RF hardware design is a complicated, time-consuming, and expertise-bound process, which constrains the development and adoption of hardware innovation. Manual workflows do not scale to emerging wireless applications, while existing design generation strategies, e.g., learning-based approaches, lack training efficiency and generalizability across hardware types, frequency bands, operation modes, and substrate materials. In this paper, we present AutoRF, the first agentic framework for automated RF hardware design, supporting metasurfaces and antennas. It allows users to specify their demands and generate corresponding designs. We introduce extensible design abstractions to enable a modular framework. The core of the framework is an efficient and generalizable algorithm for design search and optimization, which utilizes LLMs to drive both circuit model simulator and EM simulator. To boost reliability and performance, we propose custom programming interfaces and a rule reviewer agent as feedback sources to train a specialized LLM. Evaluation demonstrates high success rate and significant optimization speedup; case studies with fabricated metasurface and antennas, ranging from 2.4 GHz to sub-THz, illustrate an ability to derive novel designs for next-generation wireless infrastructure.

Wireless sensing literature has long aspired to achieve X-ray-like vision at radio frequencies. Yet, state-of-the-art wireless sensing literature has yet to generate the archetypal X-ray image: one of the bones beneath flesh. In this paper, we explore OssiSense, a penetration-based RF-imaging system for imaging bones at mm-resolution, one that significantly exceeds prior penetration-based RF imaging literature. Indeed the long wavelength, significant attenuation and complex diffraction that occur as RF propagates through flesh, have long limited imaging resolution (to several centimeters at best). We address these concerns through a novel penetration-based synthetic aperture algorithm, coupled with a learning-based pipeline to correct for diffraction-induced artifacts. A detailed evaluation of meat models demonstrates a resolution improvement from sub-decimeter to sub-centimeter over prior art in RF penetrative imaging.

Large Language Models (LLMs) have made significant progress in open-ended dialogue, yet their inability to retain and retrieve relevant information from long-term interactions limits their effectiveness in applications requiring sustained personalization. External memory mechanisms have been proposed to address this limitation, enabling LLMs to maintain conversational continuity. However, existing approaches struggle with two key challenges. First, rigid memory granularity fails to capture the natural semantic structure of conversations, leading to fragmented and incomplete representations. Second, fixed retrieval mechanisms cannot adapt to diverse dialogue contexts and user interaction patterns. In this work, we propose Reflective Memory Management (RMM), a novel mechanism for long-term dialogue agents, integrating forward- and backward-looking reflections: (1) Prospective Reflection, which dynamically summarizes interactions across granularities—utterances, turns, and sessions—into a personalized memory bank for effective future retrieval, and (2) Retrospective Reflection, which iteratively refines the retrieval in an online reinforcement learning (RL) manner based on LLMs’ cited evidence. Experiments show that RMM demonstrates consistent improvement across various metrics and benchmarks. For example, RMM shows more than 10% accuracy improvement over the baseline without memory management on the LongMemEval dataset.

Mihir Parmar, Palash Goyal, Xin Liu, Yiwen Song, Mingyang Ling, Chitta Baral, Hamid Palangi, Tomas Pfister. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In this paper, we design and fabricate cost-effective, electronics-free millimeter-wave (mmWave) fiducial markers that support clutter-free detection on radars. Fiducial markers are widely used in camera-based systems to provide spatial information in robotic navigation applications. QR-code-like fiducial markers can be readily printed on paper - low-cost, easy to produce, and electronics-free. Yet, an analogous mmWave solution is yet to appear, which stems from the unique challenge in the mmWave context: its vulnerability to clutter. Existing solutions either trade hardware simplicity for clutter resilience or stay simple but remain vulnerable to environmental multipath. This paper seeks to solve this dilemma.

Wireless research has advanced in utilizing channel diversity and beamforming for more efficient communication, sensing, and harvesting ambient energy. We demonstrate our wireless robotic platform that utilizes radio-frequency beamforming for robot actuation. The platform delivers a maximum of 60 watts of power accurately towards soft robotic actuators by efficient frequency-aware beamforming. We also engineer soft actuators to absorb microwaves of specific frequencies to enable selective actuation. In this demonstration, we show a simplified version of our system that achieves frequency-selective actuation of two actuators to enable simple robot locomotion.

