In self-consuming generative models that train on their own outputs, alignment with user preferences becomes a recursive rather than one-time process. In this paper, we provide the first formal foundation for analyzing the long-term effects of such recursive retraining on alignment. Under a two-stage curation mechanism based on the Bradley–Terry (BT) model, we model alignment as an interaction between two factions: the Model Owner, who filters which outputs should be learned by the model, and the Public User, who determines which outputs are ultimately shared and retained through interactions with the model. Our analysis reveals three structural convergence regimes: consensus collapse, compromise on shared optima, and asymmetric refinement, depending on the degree of preference alignment. We prove a fundamental impossibility theorem: no recursive BT-based curation mechanism can simultaneously preserve diversity, ensure symmetric influence, and eliminate dependence on initialization. Framing the process as dynamic social choice, we show that alignment is not a static goal but an evolving equilibrium shaped by power asymmetries and path dependence.

David H. Yang, Yuxuan Zhu, Mohammad Mohammadi Amiri, Keerthiram Murugesan, Tejaswini Pedapati, Subhajit Chaudhury, Pin-Yu Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We introduce Time-Conditioned Contraction Matching (TCCM), a novel method for semi-supervised anomaly detection in tabular data. TCCM is inspired by flow matching, a recent generative modeling framework that learns velocity fields between probability distributions and has shown strong performance compared to diffusion models and generative adversarial networks. Instead of directly applying flow matching as originally formulated, TCCM builds on its core idea—learning velocity fields between distributions—but simplifies the framework by predicting a time-conditioned contraction vector toward a fixed target (the origin) at each sampled time step. This design offers three key advantages: (1) a lightweight and scalable training objective that removes the need for solving ordinary differential equations during training and inference; (2) an efficient scoring strategy called one time-step deviation, which quantifies deviation from expected contraction behavior in a single forward pass, addressing the inference bottleneck of existing continuous-time models such as DTE (a diffusion-based model with leading anomaly detection accuracy but heavy inference cost); and (3) explainability and provable robustness, as the learned velocity field operates directly in input space, making the anomaly score inherently feature-wise attributable; moreover, the score function is Lipschitz-continuous with respect to the input, providing theoretical guarantees under small perturbations. Extensive experiments on the ADBench benchmark show that TCCM strikes a favorable balance between detection accuracy and inference cost, outperforming state-of-the-art methods—especially on high-dimensional and large-scale datasets. The source code is provided at https://github.com/ZhongLIFR/TCCM-NIPS.

In federated learning (FL), incentivizing contributions of training resources (e.g., data, compute) from potentially competitive clients is crucial. Existing incentive mechanisms often distribute post-training monetary rewards, which suffer from practical challenges of timeliness and feasibility of the rewards. Rewarding the clients after the completion of training may incentivize them to abort the collaboration, and monetizing the contribution is challenging in practice. To address these problems, we propose an incentive-aware algorithm that offers differentiated training-time model rewards for each client at each FL iteration. We theoretically prove that such a $\textit{local}$ design ensures the $\textit{global}$ objective of client incentivization. Through theoretical analyses, we further identify the issue of error propagation in model rewards and thus propose a stochastic reference-model recovery strategy to ensure theoretically that all the clients eventually obtain the optimal model in the limit. We perform extensive experiments to demonstrate the superior incentivizing performance of our method compared to existing baselines.

We study valuing the data of a data owner/seller for a data seeker/buyer. Data valuation is often carried out for a specific task assuming a particular utility metric, such as test accuracy on a validation set, that may not exist in practice. In this work, we focus on task-agnostic data valuation without any validation requirements. The data buyer has access to a limited amount of data (which could be publicly available) and seeks more data samples from a data seller. We formulate the problem as estimating the differences in the statistical properties of the data at the seller with respect to the baseline data available at the buyer. We capture these statistical differences through second moment by measuring diversity and relevance of the seller’s data for the buyer; we estimate these measures through queries to the seller without requesting the raw data. We design the queries with the proposed approach so that the seller is blind to the buyer’s raw data and has no knowledge to fabricate responses to the queries to obtain a desired outcome of the diversity and relevance trade-off. We will show through extensive experiments on real tabular and image datasets that the proposed estimates capture the diversity and relevance of the seller’s data for the buyer.

