arXiv subjects

Shengzhi Wang

Publications and source records attributed to Shengzhi Wang.

At least 19 recordsLinked to original sources

arXiv2026

World Model Control by Trajectory Reachability Metrics

Latent world models can learn representations that contain information needed for control, while the downstream controller may still rank candidate actions poorly when it relies on terminal latent distance alone. We study this failure in a fixed encoder and introduce trajectory reachability metrics (TRM), a small temporal pairwise cost trained from logged trajectories and used to rank predicted endpoints of a candidate action sequence against a goal. In the TwoRoom evaluation on 100 episodes from a high distance range, the original controller with LeWorldModel reaches 7.0% mean success. Temporal TRM trained after excluding all evaluation episodes reaches 96.7%, shuffled label controls stay at 0.0%, and TRM also improves PLDM from 32.7% to 84.0%. TopoNav, a separate pixel navigation task, shows the same link between repaired ranking and closed loop control. The selection audit with shared candidates (SASC) and rowspace interventions show that the gain comes from reweighting latent directions carrying the action decision. The XY rowspace contributes less than 1% of terminal latent MSE but carries most of the information needed for control. Coverage and boundary tests define the scope. Balanced doorway coverage restores a 0.0% coverage failure to 100.0% success, while an unseen wall orientation and contact rich PushT expose layout, dynamics, and recovery limits.

cs.LG↗

arXiv2026

Simultaneous Intensity and Frequency Control for Optical Waveform Generation using an Acousto Optic Modulator

We describe a method for calibrating the response of an acousto-optic modulator (AOM) to enable precise, arbitrary control of the intensity and frequency of optical fields. The method involves characterizing the nonlinear response of the AOM to its input RF drive voltage and applying an iterative calibration/correction algorithm to accurately map input RF amplitude and frequency drive to output optical intensity and frequency shifts. After a few calibration/correction iterations, the calibration of the AOM maintains the optical power within 1$\%$ of a constant target value over relatively wide frequency tuning range ($\approx$100 MHz), with $σ=0.2\%$ relative deviation. We apply this calibration method to generate programmable waveforms and optical pulses with tailored intensity and frequency profiles that closely match their target specifications. This technique offers a simple and robust method for applications that require high-accuracy optical modulation.

physics.optics↗

arXiv2026

MIRA: Medical Image Reflection for Agentic Diagnosis

Medical visual agents can use tools to inspect images and retrieve external knowledge, but indiscriminate tool use may introduce noisy or misleading evidence. Reliable diagnosis therefore requires not only acquiring additional observations, but also verifying whether tool actions are necessary and whether the resulting evidence supports the current hypothesis. We introduce MIRA (Medical Image Reflection for Agentic Diagnosis), a medical visual diagnostic framework for autonomous evidence search and reflective verification. MIRA dynamically invokes image-processing operations, including zooming, grounding, pointing, rotation, and measurement, as well as web search, while evaluating the relevance and consistency of the acquired evidence. We develop MIRA through a two-stage training strategy. First, a tool-augmented Monte Carlo Tree Search data engine explores diverse diagnostic hypotheses and jointly verifies visual grounding accuracy and semantic consistency to construct supervised fine-tuning trajectories. Second, reinforcement learning further improves decision-making through online reflective principle evolution: failure cases are distilled into candidate principles, and only principles that improve held-out rollout rewards are retained. Across nine medical visual reasoning benchmarks, MIRA achieves an average score of 64.73, improving its Qwen3-VL-8B backbone by 7.44 points. It also increases useful tool-use judgments from 56.2% to 73.8% and reduces harmful judgments from 8.9% to 1.6%. Qualitative analyses show that MIRA can re-examine evidence, correct premature conclusions, and adapt its tool-use strategy. Project page: https://MIRA-VL.github.io/

