International Journal of Artificial Intelligence and Robotics Research

Classic approaches to content moderation typically apply a rule-based heuristic approach to flag content. While rules are easily customizable and intuitive for humans to interpret, they are inherently fragile and lack the flexibility or robustness needed to moderate the vast amount of undesirable content found online today. Recent advances in deep learning have demonstrated the promise of using highly effective deep neural models to overcome these challenges. However, despite the improved performance, these data-driven models lack transparency and explainability, often leading to mistrust from everyday users and a lack of adoption by many platforms. In this paper, we present Rule By Example (RBE): a novel exemplar-based contrastive learning approach for learning from logical rules for the task of textual content moderation. RBE is capable of providing rule-grounded predictions, allowing for more explainable and customizable predictions compared to typical deep learning-based approaches. We demonstrate that our approach is capable of learning rich rule embedding representations using only a few data examples. Experimental results on three popular hate speech classification datasets show that RBE is able to outperform state-of-the-art deep learning classifiers as well as the use of rules in both supervised and unsupervised settings while providing explainable model predictions via rule-grounding.

Despite the impressive performance across numerous tasks, Large Language Models (LLMs) often fail in solving simple decision-making tasks due to the misalignment of the knowledge in LLMs with environments. On the contrary, Reinforcement Learning (RL) agents learn policies from scratch, which makes them always align with environments but difficult to incorporate prior knowledge for efficient explorations. To narrow the gap, we propose TWOSOME, a novel general online framework that deploys LLMs as decision-making agents to efficiently interact and align with embodied environments via RL without requiring any prepared datasets or prior knowledge of the environments. First, we query the joint probabilities of each valid action with LLMs to form behavior policies. Then, to enhance the stability and robustness of the policies, we propose two normalization methods and summarize four prompt design principles. Finally, we design a novel parameter-efficient training architecture where the actor and critic share one frozen LLM equipped with LOw-Rank Adapters (LoRA) updated by PPO. We conduct extensive experiments to evaluate TWOSOME. (i) TWOSOME exhibits significantly better sample efficiency and performance compared to the conventional RL method, PPO, and prompt tuning method, SayCan, in both classical decision-making environment, Overcooked, and simulated household environment, VirtualHome. (ii) Benefiting from LLMs’ open-vocabulary feature, TWOSOME shows superior generalization ability to unseen tasks. (iii) Under our framework, there is no significant loss of the LLMs’ original ability during online PPO finetuning.^a

ChatGPT has demonstrated its potential as a surrogate knowledge graph. Trained on extensive data sources, including open-access publications, peer-reviewed research articles, and biomedical websites, ChatGPT extracted information on gene relationships and biological pathways so that it can be used to predict them. However, a major challenge is model hallucination, that is, high false positive rates. To assess and address this challenge, we systematically evaluated ChatGPT’s capacity for predicting gene relationships using GPT-3.5-turbo, GPT-4, and GPT-4o. Benchmarking against the KEGG Pathway Database as the ground truth, we experimented with diverse prompting strategies, targeting gene relationships of activation, inhibition, and phosphorylation. We introduced an innovative iterative prompt refinement technique. By assessing prompt efficacy using metrics such as F-1 score, precision, and recall, GPT-4 suggested improved prompts. A refined prompt, which combines a specialized role with explanatory text, significantly enhanced the performance. Going beyond pairwise gene relationships, we also deciphered complex gene interplays, such as gene interaction chains and pathways pertinent to diseases such as non-small cell lung cancer. Direct prompts showed limited success, but “least-to-most” prompting exhibited significant potentials for such network constructions. The methods in this study may be used for other bioinformatics prediction problems.

The integration of robotics into domestic environments poses significant challenges due to the dynamic and varied nature of these settings. This paper introduces a new framework that combines vision-guided object recognition with adaptive grasping policies learned from human demonstrations. By harnessing computer vision technology, our system employs deep learning algorithms, particularly Convolutional Neural Networks (CNNs), to precisely detect and classify household objects. Simultaneously, the system uses imitation learning to refine grasping policies, enabling the robotic manipulator to dynamically adapt to new target objects. We validated our framework through a series of experimental setups that simulate typical kitchen tasks, such as manipulating utensils and preparing ingredients. These tasks, which primarily involve picking up and placing objects, served as practical tests for our system. The results demonstrate the system’s ability to effectively recognize a broad array of objects and adapt its grasping policies, thereby enhancing operational efficiency.