Communication efficiency is of importance for wireless federated learning systems. In this paper, we propose a communication-efficient strategy for federated learning over multiple-input multiple-output (MIMO) multiple access channels (MACs). The proposed strategy comprises two components. When sending a locally computed gradient, each device compresses a high dimensional local gradient to multiple lower-dimensional gradient vectors using block sparsification. When receiving a superposition of the compressed local gradients via a MIMO-MAC, a parameter server (PS) performs a joint MIMO detection and the sparse local-gradient recovery. Inspired by the turbo decoding principle, our joint detection-and-recovery algorithm accurately recovers the high-dimensional local gradients by iteratively exchanging their beliefs for MIMO detection and sparse local gradient recovery outputs. We then analyze the reconstruction error of the proposed algorithm and its impact on the convergence rate of federated learning. From simulations, our gradient compression and joint detection-and-recovery methods diminish the communication cost significantly while achieving identical classification accuracy for the case without any compression.

We study federated learning (FL), where power-limited wireless devices utilize their local datasets to collaboratively train a global model with the help of a remote parameter server (PS). The PS has access to the global model and shares it with the devices for local training using their datasets, and the devices return the result of their local updates to the PS to update the global model. The algorithm continues until the convergence of the global model. This framework requires downlink transmission from the PS to the devices and uplink transmission from the devices to the PS. The goal of this study is to investigate the impact of the bandwidth-limited shared wireless medium on the performance of FL with a focus on the downlink. To this end, the downlink and uplink channels are modeled as fading broadcast and multiple access channels, respectively, both with limited bandwidth. For downlink transmission, we first introduce a digital approach, where a quantization technique is employed at the PS followed by a capacity-achieving channel code to transmit the global model update over the wireless broadcast channel at a common rate such that all the devices can decode it. Next, we propose analog downlink transmission, where the global model is broadcast by the PS in an uncoded manner. We consider analog transmission over the uplink in both cases, since its superiority over digital transmission for uplink has been well studied in the literature. We further analyze the convergence behavior of the proposed analog transmission approach over the downlink assuming that the uplink transmission is error-free. Numerical experiments show that the analog downlink approach provides significant improvement over the digital one with a more notable improvement when the data distribution across the devices is not independent and identically distributed. The experimental results corroborate the convergence analysis, and show that a smaller number of local iterations should be used when the data distribution is more biased, and also when the devices have a better estimate of the global model in the analog downlink approach.

We study federated edge learning (FEEL), where wireless edge devices, each with its own dataset, learn a global model collaboratively with the help of a wireless access point acting as the parameter server (PS). At each iteration, wireless devices perform local updates using their local data and the most recent global model received from the PS, and send their local updates to the PS over a wireless fading multiple access channel (MAC). The PS then updates the global model according to the signal received over the wireless MAC, and shares it with the devices. Motivated by the additive nature of the wireless MAC, we propose an analog `over-the-air' aggregation scheme, in which the devices transmit their local updates in an uncoded fashion. However, unlike recent literature on over-the-air FEEL, here we assume that the devices do not have channel state information (CSI), while the PS has imperfect CSI. On the other hand, the PS is equipped with multiple antennas to alleviate the destructive effect of the channel, exacerbated due to the lack of perfect CSI. We design a receive beamforming scheme at the PS, and show that it can compensate for the lack of perfect CSI when the PS has a sufficient number of antennas. We also derive the convergence rate of the proposed algorithm highlighting the impact of the lack of perfect CSI, as well as the number of PS antennas. Both the experimental results and the convergence analysis illustrate the performance improvement of the proposed algorithm with the number of PS antennas, where the wireless fading MAC becomes deterministic despite the lack of perfect CSI when the PS has a sufficiently large number of antennas.