Policy exploration is critical in reinforcement learning (RL), where existing approaches include $\epsilon$-greedy, Gaussian process, etc. However, these approaches utilize preset stochastic processes and are indiscriminately applied in all kinds of RL tasks without considering task-specific features that influence policy exploration. Moreover, during RL training, the evolution of such stochastic processes is rigid, which typically only incorporates a decay in the variance, failing to adjust flexibly according to the agent's real-time learning status. Inspired by the analyzing and reasoning capability of large language models (LLMs), we design **LLM-Explorer** to adaptively generate task-specific exploration strategies with LLMs, enhancing the policy exploration in RL. In our design, we sample the learning trajectory of the agent during the RL training in a given task and prompt the LLM to analyze the agent's current policy learning status and then generate a probability distribution for future policy exploration. Updating the probability distribution periodically, we derive a stochastic process specialized for the particular task and dynamically adjusted to adapt to the learning process. Our design is a plug-in module compatible with various widely applied RL algorithms, including the DQN series, DDPG, TD3, and any possible variants developed based on them. Through extensive experiments on the Atari and MuJoCo benchmarks, we demonstrate LLM-Explorer's capability to enhance RL policy exploration, achieving an average performance improvement up to 37.27%. Our code is open-source at https://github.com/tsinghua-fib-lab/LLM-Explorer for reproducibility.

MOTIVATION: Spatial transcriptomics (ST) enables the study of gene expression within its spatial context in histopathology samples. To date, a limiting factor has been the resolution of sequencing based ST products. The introduction of the Visium High Definition (HD) technology opens the door to cell resolution ST studies. However, challenges remain in the ability to accurately map transcripts to cells and in assigning cell types based on the transcript data. RESULTS: We developed ENACT, a self-contained pipeline that integrates advanced cell segmentation with Visium HD transcriptomics data to infer cell types across whole tissue sections. Our pipeline incorporates novel bin-to-cell assignment methods, enhancing the accuracy of single-cell transcript estimates. Validated on diverse synthetic and real datasets, our approach is both scalable to samples with hundreds of thousands of cells and effective, offering a robust solution for spatially resolved transcriptomics analysis. AVAILABILITY AND IMPLEMENTATION: ENACT source code is available at https://github.com/Sanofi-Public/enact-pipeline. Experimental data are available at https://zenodo.org/records/14748859.

Trajectory generation is a vital element in AI applications. Firstly, it enables simulation such as traffic simulation and epidemic spreading modeling. Secondly, it can provide synthetic privacy-preserving data for training AI models. Notably, trajectory generation featuring controllable user profiles holds substantial value in generating customized mobility trajectories tailored to diverse requirements. However, relevant work is still lacking. On the one hand, traditional deep generative models fall short in guiding controllable trajectory generation due to the statistical nature of human mobility patterns and the corresponding insufficient control mechanisms. On the other hand, though the diffusion model has demonstrated strong generative capabilities in many fields, to achieve controllable generation on discrete trajectory data, we still need to redesign the structure of the continuous diffusion model. In this article, we introduce a controllable trajectory generation framework that leverages a continuous diffusion model and classifier guidance for more robust condition control. Our proposed framework comprises two modules: a latent trajectory diffusion model and a trajectory classifier for profile guidance. Experiments on two real-world mobility datasets consistently demonstrate its capability of generating trajectories matching given user profiles and conforming to human mobility patterns. Our source code and trained models are released at https://github.com/tsinghua-fib-lab/User-Profile-Guided-Latent-Diffusion .

Predicting the resilience of complex networks, which represents the ability to retain fundamental functionality amidst external perturbations or internal failures, plays a critical role in understanding and improving real-world complex systems. Traditional theoretical approaches grounded in nonlinear dynamical systems rely on prior knowledge of network dynamics. On the other hand, data-driven approaches frequently encounter the challenge of insufficient labeled data, a predicament commonly observed in real-world scenarios. In this paper, we introduce a novel resilience prediction framework for complex networks, designed to tackle this issue through generative data augmentation of network topology and dynamics. The core idea is the strategic utilization of the inherent joint distribution present in unlabeled network data, facilitating the learning process of the resilience predictor by illuminating the relationship between network topology and dynamics. Experiment results on three network datasets demonstrate that our proposed framework TDNetGen can achieve high prediction accuracy up to 85%-95%. Furthermore, the framework still demonstrates a pronounced augmentation capability in extreme low-data regimes, thereby underscoring its utility and robustness in enhancing the prediction of network resilience. We have open-sourced our code in the following link, https://github.com/tsinghua-fib-lab/TDNetGen.

