The rapidly evolving landscape of products, surfaces, policies, and regulations poses significant challenges for deploying state-of-the-art recommendation models at industry scale, primarily due to data fragmentation across domains and escalating infrastructure costs that hinder sustained quality improvements.

Recent advances in recommendation scaling laws have led to foundation models of unprecedented complexity. While these models offer superior performance, their computational demands make real-time serving impractical, often forcing practitioners to rely on knowledge distillation—compromising serving quality for efficiency. To address this challenge, we present SOLARIS (Speculative Offloading of Latent-bAsed Representation for Inference Scaling), a novel framework inspired by speculative decoding. SOLARIS proactively precomputes user-item interaction embeddings by predicting which user-item pairs are likely to appear in future requests, and asynchronously generating their foundation model representations ahead of time. This approach decouples the costly foundation model inference from the latency-critical serving path, enabling real-time knowledge transfer from models previously considered too expensive for online use. Deployed across Meta's advertising system serving billions of daily requests, SOLARIS achieves 0.67% revenue-driving top-line metrics gain, demonstrating its effectiveness at scale.

The volume of ads ranked impacts the performance of ad ranking systems.To enhance efficiency, multi-stage ranking systems are widely studied in academia and adopted across industry.However, as large-scale deep learning recommendation models gain prevalence, resource constraints-especially CPU and GPU limitationshave become a significant bottleneck.These constraints can hinder model iteration and lead to incomplete ranking, causing regressions in user experience and ad performance.To address these issues, we analyzed ad ranking metrics and found that ad rankings for individual users remain relatively stable over short periods.Based on this insight, we introduce Negative Exclusion Filtering, a framework that optimizes the balance between ranked ad volume and computing resources.By skipping re-ranking of consistently low-ranked ads for each user request, it reduces computing cost in large-scale social media environments.

Carrier aggregation (CA) is an important component technology in 5G and beyond. It aggregates multiple spectrum fragments to serve a mobile device. However, the current CA suffers under both high mobility and increased spectrum space. The limitations are rooted in its sequential, cell-by-cell operations. In this work, we propose CA++, which departs from the current paradigm and explores a group-based design scheme. We thus propose new algorithms that enable concurrent channel inference by measuring one or few cells but inferring all, while minimizing measurement cost via set cover approximations. Our evaluations have confirmed the effectiveness of CA++. Our solution can also be adapted to fit in the current 5G OFDM PHY and the 3GPP framework.

Extreme mobility becomes a norm rather than an exception with emergent high-speed rails, drones, industrial IoT, and many more. However, 4G/5G mobility management is not always reliable in extreme mobility, with non-negligible failures and policy conflicts. The root cause is that, existing mobility management is primarily based on wireless signal strength. While reasonable in static and low mobility, it is vulnerable to dramatic wireless dynamics from extreme mobility in triggering, decision, and execution. We deviseREM, Reliable Extreme Mobility management for beyond 5G cellular networks while maintaining backward compatibility to 4G/5G.REMshifts to movement-based mobility management in the delay-Doppler domain. Its signaling overlay relaxes feedback via cross-band estimation, simplifies policies with provable conflict freedom, and stabilizes signaling via scheduling-based OTFS modulation. Our evaluation with operational high-speed rail datasets shows that,REMreduces failures comparable to static and low mobility, with low signaling and latency cost.REMreduces the network failures by up to an order of magnitude, eliminates policy conflicts, and improves application performance by 31.8% - 88.3% compared to legacy 4G/5G.

We devise a simple, provably effective, and readily usable deterrence against intelligent, unknown DDoS threats: Demotivate adversaries to launch attacks via multi-hop traffic divergence. This new strategy is motivated by the fact that existing defenses almost always lag behind numerous emerging DDoS threats and evolving intelligent attack strategies. The root cause is if adversaries are smart and adaptive, no single-hop defenses (including optimal ones) can perfectly differentiate unknown DDoS and legitimate traffic. Instead, we formulate intelligent DDoS as a game between attackers and defenders, and prove how multi-hop traffic divergence helps bypass this dilemma by reversing the asymmetry between attackers and defenders. This insight results in EID, an Economical Intelligent DDoS Demotivation protocol. EID combines local weak (yet divergent) filters to provably null attack gains without knowing exploited vulnerabilities or attack strategies. It incentivizes multi-hop defenders to cooperate with boosted local service availability. EID is resilient to traffic dynamics and manipulations. It is readily deployable with random-drop filters in real networks today. Our experiments over a 49.8 TB dataset from a department at the Tsinghua campus network validate EID's viability against rational and irrational DDoS with negligible costs.