cs.CV↗

arXiv2026

Target-Aware Early Stage Ranking

Early Stage Ranking (ESR) in large-scale recommendation systems is dominated by ''user--item decoupling'' Two Tower architectures, which scale efficiently but cannot capture fine-grained, target-aware user--item interactions directly. We propose Target-Aware Early Stage Ranking (TESR), which augments the Two Tower with a Mixture of Attention (MoA) module trained as a request-level sequence modeling over user history. MoA combines (i) Hard Matching Attention (HMA) to capture explicit categorical-ID level overlap signals between user history and candidate item, (ii) target-aware HSTU attention for implicit affinities conditioned on the candidate, and (iii) target dependent and independent cross-attention for symmetric user-item contextualization. On top of this, a Multi-Logit Parameterized Gating (MLPG) head amplifies these signals at scoring time. To keep latency within ESR budgets, we co-design the architecture with FP8 quantization, custom kernels, and a Torch Inductor compilation path. On a production deployment, TESR delivers consistent offline NE wins and online topline gains, and is, to our knowledge, the first deployment of full target-aware attention sequence modeling in an ESR stage at this scale.

cs.LG↗

arXiv2026

PA-RNet: Perturbation-Aware Residual Network for Robust Multimodal Time Series Forecasting

In real-world applications, multimodal time-series forecasting faces a key challenge: textual information is often useful but unreliable. Auxiliary texts may contain irrelevant, ambiguous, incomplete, or structurally corrupted content, making direct text integration prone to introducing noisy semantic signals and degrading forecasting performance. Therefore, robust multimodal forecasting requires a model that can exploit useful textual context while suppressing misleading perturbations. To address this challenge, we propose PA-RNet, a carefully designed perturbation-aware residual network for robust multimodal time-series forecasting. Rather than directly fusing textual and numerical representations, PA-RNet first refines multimodal features in a perturbation-aware manner, preserving stable contextual information while reducing unstable or misleading signals. The refined textual representations are then aligned with temporal dynamics, enabling more reliable forecasting under noisy multimodal conditions. Theoretically, we prove that PA-RNet is Lipschitz continuous with respect to textual embeddings and show that the proposed spectral residual correction can reduce the expected prediction error under zero-mean textual perturbations. We further conduct supplementary experiments with injected textual perturbations to examine the robustness of PA-RNet. The results across diverse domains demonstrate that PA-RNet consistently outperforms state-of-the-art baselines and maintains stable forecasting performance under both original and noise-perturbed textual conditions.

cs.LG↗

arXiv2026

MAGIC: Multi-Step Advantage-Gated Causal Influence for Multi-agent Reinforcement Learning

A key challenge in multi-agent reinforcement learning (MARL) lies in designing learning signals that effectively promote coordination among agents. Designing such signals requires estimating how one agent's current action affects its teammates over future interaction steps. To address this, we introduce Multi-step Advantage-Gated Interventional Causal MARL (MAGIC), a framework that estimates multi-step action effects between agents and selectively converts them into intrinsic rewards. MAGIC uses counterfactual action interventions to compare teammate futures under factual and counterfactual branches, and introduces a gate based on advantage to direct exploration toward beneficial behaviors aligned with the task goal. Experiments on Multi-Agent Particle Environments (MPE) and StarCraft micromanagement benchmarks (SMAC and SMACv2) show that MAGIC consistently outperforms leading prior methods, with average relative final performance improvements of 26.9% and 10.1%, respectively.

cs.MA↗

arXiv2025

Shared Object Manipulation with a Team of Collaborative Quadrupeds

Utilizing teams of multiple robots is advantageous for handling bulky objects. Many related works focus on multi-manipulator systems, which are limited by workspace constraints. In this paper, we extend a classical hybrid motion-force controller to a team of legged manipulator systems, enabling collaborative loco-manipulation of rigid objects with a force-closed grasp. Our novel approach allows the robots to flexibly coordinate their movements, achieving efficient and stable object co-manipulation and transport, validated through extensive simulations and real-world experiments.