Microwave ovens have become an essential cooking appliance owing to their convenience and efficiency. However, microwave ovens suffer from uneven distribution of energy, which causes prolonged delays, unpleasant cooking experiences, and even safety concerns. Despite significant research efforts, current solutions remain inadequate. In this paper, we first conduct measurement studies to understand the energy distribution for 10 microwave ovens and show their energy distribution in both 2D and 3D is very skewed, with notably lower energy levels at the center of the microwave cavity, where food is commonly placed. To tackle this challenge, we propose a novel methodology to enhance the performance of microwave ovens. Our approach begins with the development of a measurement driven model of a microwave oven. We construct a detailed 3D model in the High Frequency Structure Simulator (HFSS) and use real temperature measurements from a microwave to derive critical parameters relevant to the appliance's functionality (e.g., operating frequency, waveguide specifications). We then develop a novel approach that optimizes the design and placement of a low-cost passive metasurface for a given heating objective. Using extensive experiments, we demonstrate the efficacy of our approach across diverse food, optimization objectives, and microwave ovens.

Nowadays, unmanned aerial vehicles (UAVs) are widely applied in crowd sensing. For UAV-enabled crowd sensing (UAVCS) systems, the sensing outcome and charging cost are two primary concerns. To achieve a satisfactory sensing outcome under the charging budget, we exploit joint moving, sensing, and charging scheduling of UAVs, as they all have critical impacts on such two objectives. However, the dynamically generated sensing targets and the variety of sensing tasks a UAVCS system may face make farsighted scheduling of UAVs rather challenging. To this end, we propose a novel multi-task constrained multi-agent reinforcement learning (MARL) method to help UAVs make distributed moving, sensing, and charging decisions. Specifically, we design a multi-task MARL framework to learn a single generic policy for a large collection of tasks, and propose a primal-dual training algorithm that alternates between improving the overall sensing outcome and reducing each task's constraint violation. Theoretically, we show that our algorithm provably converges, and analyze the optimality gap and constraint violation of the trained policy on unseen tasks. Extensive experiments on an incident dataset in New York City demonstrate that our method outperforms strong baselines in sensing outcome maximization and budget satisfaction, and also generalize well to unseen tasks.

Object models are gradually progressing from predicting just category labels to providing detailed descriptions of object instances. This motivates the need for large datasets which go beyond traditional object masks and provide richer annotations such as part masks and attributes. Hence, we introduce PACO: Parts and Attributes of Common Objects. It spans 75 object categories, 456 object-part categories and 55 attributes across image (LVIS) and video (Eg04D) datasets. We provide 641K part masks an-notated across 260K object boxes, with roughly half of them exhaustively annotated with attributes as well. We design evaluation metrics and provide benchmark results for three tasks on the dataset: part mask segmentation, object and part attribute prediction and zero-shot instance detection. Dataset, models, and code are open-sourced at https://github.com/jacebookresearch/paco.

Recent work on battery-free soft robotics has demonstrated the use of liquid crystal elastomers (LCE) to build shape-changing materials activated by applied external heat. However, sources of heat must typically be in direct field-of-view of the robot (i.e. NIR, laser, and visual light EM sources or convective heats guns), be tethered to an external power supply (i.e. thermoelectric heating or resistive joule heaters), or require a heavy on-board battery that limits mobility and range. This paper presents a novel battery-free soft-robotics platform that can crawl through confined, enclosed, and hard-to-reach spaces (e.g. packages, machinery, pipes, etc.), hidden from view of heating infrastructure. This is achieved through the co-design of a soft robotics platform and integrated soft conductive traces that enable wireless (microwave) heating through remote stimulation. We achieve fast actuation through a careful choice of materials and the overall mechanical structure of the robot to maximize heating efficiency. Further, the robot is actively tracked through enclosed spaces using a mm Wave radar to direct heat to its location. We provide a detailed evaluation on the robot's heating efficiency, location-tracking accuracy and crawling speed.