In cellular networks, cell selection plays a critical role in providing and maintaining ubiquitous radio access. It follows standardized procedures with operator-specific polices pre-configured by tunable parameters. These parameters specify the criteria to determine whether and how to select new serving cell(s), thus impacting access quality and user experience. Recent studies reveal that today’s cell selection fails to offer good performance as it can. This is because it is configured for seamless connectivity, and thus performance is offered at "best effort". In this work, we attempt to re-configure these parameters by taking performance into consideration. We first conduct a measurement study in one big city in the US to demonstrate that reconfiguration indeed helps improve the overall performance, without compromising connectivity. This implies that 4G/5G networks are capable of offering better performance but such potentials are under-utilized in practice. We further explore proactive reconfiguration to prevent such unnecessary performance losses. We examine technical challenges, factors and even limitations to reconfigure cell selection in a standard-compatible manner, and finally devise a simple reconfiguration algorithm based on profiling and heuristic searching to efficiently pursue promising performance gains. The evaluation over AT&T and T-Mobile in two US cities has validated its effectiveness. Performance gains outweigh losses. Reconfiguration boosts data speed in more than 30% of instances, which exceeds the ratio of losses by at least 16%; The median speed gain is at least 89.1% (up to 217 fold).

This paper reports our five-year lessons of developing and using MobileInsight, an open-source community tool to enable software-defined full-stack, runtime mobile network analytics inside our phones. We present how MobileInsight evolves from a simple monitor to a community toolset with cross-layer analytics, energy-efficient real-time user-plane analytics, and extensible user-friendly analytics at the control and user planes. These features are enabled by various novel techniques, including cross-layer state machine tracking, missing data inference, and domain-specific cross-layer sampling. Their powerfulness is exemplified with a 5-year longitudinal study of operational mobile network latency using a 6.4TB dataset with 6.1 billion over-the-air messages. We further share lessons and insights of using MobileInsight by the community, as well as our visions of MobileInsight's past, present, and future.

In this paper, we propose iCellSpeed, an on-device solution to increase data access speed by substantiating unrealized performance potentials. We find that performance potentials are missed in today's mobile networks, as the data speed a user device gets is much lower than what the device could get. The issue is rooted in the current cell selection practice, which misses good candidate cells that offer faster access speed, thus under-utilizing the available capabilities in mobile networks. We design iCellSpeed to facilitate network-controlled cell selection with proactive device-side assistance towards more desirable cells. Our evaluation over AT&T and Verizon confirms its effectiveness. iCellSpeed increases data access speed by more than 10 Mbps at 79% of test locations (> 25Mbps at 29% of locations, up to 80.6 Mbps). It doubles access speed at 62.5% of locations with the gain up to 28.4x. Datasets are available at [9].

Extreme mobility has become a norm rather than an exception. However, 4G/5G mobility management is not always reliable in extreme mobility, with non-negligible failures and policy conflicts. The root cause is that, existing mobility management is primarily based on wireless signal strength. While reasonable in static and low mobility, it is vulnerable to dramatic wireless dynamics from extreme mobility in triggering, decision, and execution. We devise REM, Reliable Extreme Mobility management for 4G, 5G, and beyond. REM shifts to movement-based mobility management in the delay-Doppler domain. Its signaling overlay relaxes feedback via cross-band estimation, simplifies policies with provable conflict freedom, and stabilizes signaling via scheduling-based OTFS modulation. Our evaluation with operational high-speed rail datasets shows that, REM reduces failures comparable to static and low mobility, with low signaling and latency cost.

Multi-carrier cellular access dynamically selects a preferred wireless carrier by leveraging the availability and diversity of multiple carrier networks at a location. It offers an alternative to the dominant single-carrier paradigm, and shows early signs of success through the operational Project Fi by Google. In this paper, we study the important, yet largely unexplored, problem of inter-carrier switching for multi-carrier access. We show that policy conflicts can arise between inter- and intra-carrier switching, resulting in oscillations among carriers in the worst case akin to BGP looping. We derive the conditions under which such oscillations occur for three categories of popular policy, and validate them with Project Fi whenever possible. We provide practical guidelines to ensure loop-freedom and assess them via trace-driven emulations.

In this paper, we call for a paradigm shift away from the wireless-access focused research efforts on 5G networked systems. We believe that the architectural limitations should share equal blame on issues of performance, reliability, and security. We thus identify architectural weakness on both sides of the mobile clients and the 4G network infrastructure. Our recent findings show that, contrary to commonly held perceptions, many design and operational issues arise not due to poor wireless link qualities. Instead, they are rooted in such architectural downsides. To address these issues, we further propose a new approach of enabling automated intelligence inside the 4G/5G network systems. We next describe our ongoing efforts along two dimensions: empowering date-driven smart clients and constructing verifiable network infrastructure. We report some early results and discuss possible next steps.