Wei Xu

Yihan Hu, Kun Li, Pingyuan Liang et al.

This paper presents our 2nd place solution for the NuPlan Challenge 2023. Autonomous driving in real-world scenarios is highly complex and uncertain. Achieving safe planning in the complex multimodal scenarios is a highly challenging task. Our approach, Imitation with Spatial-Temporal Heatmap, adopts the learning form of behavior cloning, innovatively predicts the future multimodal states with a heatmap representation, and uses trajectory refinement techniques to ensure final safety. The experiment shows that our method effectively balances the vehicle's progress and safety, generating safe and comfortable trajectories. In the NuPlan competition, we achieved the second highest overall score, while obtained the best scores in the ego progress and comfort metrics.

Wei Xu, Zaifeng Gao

Research and application have used human-AI teaming (HAT) as a new paradigm to develop AI systems. HAT recognizes that AI will function as a teammate instead of simply a tool in collaboration with humans. Effective human-AI teams need to be capable of taking advantage of the unique abilities of both humans and AI while overcoming the known challenges and limitations of each member, augmenting human capabilities, and raising joint performance beyond that of either entity. The National AI Research and Strategic Plan 2023 update has recognized that research programs focusing primarily on the independent performance of AI systems generally fail to consider the functionality that AI must provide within the context of dynamic, adaptive, and collaborative teams and calls for further research on human-AI teaming and collaboration. However, there has been debate about whether AI can work as a teammate with humans. The primary concern is that adopting the "teaming" paradigm contradicts the human-centered AI (HCAI) approach, resulting in humans losing control of AI systems. This article further analyzes the HAT paradigm and the debates. Specifically, we elaborate on our proposed conceptual framework of human-AI joint cognitive systems (HAIJCS) and apply it to represent HAT under the HCAI umbrella. We believe that HAIJCS may help adopt HAI while enabling HCAI. The implications and future work for HAIJCS are also discussed. Insights: AI has led to the emergence of a new form of human-machine relationship: human-AI teaming (HAT), a paradigmatic shift in human-AI systems; We must follow a human-centered AI (HCAI) approach when applying HAT as a new design paradigm; We propose a conceptual framework of human-AI joint cognitive systems (HAIJCS) to represent and implement HAT for developing effective human-AI teaming

Yang Wu, Huayi Zhang, Yizheng Jiao et al.

Instruction tuning has underscored the significant potential of large language models (LLMs) in producing more human controllable and effective outputs in various domains. In this work, we focus on the data selection problem for task-specific instruction tuning of LLMs. Prevailing methods primarily rely on the crafted similarity metrics to select training data that aligns with the test data distribution. The goal is to minimize instruction tuning loss on the test data, ultimately improving performance on the target task. However, it has been widely observed that instruction tuning loss (i.e., cross-entropy loss for next token prediction) in LLMs often fails to exhibit a monotonic relationship with actual task performance. This misalignment undermines the effectiveness of current data selection methods for task-specific instruction tuning. To address this issue, we introduce ROSE, a novel Reward-Oriented inStruction data sElection method which leverages pairwise preference loss as a reward signal to optimize data selection for task-specific instruction tuning. Specifically, ROSE adapts an influence formulation to approximate the influence of training data points relative to a few-shot preference validation set to select the most task-related training data points. Experimental results show that by selecting just 5\% of the training data using ROSE, our approach can achieve competitive results compared to fine-tuning with the full training dataset, and it surpasses other state-of-the-art data selection methods for task-specific instruction tuning. Our qualitative analysis further confirms the robust generalizability of our method across multiple benchmark datasets and diverse model architectures.