This paper proposes a new primitive that allows soft robots to be physically controlled in a completely non-line-of-sight context using wireless energy - a process we call wireless actuation. Soft robots, which are composed entirely of soft materials and exclude any rigid components, are highly flexible platforms that can change their shape. This paper considers a specific class of soft robots composed of liquid-crystal elastomers (LCE) that are entirely electronics-free and engineered to change shape when heated to 60 °C. Traditionally, such robotic systems must be in line-of-sight of a light source, such as infrared to be moved, or require an external power supply for Joule heating and often take several tens of seconds to heat. We present WASER, a novel RF-based heating platform that allows electronics-free robots to be actuated rapidly (within a few seconds) and potentially in non-line-of-sight. WASER achieves this through innovations in both wireless systems and material science. On the wireless front, WASER develops a new blind beamforming solution that directs high-power wireless energy at fine spatial granularity without electronics on the robot to provide feedback. On the material science front, WASER exhibits heat-responsive shape-morphing and energy-harvesting material functionalities that allow for rapid wireless heating. We implement and evaluate WASER and demonstrate diverse shape-morphing capabilities.

Channel estimation is critical to millimeter-wave capability. Unlike sub-6 GHz WiFi, commercial-off-the-shelf 60 GHz WiFi devices adopt a single RF-chain and can only report the combined received signal strength (RSS) instead of the antenna-wise channel state information (CSI). Therefore, recovering the CSI using a limited number of RSS measurements is important but faces the following challenges: (i) solving a non-convex objective is hard and computationally heavy, (ii) the estimation error is high with insufficient RSS measurements, and (iii) channel fluctuates dynamically. To jointly tackle them, we propose 2ACE, an Accelerated and Accurate Channel Estimation approach using spectral profile-driven multiresolutional compressive sensing. Our thorough experiments show that 2ACE yields 2--8 dB reduction in CSI estimation error, 1--5 dB improvement in beamforming performance, and 5° - 10° reduction in angle-of-departure estimation error over the existing schemes.

Recent years have witnessed the increasing popularity of electric buses (e-buses) around the globe due to their environment friendly nature. However, various factors, such as the prohibitive purchasing costs and the scarcity of large-scale charging facilities, hinder the wider adoption of e-buses. Thus, to effectively cut the cost of building and maintaining urban e-bus systems, we optimize the dispatching strategy for urban e-bus systems to satisfy public transportation demands with the minimum e-bus fleet. Specifically, we propose to systematically exploit at city-scale cross-line dispatching, a smart dispatching strategy allowing one bus to serve multiple bus lines when necessary. Technically, we construct a novel and generalizable graph-theoretic model for urban e-bus systems integrating e-buses non-negligible charging time, the spatio-temporal constraints of bus trips, and various other real-world factors. We prove that it is NP-hard, and has no$(2-\epsilon)$-approximation algorithm. Next, we propose a polynomial-time algorithm solving the problem with a guaranteed approximation ratio. Furthermore, we conduct extensive experiments on a large-scale real-world bus dataset from Shenzhen, China, which validate the effectiveness of our algorithms. As shown by our experimental results, to serve 300 bus lines, our dispatching strategy needs 38.2% less e-buses than the one currently used in practice.

Nowadays, the rapid development of self-driving technology and its fusion with the current vehicle electrification process has given rise to electric self-driving taxis (es-taxis). Foreseeably, es-taxis will become a major force that serves the massive urban mobility demands not far into the future. Though promising, it is still a fundamental unsolved problem of effectively deciding when and where a city-scale fleet of es-taxis should be charged, so that enough es-taxis will be available whenever and wherever ride requests are submitted. Furthermore, charging decisions are far from isolated, but tightly coupled with the order dispatch process that matches orders with es-taxis. Therefore, in this paper, we investigate the problem of joint order dispatch and charging in es-taxi systems, with the objective of maximizing the ride-hailing platform’s long-term cumulative profit. Technically, such problem is challenging in a myriad of aspects, such as long-term profit maximization, partial statistical information on future orders, etc. We address the various arising challenges by meticulously integrating a series of methods, including distributionally robust optimization, primal-dual transformation, and second order conic programming to yield far-sighted decisions. Finally, we validate the effectiveness of our proposed methods though extensive experiments based on two large-scale real-world online ride-hailing order datasets.

