Large Language Models (LLMs) are advanced AI models that can generate content and answer questions with a high level of accuracy. These models have been trained using deep learning techniques and a large amount of text data. They excel in tasks related to Natural Language Processing, Natural Language Understanding, and Natural Language Generation. LLMs are not only capable of generating coherent text quickly but can also learn from a limited amount of examples.
To create an effective robot, it is crucial to have good reasoning skills and the ability to handle uncertainty and unfamiliar environments. While LLMs have made significant progress in these areas, they still have a limitation known as hallucination. Hallucination occurs when an AI model produces results that are different from what was expected and were not part of its training data. To address this challenge, researchers from Princeton University and Google DeepMind have developed a framework called Know When You Don’t Know (KNOWNO). This framework helps LLM-based robots recognize when they are unsure and request assistance when needed.
KNOWNO utilizes the theory of Conformal Prediction (CP) to handle uncertainties in complex planning scenarios. It provides statistical guarantees for job completion while minimizing the need for human intervention. By applying conformal prediction, KNOWNO can calculate the degree of uncertainty in the predictions made by the LLM-based planner. This uncertainty measurement allows the robot to decide when to seek clarification or additional information, increasing the reliability of its operations.
Experiments conducted by the research team involved real and simulated robot setups with tasks that involved various levels of ambiguity. The tasks included linguistic riddles, numerical uncertainties, human preferences, and spatial uncertainties. Evaluation results have shown that KNOWNO outperforms other methods in terms of efficiency, autonomy, and formal assurances.
KNOWNO is a lightweight approach for modeling uncertainties that can be used with LLMs without the need for model fine-tuning. The key contribution of KNOWNO is its ability to utilize a pre-trained LLM with uncalibrated confidence and language commands to generate potential actions for the robot’s next move. The framework provides theoretical assurances on calibrated confidence for both single-step and multi-step planning problems. In instances where there is doubt, the robot can ask for assistance and complete tasks accurately with a specified level of confidence.
Overall, the KNOWNO framework shows promise in enabling robots to recognize when they are unsure and ask for help in ambiguous situations. It provides a way for robots to handle uncertainties and improve their reliability.