cs.RO↗

arXiv2025

Request-Only Optimization for Recommendation Systems

Deep Learning Recommendation Models (DLRMs) represent one of the largest machine learning applications on the planet. Industry-scale DLRMs are trained with petabytes of recommendation data to serve billions of users every day. To utilize the rich user signals in the long user history, DLRMs have been scaled up to unprecedented complexity, up to trillions of floating-point operations (TFLOPs) per example. This scale, coupled with the huge amount of training data, necessitates new storage and training algorithms to efficiently improve the quality of these complex recommendation systems. In this paper, we present a Request-Only Optimizations (ROO) training and modeling paradigm. ROO simultaneously improves the storage and training efficiency as well as the model quality of recommendation systems. We holistically approach this challenge through co-designing data (i.e., request-only data), infrastructure (i.e., request-only based data processing pipeline), and model architecture (i.e., request-only neural architectures). Our ROO training and modeling paradigm treats a user request as a unit of the training data. Compared with the established practice of treating a user impression as a unit, our new design achieves native feature deduplication in data logging, consequently saving data storage. Second, by de-duplicating computations and communications across multiple impressions in a request, this new paradigm enables highly scaled-up neural network architectures to better capture user interest signals, such as Generative Recommenders (GRs) and other request-only friendly architectures.

cs.IR↗

arXiv2025

DP-GPT4MTS: Dual-Prompt Large Language Model for Textual-Numerical Time Series Forecasting

Time series forecasting is crucial in strategic planning and decision-making across various industries. Traditional forecasting models mainly concentrate on numerical time series data, often overlooking important textual information such as events and news, which can significantly affect forecasting accuracy. While large language models offer a promise for integrating multimodal data, existing single-prompt frameworks struggle to effectively capture the semantics of timestamped text, introducing redundant information that can hinder model performance. To address this limitation, we introduce DP-GPT4MTS (Dual-Prompt GPT2-base for Multimodal Time Series), a novel dual-prompt large language model framework that combines two complementary prompts: an explicit prompt for clear task instructions and a textual prompt for context-aware embeddings from time-stamped data. The tokenizer generates the explicit prompt while the embeddings from the textual prompt are refined through self-attention and feed-forward networks. Comprehensive experiments conducted on diverse textural-numerical time series datasets demonstrate that this approach outperforms state-of-the-art algorithms in time series forecasting. This highlights the significance of incorporating textual context via a dual-prompt mechanism to achieve more accurate time series predictions.

cs.CL↗

arXiv2024

Design of a UE5-based digital twin platform

Aiming at the current mainstream 3D scene engine learning and building cost is too high, this thesis proposes a digital twin platform design program based on Unreal Engine 5 (UE5). It aims to provide a universal platform construction design process to effectively reduce the learning cost of large-scale scene construction. Taking an actual project of a unit as an example, the overall cycle work of platform building is explained, and the digital twin and data visualization technologies and applications based on UE5 are analyzed. By summarizing the project implementation into a process approach, the standardization and operability of the process pathway is improved.

cs.HC↗

arXiv2024

NoiseBoost: Alleviating Hallucination with Noise Perturbation for Multimodal Large Language Models

Multimodal large language models (MLLMs) contribute a powerful mechanism to understanding visual information building on large language models. However, MLLMs are notorious for suffering from hallucinations, especially when generating lengthy, detailed descriptions for images. Our analysis reveals that hallucinations stem from the inherent summarization mechanism of large language models, leading to excessive dependence on linguistic tokens while neglecting vision information. In this paper, we propose NoiseBoost, a broadly applicable and simple method for alleviating hallucinations for MLLMs through the integration of noise feature perturbations. Noise perturbation acts as a regularizer, facilitating a balanced distribution of attention weights among visual and linguistic tokens. Despite its simplicity, NoiseBoost consistently enhances the performance of MLLMs across common training strategies, including supervised fine-tuning and reinforcement learning. Further, NoiseBoost pioneerly enables semi-supervised learning for MLLMs, unleashing the power of unlabeled data. Comprehensive experiments demonstrate that NoiseBoost improves dense caption accuracy by 8.1% with human evaluation and achieves comparable results with 50% of the data by mining unlabeled data. Code and models are available at https://kaiwu5.github.io/noiseboost.

