AlignPxtr: Aligning Predicted Behavior Distributions for Bias-Free Video Recommendations

In video recommendation systems, user behaviors such as watch time, likes, and follows are commonly used to infer user interest. However, these behaviors are influenced by various biases, including duration bias, demographic biases, and content category biases, which obscure true user preferences. In this paper, we hypothesize that biases and user interest are independent of each other. Based on this assumption, we propose a novel method that aligns predicted behavior distributions across different bias conditions using quantile mapping, theoretically guaranteeing zero mutual information between bias variables and the true user interest. By explicitly modeling the conditional distributions of user behaviors under different biases and mapping these behaviors to quantiles, we effectively decouple user interest from the confounding effects of various biases. Our approach uniquely handles both continuous signals (e.g., watch time) and discrete signals (e.g., likes, comments), while simultaneously addressing multiple bias dimensions. Additionally, we introduce a computationally efficient mean alignment alternative technique for practical real-time inference in large-scale systems. We validate our method through online A/B testing on two major video platforms: Kuaishou Lite and Kuaishou. The results demonstrate significant improvements in user engagement and retention, with \textbf{cumulative lifts of 0.267\% and 0.115\% in active days, and 1.102\% and 0.131\% in average app usage time}, respectively. The results demonstrate that our approach consistently achieves significant improvements in long-term user retention and substantial gains in average app usage time across different platforms. Our core code will be publised at https://github.com/justopit/CQE.

Conditional Quantile Estimation for Uncertain Watch Time in Short-Video Recommendation

Within the domain of short video recommendation, predicting users' watch time is a critical but challenging task. Prevailing deterministic solutions obtain accurate debiased statistical models, yet they neglect the intrinsic uncertainty inherent in user environments. In our observation, we found that this uncertainty could potentially limit these methods' accuracy in watch-time prediction on our online platform, despite that we have employed numerous features and complex network architectures. Consequently, we believe that a better solution is to model the conditional distribution of this uncertain watch time. In this paper, we introduce a novel estimation technique -- Conditional Quantile Estimation (CQE), which utilizes quantile regression to capture the nuanced distribution of watch time. The learned distribution accounts for the stochastic nature of users, thereby it provides a more accurate and robust estimation. In addition, we also design several strategies to enhance the quantile prediction including conditional expectation, conservative estimation, and dynamic quantile combination. We verify the effectiveness of our method through extensive offline evaluations using public datasets as well as deployment in a real-world video application with over 300 million daily active users.

What Makes for Good Visual Instructions? Synthesizing Complex Visual Reasoning Instructions for Visual Instruction Tuning

Visual instruction tuning is an essential approach to improving the zero-shot generalization capability of Multi-modal Large Language Models (MLLMs). A surge of visual instruction datasets with various focuses and characteristics have been proposed recently, enabling MLLMs to achieve surprising results on evaluation benchmarks. To develop more capable MLLMs, in this paper, we aim to investigate a more fundamental question: ``what makes for good visual instructions?''. By conducting a comprehensive empirical study, we find that instructions focused on complex visual reasoning tasks are particularly effective in improving the performance of MLLMs on evaluation benchmarks. Building upon this finding, we design a systematic approach to automatically creating high-quality complex visual reasoning instructions. Our approach employs a synthesis-complication-reformulation paradigm, leveraging multiple stages to gradually increase the complexity of the instructions while guaranteeing quality. Based on this approach, we create the synthetic visual reasoning instruction dataset consisting of 32K examples, namely ComVint, and fine-tune four MLLMs on it. Experimental results demonstrate that our dataset consistently enhances the performance of all the compared MLLMs, e.g., improving the performance of MiniGPT-4 and BLIP-2 on MME-Cognition by 32.6% and 28.8%, respectively. Our code and data are publicly available at the link: https://github.com/RUCAIBox/ComVint.

PREFER: Prompt Ensemble Learning via Feedback-Reflect-Refine

As an effective tool for eliciting the power of Large Language Models (LLMs), prompting has recently demonstrated unprecedented abilities across a variety of complex tasks. To further improve the performance, prompt ensemble has attracted substantial interest for tackling the hallucination and instability of LLMs. However, existing methods usually adopt a two-stage paradigm, which requires a pre-prepared set of prompts with substantial manual effort, and is unable to perform directed optimization for different weak learners. In this paper, we propose a simple, universal, and automatic method named PREFER (Pompt Ensemble learning via Feedback-Reflect-Refine) to address the stated limitations. Specifically, given the fact that weak learners are supposed to focus on hard examples during boosting, PREFER builds a feedback mechanism for reflecting on the inadequacies of existing weak learners. Based on this, the LLM is required to automatically synthesize new prompts for iterative refinement. Moreover, to enhance stability of the prompt effect evaluation, we propose a novel prompt bagging method involving forward and backward thinking, which is superior to majority voting and is beneficial for both feedback and weight calculation in boosting. Extensive experiments demonstrate that our PREFER achieves state-of-the-art performance in multiple types of tasks by a significant margin. We have made our code publicly available.