We study federated learning (FL) at the wireless edge, where power-limited devices with local datasets collaboratively train a joint model with the help of a remote parameter server (PS). We assume that the devices are connected to the PS through a bandwidth-limited shared wireless channel. At each iteration of FL, a subset of the devices are scheduled to transmit their local model updates to the PS over orthogonal channel resources, while each participating device must compress its model update to accommodate to its link capacity. We design novel scheduling and resource allocation policies that decide on the subset of the devices to transmit at each round, and how the resources should be allocated among the participating devices, not only based on their channel conditions, but also on the significance of their local model updates. We then establish convergence of a wireless FL algorithm with device scheduling, where devices have limited capacity to convey their messages. The results of numerical experiments show that the proposed scheduling policy, based on both the channel conditions and the significance of the local model updates, provides a better long-term performance than scheduling policies based only on either of the two metrics individually. Furthermore, we observe that when the data is independent and identically distributed (i.i.d.) across devices, selecting a single device at each round provides the best performance, while when the data distribution is non-i.i.d., scheduling multiple devices at each round improves the performance. This observation is verified by the convergence result, which shows that the number of scheduled devices should increase for a less diverse and more biased data distribution.

Federated learning is a privacy-preserving approach to train a global model at a central server by collaborating with wireless devices, each with its own local training data set. In this paper, we present a compressive sensing approach for federated learning over massive multiple-input multiple-output communication systems in which the central server equipped with a massive antenna array communicates with the wireless devices. One major challenge in system design is to reconstruct local gradient vectors accurately at the central server, which are computed-and-sent from the wireless devices. To overcome this challenge, we first establish a transmission strategy to construct sparse transmitted signals from the local gradient vectors at the devices. We then propose a compressive sensing algorithm enabling the server to iteratively find the linear minimum-mean-square-error (LMMSE) estimate of the transmitted signal by exploiting its sparsity. We also derive an analytical threshold for the residual error at each iteration, to design the stopping criterion of the proposed algorithm. We show that for a sparse transmitted signal, the proposed algorithm requires less computationally complexity than LMMSE. Simulation results demonstrate that the presented approach outperforms conventional linear beamforming approaches and reduces the performance gap between federated learning and centralized learning with perfect reconstruction.

We study federated learning (FL) at the wireless edge, where power-limited devices with local datasets train a joint model with the help of a remote parameter server (PS). We assume that the devices are connected to the PS through a bandwidth-limited shared wireless channel. At each iteration of FL, a subset of the devices are scheduled to transmit their local model updates to the PS over orthogonal channel resources. We design novel scheduling policies, that decide on the subset of devices to transmit at each round not only based on their channel conditions, but also on the significance of their local model updates. Numerical results show that the proposed scheduling policy provides a better long-term performance than scheduling policies based only on either of the two metrics individually. We also observe that when the data is independent and identically distributed (i.i.d.) across devices, selecting a single device at each round provides the best performance, while when the data distribution is non-i.i.d., more devices should be scheduled.

We study federated machine learning at the wireless network edge, where limited power wireless devices, each with its own dataset, build a joint model with the help of a remote parameter server (PS). We consider a bandwidth-limited fading multiple access channel (MAC) from the wireless devices to the PS, and propose various techniques to implement distributed stochastic gradient descent (DSGD) over this shared noisy wireless channel. We first propose a digital DSGD (D-DSGD) scheme, in which one device is selected opportunistically for transmission at each iteration based on the channel conditions; the scheduled device quantizes its gradient estimate to a finite number of bits imposed by the channel condition, and transmits these bits to the PS in a reliable manner. Next, motivated by the additive nature of the wireless MAC, we propose a novel analog communication scheme, referred to as the compressed analog DSGD (CA-DSGD), where the devices first sparsify their gradient estimates while accumulating error from previous iterations, and project the resultant sparse vector into a low-dimensional vector for bandwidth reduction. We also design a power allocation scheme to align the received gradient vectors at the PS in an efficient manner. Numerical results show that D-DSGD outperforms other digital approaches in the literature; however, in general the proposed CA-DSGD algorithm converges faster than the D-DSGD scheme, and reaches a higher level of accuracy. We have observed that the gap between the analog and digital schemes increases when the datasets of devices are not independent and identically distributed (i.i.d.). Furthermore, the performance of the CA-DSGD scheme is shown to be robust against imperfect channel state information (CSI) at the devices. Overall these results show clear advantages for the proposed analog over-the-air DSGD scheme, which suggests that learning and communication algorithms should be designed jointly to achieve the best end-to-end performance in machine learning applications at the wireless edge.

