Abstract:The reward model (RM) that represents human preferences plays a crucial role in optimizing the outputs of large language models (LLMs), e.g., through reinforcement learning from human feedback (RLHF) or rejection sampling. However, a long challenge for RM is its uncertain reliability, i.e., LLM outputs with higher rewards may not align with actual human preferences. Currently, there is a lack of a convincing metric to quantify the reliability of RMs. To bridge this gap, we propose the \textit{\underline{R}eliable at \underline{$\eta$}} (RETA) metric, which directly measures the reliability of an RM by evaluating the average quality (scored by an oracle) of the top $\eta$ quantile responses assessed by an RM. On top of RETA, we present an integrated benchmarking pipeline that allows anyone to evaluate their own RM without incurring additional Oracle labeling costs. Extensive experimental studies demonstrate the superior stability of RETA metric, providing solid evaluations of the reliability of various publicly available and proprietary RMs. When dealing with an unreliable RM, we can use the RETA metric to identify the optimal quantile from which to select the responses.
Abstract:Recent E-commerce applications benefit from the growth of deep learning techniques. However, we notice that many works attempt to maximize business objectives by closely matching offline labels which follow the supervised learning paradigm. This results in models obtain high offline performance in terms of Area Under Curve (AUC) and Normalized Discounted Cumulative Gain (NDCG), but cannot consistently increase the revenue metrics such as purchases amount of users. Towards the issues, we build a simulated search engine AESim that can properly give feedback by a well-trained discriminator for generated pages, as a dynamic dataset. Different from previous simulation platforms which lose connection with the real world, ours depends on the real data in AliExpress Search: we use adversarial learning to generate virtual users and use Generative Adversarial Imitation Learning (GAIL) to capture behavior patterns of users. Our experiments also show AESim can better reflect the online performance of ranking models than classic ranking metrics, implying AESim can play a surrogate of AliExpress Search and evaluate models without going online.
Abstract:Modern large-scale knowledge graphs, such as DBpedia, are datasets which require large computational resources to serve and process. Moreover, they often have longer release cycles, which leads to outdated information in those graphs. In this paper, we present DBpedia on Demand -- a system which serves DBpedia resources on demand without the need to materialize and store the entire graph, and which even provides limited querying functionality.