Companies today increasingly rely on microservice architectures to deliver service for their large-scale mobile or web applications. However, not all developers working on these applications are distributed systems engineers and therefore do not anticipate partial failure: where one or more of the dependencies of their service might be unavailable once deployed into production. Therefore, it is paramount that these issues be raised early and often, ideally in a testing environment or before the code ships to production.

The label noise transition matrix, characterizing the probabilities of a training instance being wrongly annotated, is crucial to designing popular solutions to learning with noisy labels. Existing works heavily rely on finding “anchor points” or their approximates, defined as instances belonging to a particular class almost surely. Nonetheless, finding anchor points remains a non-trivial task, and the estimation accuracy is also often throttled by the number of available anchor points. In this paper, we propose an alternative option to the above task. Our main contribution is the discovery of an efficient estimation procedure based on a clusterability condition. We prove that with clusterable representations of features, using up to third-order consensuses of noisy labels among neighbor representations is sufficient to estimate a unique transition matrix. Compared with methods using anchor points, our approach uses substantially more instances and benefits from a much better sample complexity. We demonstrate the estimation accuracy and advantages of our estimates using both synthetic noisy labels (on CIFAR-10/100) and real human-level noisy labels (on Clothing1M and our self-collected human-annotated CIFAR-10). Our code and human-level noisy CIFAR-10 labels are available at https://github.com/UCSC-REAL/HOC.

Nowadays, crowdsourcing has become an increasingly popular paradigm for large-scale data collection, annotation, and classification. Today's rapid growth of crowdsourcing platforms calls for effective worker selection mechanisms, which oftentimes have to operate with a priori unknown worker reliability. We discover that the empirical entropy of workers' results, which measures the uncertainty in the final aggregated results, naturally becomes a suitable metric to evaluate the outcome of crowdsourcing tasks. Therefore, this paper designs a worker selection mechanism that minimizes the empirical entropy of the results submitted by participating workers. Specifically, we formulate worker selection under sequentially arriving tasks as a combinatorial multi-armed bandit problem, which treats each worker as an arm, and aims at learning the best combination of arms that minimize the cumulative empirical entropy. By information theoretic methods, we carefully derive an estimation of the upper confidence bound for empirical entropy minimization, and leverage it in our minimum entropy upper confidence bound (ME-UCB) algorithm to balance exploration and exploitation. Theoretically, we prove that ME-UCB has a regret upper bound of O(1), which surpasses existing submodular UCB algorithms. Our extensive experiments with both a synthetic and real-world dataset empirically demonstrate that our ME-UCB algorithm outperforms other state-of-the-art approaches.

The rapid development of information and communication technologies has given rise to mobility-on-demand (MoD) systems (e.g., Uber, Didi) that have fundamentally revolutionized urban transportation. One common feature of today's MoD systems is the integration of ridesharing due to its cost-efficient and environment-friendly natures. However, a fundamental unsolved problem for such systems is how to serve people's heterogeneous transportation demands with as few vehicles as possible. Naturally, solving such minimum fleet problem is essential to reduce the vehicles on the road to improve transportation efficiency. Therefore, we investigate the fleet minimization problem in ridesharing-aware MoD systems. We use graph-theoretic methods to construct a novel order graph capturing the complicated inter-order shareability, each order's spatial-temporal features, and various other real-world factors. We then formulate the problem as a tree cover problem over the order graph, which differs from the traditional coverage problems. Theoretically, we prove the problem is NP-hard, and propose a polynomial-time algorithm with a guaranteed approximation ratio. Besides, we address the online fleet minimization problem, where orders arrive in an online manner. Finally, extensive experiments on a city-scale dataset from Shenzhen, containing 21 million orders from June 1st to 30th, 2017, validate the effectiveness of our algorithms.