We study downlink beamforming in a single-cell network with a multi-antenna base station serving cache-enabled users. Assuming a library of files with a common rate, we formulate the minimum transmit power with proactive caching and coded delivery as a non-convex optimization problem. While this multiple multicast problem can be efficiently solved by successive convex approximation (SCA), the complexity of the problem grows exponentially with the number of subfiles delivered to each user in each time slot, which itself grows exponentially with the number of users. We introduce a low-complexity alternative through time-sharing that limits the number of subfiles received by a user in each time slot. We then consider the joint design of beamforming and content delivery with sparsity constraints to limit the number of subfiles received by a user in each time slot. Numerical simulations show that the low-complexity scheme has only a small performance gap to that obtained by solving the joint problem with sparsity constraints, and outperforms state-of-the-art results at all signal-to-noise ratio (SNR) and rate values with a sufficient number of transmit antennas. A lower bound on the achievable degrees-of-freedom (DoF) of the low-complexity scheme is derived to characterize its performance in the high SNR regime.

We study scheduling of computation tasks across n workers in a large scale distributed learning problem. Computation speeds of the workers are assumed to be heterogeneous and unknown to the master, and redundant computations are assigned to the workers in order to tolerate straggling workers. We consider sequential computation and instantaneous communication from each worker to the master, and each computation round, which can model a single iteration of the stochastic gradient descent (SGD) algorithm, is completed once the master receives k ≤ n distinct computations, referred to as the computation target. Our goal is to characterize the average completion time as a function of the computation load, which denotes the portion of the dataset available at each worker, and the computation target. We propose two computation scheduling schemes that specify the computation tasks assigned to each worker, as well as their order of execution. We also establish a lower bound on the minimum average completion time. Numerical results show a significant reduction in the average computation time over the existing coded and uncoded computing schemes.

We study collaborative machine learning at the wireless edge, where power and bandwidth-limited devices (workers), with limited local datasets, implement distributed stochastic gradient descent (DSGD) over-the-air with the help of a remote parameter server (PS). We consider a wireless multiple access channel (MAC) from the workers to the PS for communicating the local gradient estimates. We first introduce a digital DSGD (D-DSGD) scheme, assuming that the workers operate on the boundary of the MAC capacity region at each iteration of the DSGD algorithm, and digitize their estimates within the bit budget allowed by the employed power allocation. We then introduce an analog scheme, called A-DSGD, motivated by the additive nature of the wireless MAC, where the workers send their gradient estimates over the MAC through the available channel bandwidth without employing any digital code. Numerical results show that A-DSGD converges much faster than D-DSGD. The improvement is particularly compelling at low power and low bandwidth regimes. We also observe that the performance of A-DSGD improves with the number of workers, while D-DSGD deteriorates, limiting the ability of the latter in harnessing the computation power of many edge devices.

Most of the current literature on coded caching focus on a static scenario in which a fixed number of users synchronously place their requests from a content library, and the performance is measured in terms of the latency in satisfying all of these requests. In practice, however, users start watching an online video content asynchronously over time, and often abort watching a video before it is completed. The latter behavior is captured by the notion of audience retention rate, which measures the portion of a video content watched on average. In order to bring coded caching one step closer to practice, asynchronous user demands are considered in this paper by allowing user demands to arrive randomly over time, and both the popularity of video files, and the audience retention rates are taken into account. A decentralized partial coded delivery (PCD) scheme is proposed, and two cache allocation schemes are employed; namely homogeneous cache allocation (HoCA) and heterogeneous cache allocation (HeCA), which allocate users' caches among different chunks of the video files in the library. Numerical results validate that the proposed PCD scheme, either with HoCA or HeCA, outperforms conventional uncoded caching as well as the state-of-the-art decentralized caching schemes, which consider only the file popularities, and are designed for synchronous demand arrivals. An information-theoretical lower bound on the average delivery rate is also presented.

