Zhuofeng Li, Yi Lu, Dongfu Jiang, Haoxiang Zhang, Yuyang Bai, Chuan Li, Yu Wang, Shuiwang Ji, Jianwen Xie, Yu Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Gaze input in virtual reality face two main challenges: avoiding false triggering from continuous eye tracking and minimizing the effects of eye jitter, all while maintaining interactive performance. We present GazeBubble, a technique that removes the need for users to precisely fixate on the target by dynamically identifying and selecting the closest object. We created a dataset and collected natural gaze pointing data from 24 participants in various 3D scenes. By analyzing the eye-in-head angle, we initially identified a trigger angular threshold for GazeBubble. Subsequent experiments confirmed that this trigger angular threshold enhances the overall performance of GazeBubble, establishing it as the optimal trigger angular threshold. We also developed Diffuse as an assistive technique for GazeBubble, incorporating various interaction modes, and evaluated its performance in a dense 3D target acquisition task. GazeBubble significantly enhances both performance and user preference in gaze pointing, and improves interaction synergy.

In a multi-stage recommendation system, reranking plays a crucial role in modeling intra-list correlations among items. A key challenge lies in exploring optimal sequences within the combinatorial space of permutations. Recent research follows a two-stage (generator-evaluator) paradigm, where a generator produces multiple feasible sequences, and an evaluator selects the best one. In practice, the generator is typically implemented as an autoregressive model. However, these two-stage methods face two main challenges. First, the separation of the generator and evaluator hinders end-to-end training. Second, autoregressive generators suffer from inference efficiency. In this work, we propose a Unified Generative Efficient Reranking Framework (GReF) to address the two primary challenges. Specifically, we introduce Gen-Reranker, an autoregressive generator featuring a bidirectional encoder and a dynamic autoregressive decoder to generate causal reranking sequences. Subsequently, we pre-train Gen-Reranker on the item exposure order for high-quality parameter initialization. To eliminate the need for the evaluator while integrating sequence-level evaluation during training for end-to-end optimization, we propose post-training the model through Rerank-DPO. Moreover, for efficient autoregressive inference, we introduce ordered multi-token prediction (OMTP), which trains Gen-Reranker to simultaneously generate multiple future items while preserving their order, ensuring practical deployment in real-time recommender systems. Extensive offline experiments demonstrate that GReF outperforms state-of-the-art reranking methods while achieving latency that is nearly comparable to non-autoregressive models. Additionally, GReF has also been deployed in a real-world video app Kuaishou with over 300 million daily active users, significantly improving online recommendation quality.

Text-Attributed Graphs (TAGs) augment graph structures with natural language descriptions, facilitating detailed depictions of data and their interconnections across various real-world settings. However, existing TAG datasets predominantly feature textual information only at the nodes, with edges typically represented by mere binary or categorical attributes. This lack of rich textual edge annotations significantly limits the exploration of contextual relationships between entities, hindering deeper insights into graph-structured data. To address this gap, we introduce Textual-Edge Graphs Datasets and Benchmark (TEG-DB), a comprehensive and diverse collection of benchmark textual-edge datasets featuring rich textual descriptions on nodes and edges. The TEG-DB datasets are large-scale and encompass a wide range of domains, from citation networks to social networks. In addition, we conduct extensive benchmark experiments on TEG-DB to assess the extent to which current techniques, including pre-trained language models, graph neural networks, and their combinations, can utilize textual node and edge information. Our goal is to elicit advancements in textual-edge graph research, specifically in developing methodologies that exploit rich textual node and edge descriptions to enhance graph analysis and provide deeper insights into complex real-world networks. The entire TEG-DB project is publicly accessible as an open-source repository on Github, accessible at https://github.com/Zhuofeng-Li/TEG-Benchmark.

