Speculative decoding generally dictates having a small, efficient draft model that is either pretrained or distilled offline to a particular target model series, for instance, Llama or Qwen models. However, within online deployment settings, there are two major challenges: 1) usage of a target model that is incompatible with the draft model; 2) expectation of latency improvements over usage and time. In this work, we propose OmniDraft, a unified framework that enables a single draft model to operate with any target model and adapt dynamically to user data. We introduce an online n-gram cache with hybrid distillation fine-tuning to address the cross-vocabulary mismatch across draft and target models; and further improve decoding speed by leveraging adaptive drafting techniques. OmniDraft is particularly suitable for on-device LLM applications where model cost, efficiency and user customization are the major points of contention. This further highlights the need to tackle the above challenges and motivates the “one drafter for all” paradigm. We showcase the proficiency of the OmniDraft framework by performing online learning on math reasoning, coding and text generation tasks. Notably, OmniDraft enables a single Llama-68M model to pair with various target models including Vicuna-7B, Qwen2-7B and Llama3-8B models for speculative decoding; and additionally provides up to 1.5-2x speedup.

Recent advances in cancer immunotherapy have highlighted the potential of neoantigen-based vaccines. However, the design of such vaccines is hindered by the possibility of weak binding affinity between the peptides and the patient's specific human leukocyte antigen (HLA) alleles, which may not elicit a robust adaptive immune response. Triggering cross-immunity by utilizing peptide mutations that have enhanced binding affinity to target HLA molecules, while preserving their homology with the original one, can be a promising avenue for neoantigen vaccine design. In this study, we introduced UltraMutate, a novel algorithm that combines Reinforcement Learning and Monte Carlo Tree Search, which identifies peptide mutations that not only exhibit enhanced binding affinities to target HLA molecules but also retains a high degree of homology with the original neoantigen. UltraMutate outperformed existing state-of-the-art methods in identifying affinity-enhancing mutations in an independent test set consisting of 3660 peptide-HLA pairs. UltraMutate further showed its applicability in the design of peptide vaccines for Human Papillomavirus and Human Cytomegalovirus, demonstrating its potential as a promising tool in the advancement of personalized immunotherapy.

Tracking features in image sequences suffers from varying illumination and viewpoints. In recent years, learning-based features have achieved higher repeatability in challenging scenes and are considered to have the potential to solve this problem. However, features with high repeatability are not always easy to track. There is a gap between repeatability and trackability. To obtain features that are easily tracked in illumination and viewpoints, a data-driven approach expands the definition of good features. Trackability is defined end-to-end as the tracking error. According to this definition, a complete feature tracking process is used to compute the tracking loss and train the network. A four-layer convolutional network is used to extract low dimensional image information and obtain features. To validate the proposed method, we compare the tracking errors of several mainstream methods on a challenging test dataset, and the proposed method shows significant advantages. Then, fundamental matrix estimation and visual odometry experiments demonstrate the feature excels in practical tasks. Finally, the features were used in a visual inertial odometry system and achieved a 43% improvement in absolute trajectory error on the challenging dataset. All code will be open source for the benefit of the community.

The fifth generation of cellular wireless networks known as 5G is based on user-centric non-cellular concept where the users are surrounded by many network nodes cooperating to serve the users providing a “cell-center” experience throughout the network. Traditional multi point cooperation techniques often rely on centralized coordination and require different and often stringent requirement on the central controller, backhaul capacity and overall network synchronization. A novel two tiered, open loop and distributed cooperation technique is proposed in this paper where the lower tier with fixed or slowly changing parameters and preferably using SCMA provide a ubiquitous performance to mobile users and users exposed to multiple network nodes, while the higher tier provides service to the users close to the network nodes and maintain the overall network throughput. Users scheduled by the lower tier signaling use joint detection techniques from multiple network nodes. Other users scheduled to the higher tier jointly decode the lower tier signal from one or multiple network nodes before proceeding with the detection of their intended signal. The proposed algorithm is based on distributed scheduling and imposes limited requirements on the central controller and backhaul capacity and does not require stringent network time and frequency synchronization among network nodes. Unlike traditional cooperation techniques, the proposed method is open loop and requires very limited feedback and signaling overhead. Performance evaluations show that the proposed technique provides significant network coverage enhancement especially for high speed mobile users.

