NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 26·5 min read
Authority Inversion in LLM-Mediated Ubiquitous Systems: When Models Trust Users Over Sensors
TL;DR
Large language models (LLMs) often prioritize user claims over sensor data, leading to unreliable decision-making in systems that require accurate physical sensing. To address this, a geometric framework called Geometric Authority Calibration (GAC) was developed, along with two metrics: Context Integration Ratio (CIR) and Authority Alignment Index (AAI).
✦ Why It Matters
Engineers should implement GAC to enhance the reliability of LLMs in sensor-dependent applications.
Key Takeaways
- Implement GAC to improve decision-making in LLMs.
- Use CIR and AAI to audit authority allocation.
- Prioritize sensor data in conflicting scenarios.
- Evaluate model performance across diverse datasets.
Full Summary
As large language models (LLMs) are increasingly used in ubiquitous systems, a critical issue arises when these models face conflicting inputs from sensors and user claims. The study introduces a geometric framework for context integration, which includes two metrics: the Context Integration Ratio (CIR) and the Authority Alignment Index (AAI).
The researchers developed Geometric Authority Calibration (GAC) to adjust the influence of user claims during inference. Testing on four models across various datasets revealed a concerning trend where LLMs showed almost no trust in sensor data, with an AAI of -0.805.
By applying GAC, the accuracy of human activity recognition (HAR) tasks improved significantly, from 0-1.6% to 21.9-27.5%. These findings highlight the importance of explicitly managing authority allocation in LLM-mediated systems to ensure reliable outcomes.