A cache-aided K-user Gaussian broadcast channel (BC) is studied. The transmitter has a library of N files, from which each user requests one. The users are equipped with caches of different sizes, which are filled without the knowledge of the user requests in a centralized manner. Differently from the literature, it is assumed that each file can be delivered to different users at different rates, which may correspond to different quality representations of the underlying content, e.g., scalable coded video segments. Accordingly, instead of a single achievable rate, the system performance is characterized by a rate tuple, which corresponds to the vector of rates users' requests can be delivered at. The goal is to characterize the set of all achievable rate tuples for a given total cache capacity by designing joint cache and channel coding schemes together with cache allocation across users. Assuming that the users are ordered in increasing channel quality, each file is coded into K layers, and only the first k layers of the requested file are delivered to user k, k=1, ..., K. Three different coding schemes are proposed, which differ in the way they deliver the coded contents over the BC; in particular, time-division, superposition, and dirty paper coding schemes are studied. Corresponding achievable rate regions are characterized, and compared with a novel outer bound. To the best of our knowledge, this is the first work studying the delivery of files at different rates over a cache-aided noisy BC.

A cache-aided K-user Gaussian broadcast channel is considered. The transmitter has a library of N equal-rate files, from which each user demands one. The impact of the equal-capacity receiver cache memories on the minimum required transmit power to satisfy all user demands is studied. Considering uniformly random demands across the library, both the minimum average power (averaged over all demand combinations) and the minimum peak power (minimum power required to satisfy all demand combinations) are studied. Upper bounds are presented on the minimum required average and peak transmit power as a function of the cache capacity considering both centralized and decentralized caching. The lower bounds on the minimum required average and peak power values are also derived assuming uncoded cache placement. The bounds for both the peak and average power values are shown to be tight in the centralized scenario through numerical simulations. The results in this paper show that proactive caching and coded delivery can provide significant energy savings in wireless networks.

A cache-aided broadcast network is studied, in which a server delivers contents to a group of receivers over a packet erasure broadcast channel. The receivers are divided into two sets with regards to their channel qualities: the weak and the strong receivers, where all the weak receivers have statistically worse channel qualities than all the strong receivers. The weak receivers, in order to compensate for the high erasure probability they encounter over the channel, are equipped with cache memories of equal size, while the receivers in the strong set have no caches. Data can be pre-delivered to the weak receivers' caches over the off-peak traffic period before the receivers reveal their demands. Allowing arbitrary erasure probabilities for the weak and strong receivers, a joint caching and channel coding scheme, which divides each file into several subfiles, and applies a different caching and delivery scheme for each subfile, is proposed. It is shown that all the receivers, even those without any cache memories, benefit from the presence of caches across the network. An information theoretic tradeoff between the cache size and the achievable rate is formulated. It is shown that the proposed scheme improves upon the state-of-the-art in terms of the achievable tradeoff.

A cache-aided erasure broadcast channel is studied. The receivers are divided into two sets: the weak and strong receivers, where the receivers in the same set all have the same erasure probability. The weak receivers, in order to compensate for the high erasure probability, are equipped with cache memories of equal size, while the receivers in the strong set have no caches. Data can be pre-delivered to weak receivers' caches over the off-peak traffic period before the receivers reveal their demands. A joint caching and channel coding scheme is proposed such that all the receivers, even the receivers without any cache memories, benefit from the presence of caches across the network. The trade-off between the cache size and the achievable rate is studied, and it is shown that the proposed scheme significantly improves the achievable trade-off upon the state-of-the-art.