The joint user association and spectrum allocation problem is studied for multi-tier heterogeneous networks (HetNets) in both downlink and uplink in the interference-limited regime. Users are associated with base-stations (BSs) based on the biased downlink received power. Spectrum is either shared or orthogonally partitioned among the tiers. This paper models the placement of BSs in different tiers as spatial point processes and adopts stochastic geometry to derive the theoretical mean proportionally fair utility of the network based on the coverage rate. By formulating and solving the network utility maximization problem, the optimal user association bias factors and spectrum partition ratios are analytically obtained for the multi-tier network. The resulting analysis reveals that the downlink and uplink user associations do not have to be symmetric. For uplink under spectrum sharing, if all tiers have the same target signal-to-interference ratio (SIR), distance-based user association is shown to be optimal under a variety of path loss and power control settings. For both downlink and uplink, under orthogonal spectrum partition, it is shown that the optimal proportion of spectrum allocated to each tier should match the proportion of users associated with that tier. Simulations validate the analytical results. Under typical system parameters, simulation results suggest that spectrum partition performs better for downlink in terms of utility, while spectrum sharing performs better for uplink with power control.

This paper studies the downlink spectral efficiency of distributed antenna system (DAS) where antenna ports are distributed as a Poisson point process (PPP), while assuming channel state information is not available at the transmitter and each antenna has an individual power constraint. We first consider the case with a single user per cell and analyze regular DAS with fixed cell boundaries, and study both blanket transmission where the user is served by all the antenna ports within each cell, and selective transmission where only the closest antenna port to the user within each cell is selected. We derive efficiently computable spectral efficiency expressions as a function of the user location, and show the limitation of blanket transmission by establishing that the cell-edge spectral efficiency under blanket transmission is upper bounded by a constant. Further, from a network perspective, we also model users as a PPP and assume a time-division multiple-access (TDMA) scheme, and give analytical expressions for and compare the average spectral efficiencies of regular DAS and user-centric DAS where no fixed cell boundaries exist. We validate our models with simulation, and show that selective transmission outperforms blanket transmission for regular DAS, and user-centric DAS with selective transmission achieves a higher spectral efficiency averaged over the network than regular DAS.

This paper considers the joint optimization of frequency reuse and base-station (BS) bias for user association in downlink heterogeneous networks for load balancing and intercell interference management. To make the analysis tractable, we assume that BSs are randomly deployed as point processes in multiple tiers, where BSs in each tier have different transmission powers and spatial densities. A utility maximization framework is formulated based on the user coverage rate, which is a function of the different BS biases for user association and different frequency reuse factors across BS tiers. Compared to previous works where the bias levels are heuristically determined and full reuse is adopted, we quantitatively compute the optimal user association bias and obtain the closed-form solution of the optimal frequency reuse. Interestingly, we find that the optimal bias and the optimal reuse factor of each BS tier have an inversely proportional relationship. Further, we also propose an iterative method for optimizing these two factors. In contrast to system-level optimization solutions based on specific channel realization and network topology, our approach is off-line and is useful for deriving deployment insights. Numerical results show that optimizing user association and frequency reuse for multi-tier heterogeneous networks can effectively improve cell-edge user rate performance and utility.

This paper studies the ergodic capacity of a multicell distributed antenna system (DAS), where remote antenna ports are spread within each cell to cooperatively transmit to user terminals. Unlike most prior studies which assume the antenna ports to be deployed at fixed locations, this paper assumes the antenna ports to be distributed as a spatial Poisson point process (PPP) to account for the fact that in practice the antenna ports are randomly placed to cover wherever the dead spots are. We first model DAS within each cell as a downlink multiple-input single-output (MISO) channel with per-antenna power constraint while accounting for inter-cell (inter-cluster) interference. Two DAS layouts are considered: the “regular” layout where the antenna ports are randomly deployed within regular cellular boundary to serve a given user, and the “user-centric” layout where the antenna ports are distributed over a wide area and the users choose the surrounding antenna ports to form a “virtual cell” as its own serving antenna subset. Using the tool of stochastic geometry, we analytically derive efficiently computable ergodic capacity expressions for the two layouts of DAS. Using these expressions, the cell-edge capacity of DAS under the regular layout is shown to be upper-bounded by α/2, where α is the pathloss exponent. Numerical results show that the proposed analytical model can accurately model the first layout, and can well approximate the second layout when the serving radius of users is not large. Compared to the traditional cellular system where all antennas are co-located at the cell center, DAS has better cell-edge performance. Further, the user-centric DAS has higher capacity than the DAS under regular layout.