cs.CV↗

arXiv2023

A Dual-mode Local Search Algorithm for Solving the Minimum Dominating Set Problem

Given a graph, the minimum dominating set (MinDS) problem is to identify a smallest set $D$ of vertices such that every vertex not in $D$ is adjacent to at least one vertex in $D$. The MinDS problem is a classic $\mathcal{NP}$-hard problem and has been extensively studied because of its many disparate applications in network analysis. To solve this problem efficiently, many heuristic approaches have been proposed to obtain a good solution within an acceptable time limit. However, existing MinDS heuristic algorithms are always limited by various tie-breaking cases when selecting vertices, which slows down the effectiveness of the algorithms. In this paper, we design an efficient local search algorithm for the MinDS problem, named DmDS -- a dual-mode local search framework that probabilistically chooses between two distinct vertex-swapping schemes. We further address limitations of other algorithms by introducing vertex selection criterion based on the frequency of vertices added to solutions to address tie-breaking cases, and a new strategy to improve the quality of the initial solution via a greedy-based strategy integrated with perturbation. We evaluate DmDS against the state-of-the-art algorithms on seven datasets, consisting of 346 instances (or families) with up to tens of millions of vertices. Experimental results show that DmDS obtains the best performance in accuracy for almost all instances and finds much better solutions than state-of-the-art MinDS algorithms on a broad range of large real-world graphs.

cs.SI↗

arXiv2023

Generation of highly retrievable atom photon entanglement with a millisecond lifetime via a spatially multiplexed cavity

Qubit memory that is entangled with photonic qubit is the building block for long distance quantum repeaters. Cavity enhanced and long lived spin wave photon entanglement has been demonstrated by applying dual laser beams onto optical-lattice atoms. However, owing to cross readouts by two beams, retrieval efficiency of spin wave qubit is decreased by one quarter compared to that of single mode spin wave at all storage times. Here, by coupling cold atoms to two modes of a polarization interferometer based cavity, we achieve perfect qubit retrieval in cavity enhanced and long lived atom photon entanglement. A write laser beam is applied onto cold atoms, we then create a magnetic field insensitive spin wave qubit that is entangled with the photonic qubit encoded onto two arms of the interferometer. The spin wave qubit is retrieved by a read beam, which avoids the cross readouts. Our experiment demonstrates 540μs storage time at 50% intrinsic qubit retrieval efficiency, which is 13.5 times longer than the best reported result.

quant-ph↗

arXiv2022

Towards Exact Interaction Force Control for Underactuated Quadrupedal Systems with Orthogonal Projection and Quadratic Programming

Projected Inverse Dynamics Control (PIDC) is commonly used in robots subject to contact, especially in quadrupedal systems. Many methods based on such dynamics have been developed for quadrupedal locomotion tasks, and only a few works studied simple interactions between the robot and environment, such as pressing an E-stop button. To facilitate the interaction requiring exact force control for safety, we propose a novel interaction force control scheme for underactuated quadrupedal systems relying on projection techniques and Quadratic Programming (QP). This algorithm allows the robot to apply a desired interaction force to the environment without using force sensors while satisfying physical constraints and inducing minimal base motion. Unlike previous projection-based methods, the QP design uses two selection matrices in its hierarchical structure, facilitating the decoupling between force and motion control. The proposed algorithm is verified with a quadrupedal robot in a high-fidelity simulator. Compared to the QP designs without the strategy of using two selection matrices and the PIDC method for contact force control, our method provided more accurate contact force tracking performance with minimal base movement, paving the way to approach the exact interaction force control for underactuated quadrupedal systems.

cs.RO↗

arXiv2022

Towards Safe Landing of Falling Quadruped Robots Using a 3-DoF Morphable Inertial Tail