Gaze input is a popular hands-free input method that allows for intuitive and rapid pointing but lacks a confirmation mechanism. This study introduces GazePuffer, an interaction method that combines puffing cheeks with gaze. We explored the design space of mouth gestures, proposed a set of candidate gestures, filtered them through user subjective evaluation, and selected five basic gestures and four variations. We determined the corresponding virtual reality (VR) actions for these gestures through brainstorming. We achieved an accuracy of 93.8% in recognizing the five basic mouth gestures using the built-in sensors of the head-mounted display devices. We compared GazePuffer with two baseline methods in target selection tasks, demonstrating that GazePuffer is on par with Gaze&Pinch in throughput and speed, slightly outperforming Gaze&Dwell. Finally, we showcased the applicability of GazePuffer in real VR interaction tasks, with users generally finding it usable and effortless.

Microservices are becoming more complicated, posing new challenges for traditional performance monitoring solutions. On the one hand, the rapid evolution of microservices places a significant burden on the utilization and maintenance of existing distributed tracing frameworks. On the other hand, complex infrastructure increases the probability of network performance problems and creates more blind spots on the network side. In this paper, we present DeepFlow, a network-centric distributed tracing framework for troubleshooting microservices. DeepFlow provides out-of-the-box tracing via a network-centric tracing plane and implicit context propagation. In addition, it eliminates blind spots in network infrastructure, captures network metrics in a low-cost way, and enhances correlation between different components and layers. We demonstrate analytically and empirically that DeepFlow is capable of locating microservice performance anomalies with negligible overhead. DeepFlow has already identified over 71 critical performance anomalies for more than 26 companies and has been utilized by hundreds of individual developers. Our production evaluations demonstrate that DeepFlow is able to save users hours of instrumentation efforts and reduce troubleshooting time from several hours to just a few minutes.

The growing availability of various wireless access technologies promotes increasing demand for mobile video applications. Stream control transmission protocol (SCTP)-based concurrent multipath transfer (CMT) improves the wireless video delivery performance with its parallel transmission and bandwidth (BW) aggregation features. However, the existing CMT solutions deployed at the transport layer only are not accurate enough due to lower layer uncertainties, such as variations of the wireless channel. In addition, CMT-based video transmission may use excessive BW in comparison with the popular Transmission Control Protocol (TCP)-based flows, which results in unfair sharing of network resources. This paper proposes a novel cross-layer fairness-driven (CL/FD) SCTP-based CMT solution (CMT-CL/FD) to improve video delivery performance, while remaining fair to the competing TCP flows. CMT-CL/FD utilizes a cross-layer approach to monitor and analyze path quality, which includes wireless channel measurements at the data-link layer and rate/BW estimations at the transport layer. Furthermore, an innovative window-based mechanism is applied for flow control to balance delivery fairness and efficiency. Finally, CMT-CL/FD intelligently distributes video data over different paths depending on their estimated quality to mitigate packet reordering and loss, under the constraint of TCP-friendly flow control. Simulation results show how CMT-CL/FD outperforms existing solutions in terms of both video delivery performance and TCP-friendliness.

Concurrent Multipath Transfer (CMT), enabled by Stream Control Transmission Protocol (SCTP), is considered as one preferred data-transport mode due to increased available bandwidth. However, CMT performance degrades seriously in terms of data reordering due to path dissimilarity and frequent packet loss from wireless unreliability. Most relevant solutions follow the packet sequence numbers and thereby focus on strict in-order reception and packet-specific retransmission. Passively adapting to the network conditions, those approaches are not general and well enough responding to the dynamicity of wireless environment. By applying Network Coding (NC) to CMT, this paper proposes a progressive end-to-end solution (CMT-NC) to those problems in heterogeneous wireless networks. CMT-NC is capable of avoiding reordering and compensating lost packets. Further, an innovative group-based transmission management mechanism enhances the robustness and reliability of data transfer. Simulation results show how by using CMT-NC significant improvements in comparison to another state-of-the-art solution are obtained.