The placement of the access points (APs) has a significant impact on the wireless system performance, especially for irregular heterogeneous networks with hierarchical APs such as macro/micro base-stations, pico-stations, and femto-stations. Traditional system modeling and optimization are based on the regular 2D hexagonal cellular topology and a set of predefined large-scale propagation models, which is highly abstract and may be inaccurate. This paper considers the AP placement optimization problem in realistic deployment environments, where radio-wave propagation characteristics are accurately modeled using ray-tracing techniques. Toward this end, this paper proposes a novel concept of area proportional fairness utility for the entire network under a given user geographic distribution, and proposes an iterative method to optimize the placement of the APs for utility improvement while taking into account the mutual interference between the APs. The significant benefit of the placement optimization is shown in a numerical experiment conducted in a realistic Chicago downtown topology, where placement optimization is shown to improve the sum rate by up to 40%.

This paper examines the shared relay architecture for the wireless cellular network, where instead of deploying multiple separate relays within each cell sector, a single relay with multiple antennas is placed at the cell edge and is shared by multiple sectors. The advantage of shared relaying is that the joint processing of signals at the relay enables the mitigation of intercell interference. To maximize the benefit of shared relaying, the resource allocation and the scheduling of users among adjacent cell sectors need to be optimized jointly. Based on this motivation, this paper formulates a network utility maximization problem for the shared relay system that considers the practical wireless backhaul constraint of matching the relay-to-user rate demand with the base-station-to-relay rate supply using a set of pricing variables. In addition, zero-forcing beamforming is used at the shared relay to separate users spatially; multiple users are scheduled in the frequency domain to maximize frequency reuse. A heuristic but efficient scheduling and resource allocation algorithm is proposed accordingly. System-level simulations quantify the effectiveness of the proposed approach, and show that the incorporation of the shared relay can improve the overall network performance and in particular significantly increase the throughput of cell edge users as compared to separate relaying.

This paper considers subcarrier allocation for the multihop orthogonal frequency division multiple-access (OFDMA) broadcast networks consisting of one source, multiple destinations, and one amplify-and-forward (AF) relay. In this paper, proportional fair (PF) based subcarrier allocation is discussed to get a tradeoff between the system transmission rate and fairness. First, the problem is formulated as an optimization problem with prohibitive complexity. Second, by analyzing the optimal solution for the subcarrier allocation problem without PF constraint, two suboptimal schemes are proposed. The simulation results indicate that with the same fairness performance, the proposed schemes achieve considerable capacity gain compared with the conventional PF scheduling method which is extended simply from the single-hop system.

In this paper, we consider the optimal resource allocation problem for dual-hop systems with a source-destination pair and multiple decode-and-forward (DF) relays. Orthogonal frequency division multiplexing (OFDM) is adopted in both hops. To maximize the system capacity, we formulate a mixed binary integer programming problem to optimally allocate the subchannels and transmit power for both the source and the relay nodes. Because of the prohibitive complexity, a greedy heuristic algorithm is proposed to decompose the original problem into solvable subproblems. Simulation results demonstrate that the proposed algorithm has a near-optimal performance.

This paper investigates the spatial resource allocation algorithms for the uplink Virtual-MIMO (Multiple Input and Multiple Output) systems, where K client users (each with one or multiple antennas) are served by one multiple-antenna BS (Base Station). To fully exploit the spatial multi-user diversity gain, the finding of the optimal spatial co-channel users is discussed, which is formulated as a mixed binary integer programming problem. Based on the property of hermitian matrix, simulated annealing is used to recursively find the suboptimal spatial co-channel user group. By simulation results, our proposed strategy has almost the same system throughput performance as the optimal greedy algorithm, and can maintain lower computation complexity.

A two-hop relaying system where a source communicates with its destination through multiple relay nodes is studied for capacity maximization. Amplify-and-forward (AF) relaying is adopted for its simplicity. Taking the independent channel transfer function of different subcarriers and links into consideration, we formulate a resource allocation optimization problem for joint subcarrier assignment, subcarrier matching and power allocation of the two-hop transmission. A heuristic suboptimal algorithm which decomposes the joint problem is proposed. First, equal power allocation is assumed in a partial- update manner to help assigning subcarriers to each relay for matching; then optimal power allocation is achieved by joint waterfilling of source and relays under separate power constraints to further increase system capacity. Numerical results show that the proposed algorithm has a near-optimal performance.