Falling cat problem is well-known where cats show their super aerial reorientation capability and can land safely. For their robotic counterparts, a similar falling quadruped robot problem, has not been fully addressed, although achieving safe landing as the cats has been increasingly investigated. Unlike imposing the burden on landing control, we approach to safe landing of falling quadruped robots by effective flight phase control. Different from existing work like swinging legs and attaching reaction wheels or simple tails, we propose to deploy a 3-DoF morphable inertial tail on a medium-size quadruped robot. In the flight phase, the tail with its maximum length can self-right the body orientation in 3D effectively; before touch-down, the tail length can be retracted to about 1/4 of its maximum for impressing the tail's side-effect on landing. To enable aerial reorientation for safe landing in the quadruped robots, we design a control architecture, which has been verified in a high-fidelity physics simulation environment with different initial conditions. Experimental results on a customized flight-phase test platform with comparable inertial properties are provided and show the tail's effectiveness on 3D body reorientation and its fast retractability before touch-down. An initial falling quadruped robot experiment is shown, where the robot Unitree A1 with the 3-DoF tail can land safely subject to non-negligible initial body angles.

cs.RO↗

arXiv2021

Generation of entanglement between a highly wave-packet-tunable photon and a spin-wave memory in cold atoms

Controls of waveforms (pulse durations) of single photons are important tasks for effectively interconnecting disparate atomic memories in hybrid quantum networks. So far, the waveform control of single photon that is entangled with an atomic memory remains unexplored. Here, we demonstrated control of waveform length of the photon that is entangled with an atomic spin-wave memory by varying light-atom interaction time in cold atoms. The Bell parameter S as a function of the duration of photon pulse is measured, which shows that violations of Bell equality can be achieved for the photon pulse in the duration range from 40 ns to 50 us, where, S=2.64+/-0.02 and S=2.26+/-0.05 for the 40-ns and 50-μs durations, respectively. The measured results show that S parameter decreases with the increase in the pulse duration. We confirm that the increase in photon noise probability per pulse with the pulse-duration is responsible for the S decrease.

quant-ph↗

arXiv2021

Generation of perfect-cavity-enhanced atom-photon entanglement with a millisecond lifetime via a spatially-multiplexed cavity

A qubit memory is the building block for quantum information. Cavity-enhanced spin-wave-photon entanglement has been achieved by applying dual-control modes. However, owing to cross readouts between the modes, the qubit retrieval efficiency is about one quarter lower than that for a single spin-wave mode at all storage times. Here, we overcome cross readouts using a multiplexed ring cavity. The cavity is embedded with a polarization interferometer, and we create a write-out photonic qubit entangled with a magnetic-field-insensitive spin-wave qubit by applying a single-mode write-laser beam to cold atoms. The spin-wave qubit is retrieved with a single-mode read-laser beam, and the quarter retrieval-efficiency loss is avoided at all storage times. Our experiment demonstrates 50% intrinsic retrieval efficiency for 540 microsecond storage time, which is 13.5 times longer than the best reported result. Importantly, our multiplexed-cavity scheme paves one road to generate perfect-cavity-enhanced and large-scale multiplexed spin-wave-photon entanglement with a long lifetime.

quant-ph↗

arXiv2020

Robot in the mirror: toward an embodied computational model of mirror self-recognition

Self-recognition or self-awareness is a capacity attributed typically only to humans and few other species. The definitions of these concepts vary and little is known about the mechanisms behind them. However, there is a Turing test-like benchmark: the mirror self-recognition, which consists in covertly putting a mark on the face of the tested subject, placing her in front of a mirror, and observing the reactions. In this work, first, we provide a mechanistic decomposition, or process model, of what components are required to pass this test. Based on these, we provide suggestions for empirical research. In particular, in our view, the way the infants or animals reach for the mark should be studied in detail. Second, we develop a model to enable the humanoid robot Nao to pass the test. The core of our technical contribution is learning the appearance representation and visual novelty detection by means of learning the generative model of the face with deep auto-encoders and exploiting the prediction error. The mark is identified as a salient region on the face and reaching action is triggered, relying on a previously learned mapping to arm joint angles. The architecture is tested on two robots with a completely different face.

cs.RO↗