Recommender systems aim to leverage user interaction signals to recommend items that users are likely to be interested in. Motivated by the success of Large Language Model (LLMs), generative recommendation (GR) has recently gained increasing attention, typically following a two-stage paradigm: supervised fine-tuning followed by preference alignment. However, aligning generative recommenders with users' personalized preferences remains challenging, as user feedback is inherently heterogeneous and uncertain. Different types of user interaction signals reflect varying levels of intent and should therefore be modeled differently. In this work, we propose Adaptive Mix Preference Optimization (AMPO), an adaptive alignment framework that mixes likelihood and preference objectives with self-calibrated, sample-wise confidence adjustment. AMPO introduces an adaptive target margin that leverages the model's own probability ratio to modulate optimization strength: confident pairs receive full margins that reinforce correct rankings, while uncertain pairs receive reduced margins that prevent overfitting to ambiguous signals. Additionally, AMPO incorporates negative log-likelihood regularization on preferred items to counteract likelihood displacement, a phenomenon where contrastive objectives cause preferred and non-preferred probabilities to collapse simultaneously. Such a design eliminates the need for a reference model, yielding up to 2.5x speedup and 35% memory reduction. Extensive experiments on public benchmarks and a large-scale industrial dataset demonstrate consistent improvements in ranking metrics. Online A/B tests on a major e-commerce platform further confirm statistically significant gains in click-through and conversion rates. The code is available at https://github.com/jumbo-q/ampo.

Novelty recommendation sustains long-term user engagement by exposing users to content that is both relevant and meaningfully different from their recent consumption. In large-scale e-commerce, this requires composing coherent yet non-redundant recommendation lists, a task fundamentally constrained by the relevance-diversity trade-off. Large language models (LLMs) offer a unified generative paradigm for inferring user intent and producing semantically coherent candidates, yet industrial deployment faces two critical challenges: (i) scarce supervision for modeling novelty transitions and diversity-aware list construction, and (ii) reward granularity mismatch, where standard RL assigns coarse sequence-level rewards that fail to capture item-level redundancy and complementarity. We present BALANCE, a hierarchical reasoning-and-generation framework that decomposes novelty recommendation into three structured stages: generating a Novelty Tag for exploration direction, refining an Interest Topic for intent specification, and constructing a Recommendation List for facet coverage. We address data scarcity through a self-reflection pipeline that synthesizes high-quality supervision by integrating real behavior logs with structured rationales. We resolve granularity mismatch through Sequence-Item Policy Optimization (SIPO), which jointly optimizes sequence- and item-level objectives via granularity-aware advantage fusion. Extensive offline experiments and online A/B test on the JD.com recommender system, validate the performance of our method, highlighting its superior novelty and diversity without compromising relevance.

Generative Retrieval (GR) offers a promising paradigm for recommendation through next-token prediction (NTP). However, scaling it to large-scale industrial systems introduces three challenges: (i) within a single request, the identical model inputs may produce inconsistent outputs due to the pagination request mechanism; (ii) the prohibitive cost of encoding long user behavior sequences with multi-token item representations based on semantic IDs, and (iii) aligning the generative policy with nuanced user preference signals. We present GenRec, a preference-oriented generative framework deployed on the JD App https://www.jd.com that addresses above challenges within a single decoder-only architecture. For training objective, we propose Page-wise NTP task, which supervises over an entire interaction page rather than each interacted item individually, providing denser gradient signal and resolving the one-to-many ambiguity of point-wise training. On the prefilling side, an asymmetric linear Token Merger compresses multi-token Semantic IDs in the prompt while preserving full-resolution decoding, reducing input length by ~2× with negligible accuracy loss. To further align outputs with user satisfaction, we introduce GRPO-SR, a reinforcement learning method that pairs Group Relative Policy Optimization with NLL regularization for training stability, and employs Hybrid Rewards combining a dense reward model with a relevance gate to mitigate reward hacking. In month-long online A/B tests serving production traffic, GenRec achieves 9.5% improvement in click count and 8.7% in transaction count over the existing pipeline.

The global aging demographic underscores the imperative for continuous in-home monitoring of the empty-nest elderly, ensuring their safety and well-being. The widespread deployment of Wi-Fi infrastructure has paved the way to monitor the elderly in a non-intrusive and privacy-preserving manner. Numerous studies have explored the potential of utilizing Wi-Fi signals to address urgent life safety concerns such as fall detection and vital sign monitoring. However, apart from these acute safety issues, the early detection of potential disease symptoms and managing the progression of chronic diseases are also crucial for elderly care, which calls for long-term and continuous monitoring of the elderly’s daily routines. Unfortunately, challenges like continuous activity segmentation and location/orientation dependencies have hindered the implementation of a long-term, around-the-clock activity monitoring system for the elderly. This work introduces “WiLife,” a cutting-edge Wi-Fi-based framework for continuous monitoring of the elderly’s spatio-temporal daily status information. Specifically, WiLife adopts a strategy of partitioning living spaces into functional areas and categorizing daily activities into atomic states . By encapsulating daily life status into a unique series of triple unit format: \(\left\langle\textit{Time, Area, State}\right\rangle\) , WiLife is able to offer valuable insights into when, where, and how activities occur. Field implementations spanning 1,080 hours (45 days \(\times\) 24 hours) in real-world home environments highlight WiLife’s exceptional capability in understanding individual living habits and timely detection of irregularities.