A cache-aided K-user Gaussian broadcast channel (BC) is considered. The transmitter has a library of N equal-rate files, from which each user demands one. The impact of the equal-capacity receiver cache memories on the minimum required transmit power to satisfy all user demands is studied. Decentralized caching with uniformly random demands is considered, and both the minimum average power (averaged over all demand combinations) and the minimum peak power (minimum power required to satisfy the worst-case demand combination) are studied. Upper and lower bounds are presented on the minimum required average and peak transmit power as a function of the cache capacity, assuming uncoded cache placement. The gaps between the upper and lower bounds on both the minimum peak and average power values are shown to be relatively small through numerical results, particularly for large cache capacities.

A centralized coded caching system, consisting of a server delivering N popular files, each of size F bits, to K users through an error-free shared link, is considered. It is assumed that each user is equipped with a local cache memory with capacity MF bits, and contents can be proactively cached into these caches over a low traffic period, however, without the knowledge of the user demands. During the peak traffic period, each user requests a single file from the server. The goal is to minimize the number of bits delivered by the server over the shared link, known as the delivery rate, over all user demand combinations. A novel coded caching scheme for the cache capacity of M = (N-1)/K is proposed. It is shown that the proposed scheme achieves a smaller delivery rate than the existing coded caching schemes in the literature, when K > N ≥ 3. Furthermore, we argue that the delivery rate of the proposed scheme is within a constant multiplicative factor of 2 of the optimal delivery rate for cache capacities 1/K ≤ M ≤ (N -1)/K, when K > N ≥ 3.

Decentralized proactive caching and coded delivery is studied in a content delivery network, where each user is equipped with a cache memory, not necessarily of equal capacity. Cache memories are filled in advance during the off-peak traffic period in a decentralized manner, i.e., without the knowledge of the number of active users, their identities, or their particular demands. User demands are revealed during the peak traffic period, and are served simultaneously through an error-free shared link. The goal is to find the minimum delivery rate during the peak traffic period that is sufficient to satisfy all possible demand combinations. A group-based decentralized caching and coded delivery scheme is proposed, and it is shown to improve upon the state of the art in terms of the minimum required delivery rate when there are more users in the system than files. Numerical results indicate that the improvement is more significant as the cache capacities of the users become more skewed. A new lower bound on the delivery rate is also presented, which provides a tighter bound than the classical cut-set bound.

We consider the coded caching problem with a central server containing N files, each of length F bits, and K users, each equipped with a cache of capacity MF bits. We assume that coded contents can be proactively placed into users' caches at no cost during the placement phase. During the delivery phase, each user requests exactly one file from the database, and all the requests are served simultaneously by the server over an error-free common link. The goal is to utilize the local cache memories at the users to reduce the delivery rate from the server during the peak period. Here, we focus on a system which has more users than files, i.e., K > N. We first consider the centralized caching problem, in which the number and identity of active users are known in advance, and propose a group-based coded caching scheme for M = N/K, which improves upon the best achievable scheme in the literature. The proposed centralized caching scheme is then exploited in a decentralized setting, in which neither the number nor the identity of the active users are known during the placement phase. It is shown that the proposed coded caching scheme improves upon the best known decentralized delivery rate as well.

A cooperative network consisting of a source, one relay, and a destination is considered. The source and destination are both equipped with a single antenna. The relay is equipped with N antennas and operates in decode-and-forward (DF) mode. At the relay, the detected symbols are retransmitted using a novel beamforming method proposed in this paper that assumes the relay has 2-bit quantized information about the phase of all links from the relay to the destination. At the destination, the detector combines the received signals from the source and the relay in the first and second phases of transmission, respectively. An upper bound for the bit error rate (BER) performance of the detector is derived, and it is shown that the proposed beamforming and detector achieve full diversity order. The detector is extended to a multicast system with one single antenna source, one N-antenna DF relay and M single antenna destinations. It is assumed that the relay has 2-bit quantized phase information and exact signal-to-noise ratio (SNR) information at all the links from the relay to the destinations. An antenna selection algorithm based on a MAX-MIN criterion at the relay is proposed, and after beamforming, the detector is used at all destinations. By deriving a rigorous upper bound for BER performance at all destinations, it is shown that the proposed scheme achieves full diversity order at all destinations. Simulation results show that the BER performance of the proposed antenna selection algorithm outperforms the well known MAX-MIN antenna selection scheme.