enable efficient and stable decision-making. To support personalization, we further propose a preference adaptation mechanism that learns from user feedback over time without requiring prompt augmentation or model retraining. We evaluate AdaHome against representative LLM-based baselines under a unified small model setting. AdaHome achieves substantially higher accuracy on direct commands (86.7%) while reducing latency by up to 3$\times$. Furthermore, it maintains competitive performance on ambiguous inputs with lower computational cost. In multi-turn scenarios, AdaHome achieves 88% preference consistency, compared to 52.5% for a prompt augmentation baseline.

AdaHome is an adaptive smart home assistant designed to run locally on small language models (SLMs). By moving away from cloud-based, resource-heavy systems, it addresses privacy concerns and high latency while maintaining the ability to interpret complex, ambiguous user commands. The system intelligently balances efficiency and reasoning, allowing it to provide personalized, reliable home automation on low-cost hardware.

Intent-Aware Planning

Not every command requires the same amount of processing power. AdaHome uses an intent-aware framework to categorize user inputs into direct, indirect, or ambiguous commands. Direct commands—such as "turn on the light"—are handled by a straightforward planner that executes actions immediately. For more complex or ambiguous requests, the system routes the command to a reasoning planner. This selective approach ensures that the system only uses computational resources when necessary, significantly reducing latency.

Efficient Reasoning with Chain-of-Draft

To handle complex commands without the overhead of traditional, verbose reasoning methods, AdaHome employs a "Chain-of-Draft" strategy. Instead of generating long, multi-step explanations that can lead to errors in small models, the system produces a concise "draft" of the intent followed by a structured JSON action plan. This compact representation provides enough context for the model to make accurate decisions while minimizing token usage and preventing the hallucinations often associated with over-reasoning in smaller models.

Continual Personalization

AdaHome adapts to individual user habits over time without the need for expensive model retraining or complex prompt augmentation. It maintains a local "preference memory" that stores past interactions. When a user issues a command, the system retrieves semantically similar historical actions and applies a decay-based weighting system. This ensures that recent and highly relevant past behaviors have a stronger influence on current decisions. As the system gathers more feedback through user confirmations or corrections, it becomes increasingly personalized to the specific user's environment and preferences.

Performance and Efficiency

In evaluations using a unified small model setting, AdaHome demonstrated significant improvements over existing LLM-based baselines. It achieved an 86.7% accuracy rate on direct commands while reducing latency by up to 3 times compared to other systems. Furthermore, in multi-turn scenarios, AdaHome maintained an 88% preference consistency, proving its ability to learn and retain user habits effectively. By operating locally, the system provides a responsive and private alternative to cloud-dependent smart home assistants.