Indoor intrusion detection is a critical task for home security. Previous works in intrusion detection suffer from the problems such as blind spots in non-line-of-sight (NLOS) areas, restricted device locations, massive offline training required, and privacy concern. In this article, we design and implement an omnidirectional indoor intrusion detection system, named AudioGuard , using only a pair of speaker and microphone. AudioGuard is able to detect both line-of-sight (LOS) and NLOS intrusions. Our observation of acoustic signal propagation in an indoor environment shows that there exist abundant multipath reflections and human movement introduces Doppler shift in echo signals. We hence capture periodical Doppler shift caused by intruder's walking motion to detect intrusion. Specifically, we first extract the Doppler shift embedded in echo signals, and we then propose a periodicity polarization method to cancel out the impact of the change of radial angle and the distance on periodicity of Doppler shift. Finally, we detect intrusion by measuring periodicity of Doppler shift over time. Extensive experiments show that AudioGuard achieves a miss report rate of 0% and 1.75% for LOS and NLOS intrusion, respectively, and a false alarm rate of 4.17%.

WiFi-based device-free motion tracking systems track persons without requiring them to carry any device. Existing work has explored signal parameters such as time-of-flight (ToF), angle-of-arrival (AoA), and Doppler-frequency-shift (DFS) extracted from WiFi channel state information (CSI) to locate and track people in a room. However, they are not robust due to unreliable estimation of signal parameters. ToF and AoA estimations are not accurate for current standards-compliant WiFi devices that typically have only two antennas and limited channel bandwidth. On the other hand, DFS can be extracted relatively easily on current devices but is susceptible to the high noise level and random phase offset in CSI measurement, which results in a speed-sign-ambiguity problem and renders ambiguous walking speeds. This paper proposes WiTraj, a device-free indoor motion tracking system using commodity WiFi devices. WiTraj improves tracking robustness from three aspects: 1) It significantly improves DFS estimation quality by using the ratio of the CSI from two antennas of each receiver, 2) To better track human walking, it leverages multiple receivers placed at different viewing angles to capture human walking and then intelligently combines the best views to achieve a robust trajectory reconstruction, and, 3) It differentiates walking from in-place activities, which are typically interleaved in daily life, so that non-walking activities do not cause tracking errors. Experiments show that WiTraj can significantly improve tracking accuracy in typical environments compared to existing DFS-based systems. Evaluations across 9 participants and 3 different environments show that the median tracking error$<2.5\%$for typical room-sized trajectories.

Abstractive dialogue summarization aims to convert a long dialogue content into its short form where the salient information is preserved while the redundant pieces are ignored. Different from the well-structured text, such as news and scientific articles, dialogues often consist of utterances coming from two or more interlocutors, where the conversations are often informal, verbose, and repetitive, sprinkled with false-starts, backchanneling, reconfirmations, hesitations, speaker interruptions and the salient information is often scattered across the whole chat. The above properties of conversations make it difficult to directly concentrate on scattered outstanding utterances and thus present new challenges of summarizing dialogues. In this work, rather than directly forcing a summarization system to merely pay more attention to the salient pieces, we propose to explicitly have the model perceive the redundant parts of an input dialogue history during the training phase. To be specific, we design two strategies to construct examples without salient pieces as negative cues. Then, the sequence-to-sequence likelihood loss is cooperated with the unlikelihood objective to drive the model to focus less on the unimportant information and also pay more attention to the salient pieces. Extensive experiments on the benchmark dataset demonstrate that our simple method significantly outperforms the baselines with regard to both semantic matching and factual consistent based metrics. The human evaluation also proves the performance gains.

This paper presents the design and implementation of RT-Fall, a real-time, contactless, low-cost yet accurate indoor fall detection system using the commodity WiFi devices. RT-Fall exploits the phase and amplitude of the fine-grained Channel State Information (CSI) accessible in commodity WiFi devices, and for the first time fulfills the goal of segmenting and detecting the falls automatically in real-time, which allows users to perform daily activities naturally and continuously without wearing any devices on the body. This work makes two key technical contributions. First, we find that the CSI phase difference over two antennas is a more sensitive base signal than amplitude for activity recognition, which can enable very reliable segmentation of fall and fall-like activities. Second, we discover the sharp power profile decline pattern of the fall in the time-frequency domain and further exploit the insight for new feature extraction and accurate fall segmentation/detection. Experimental results in four indoor scenarios demonstrate that RT-fall consistently outperforms the state-of-the-art approach WiFall with 14 percent higher sensitivity and 10 percent higher specificity on average.

Device-free passive indoor localization is playing a critical role in many applications such as elderly care, intrusion detection, smart home, etc. However, existing device-free localization systems either suffer from labor-intensive offline training or require dedicated special-purpose devices. To address the challenges, we present our system named MaTrack, which is implemented on commodity off-the-shelf Intel 5300 Wi-Fi cards. MaTrack proposes a novel Dynamic-MUSIC method to detect the subtle reflection signals from human body and further differentiate them from those reflected signals from static objects (furniture, walls, etc.) to identify the human target's angle for localization. MaTrack does not require any offline training compared to existing signature-based systems and is insensitive to changes in environment. With just two receivers, MaTrack is able to achieve a median localization accuracy below 0.6 m when the human is walking, outperforming the state-of-the-art schemes.