arXiv subjects

Yixian Zhang

Publications and source records attributed to Yixian Zhang.

17 recordsLinked to original sources

arXiv2026

Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents

Language-conditioned manipulation requires both precise contact-rich control and robust reasoning over language, scenes, and long horizons. End-to-end Vision-Language-Action (VLA) models provide strong local visuomotor skills, but they are trained on in-distribution task trajectories and often fail under deployment perturbations such as semantic retargeting, goal re-binding, spatial-layout shifts, and unstable local contacts. LLM coding agents provide complementary semantic and compositional reasoning, but purely analytic primitives struggle with irregular grasping, constrained placement, and articulated-object interaction. We present Harness VLA, a memory-augmented agentic framework that exposes a frozen VLA as a retryable contact-rich primitive and composes it with a small fixed library of analytic primitives for grounding, staging, transport, navigation, and release. Rather than expanding the skill library, the harness learns the operating range of these fixed primitives from task-specific execution traces, global success rules, and failure models. By lifting semantic re-grounding, non-contact execution, and VLA re-staging to the planner while reserving the frozen VLA for local contact-rich phases, Harness VLA extends pretrained VLAs beyond their original trajectory distribution without finetuning. Across perturbed tabletop, household kitchen, and clean-to-randomized bimanual manipulation, Harness VLA improves over the strongest relevant baselines by 38.6 and 25.4 percentage points on LIBERO-Pro and RoboCasa365, respectively, and reaches 58.4% on RoboTwin C2R. Code is available at https://github.com/RLinf/RPent.

cs.RO↗

arXiv2026

Critic-Free Pretraining for Efficient Online Reinforcement Learning Fine-Tuning

Offline-to-online (O2O) reinforcement learning aims to leverage policies pretrained on static datasets while improving them through online interaction. However, directly reusing an offline-trained critic can hinder online fine-tuning: as the policy and data distribution change rapidly, value estimates inherited from offline training may become misaligned with the online environment, leading to inaccurate policy improvement and inefficient exploration. To address this problem, we introduce Critic-Free Pretraining: an efficient paradigm that completely abandons the approach of offline critic training, allowing a freshly initialized critic to adapt without inheriting biased estimates. CFP is compatible with various mainstream O2O algorithms and consistently matches or improves upon conventional O2O algorithms across a diverse set of tasks, with particularly pronounced gains on several challenging tasks.

cs.LG↗

arXiv2026

STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning

Real-world robot learning increasingly relies on heterogeneous data, but demonstrations and rollouts often mix useful progress with stalls, corrections, and suboptimal behavior. Effective policy learning therefore requires frame-level advantages that distinguish reliable local progress from failures and regressions. We propose Self-supervised Temporal Ensemble Advantage Modeling (STEAM), a label-free method that learns such advantages from expert demonstrations. STEAM trains an ensemble of temporal-offset predictors on frame pairs within expert trajectories, using the normalized temporal offset between two frames as a self-supervised signal. Each predictor maps a frame pair to a distribution over temporal offsets, which is converted into a scalar advantage. STEAM then takes the minimum advantage across the ensemble to score mixed-quality rollout data conservatively. Across real-world bimanual towel folding, chip checkout, cola restocking, and single-arm pick-and-place tasks, STEAM identifies stalls, failures, and recoveries. When combined with CFGRL, STEAM further improves policy success rate by 59%, 54.3%, 23% and 16.2% over baselines, respectively.

cs.RO↗

arXiv2026

CXR-ContraBench: Benchmarking Negated-Option Attraction in Medical VLMs

When a chest X-ray shows consolidation but the question asks which finding is present, a medical vision-language model may answer "No consolidation." This is more than an incorrect choice: it is a polarity reversal that emits a clinical statement contradicting the image. We study this failure as negated-option attraction, where a model is drawn to a negated answer option even when it conflicts with both the visual evidence and the question. We introduce CXR-ContraBench (Chest X-Ray Contradiction Benchmark), a diagnostic benchmark spanning internal ReXVQA slices and external OpenI and CheXpert protocols. The benchmark centers on present-finding questions, where selecting "No X" despite visible X creates the main clinical risk, and uses absent-finding questions as secondary tests of whether models copy negated wording. Across CheXpert protocols, the failure is substantial and persistent. On a strict direct presence probe, MedGemma and Qwen2.5-VL reach only 31.49% and 30.21% accuracy, respectively; on a matched 135,754-record CheXpert training-split protocol, both models select negated options on over 62% of presence questions. Chain-of-thought prompting reduces some presence-side reversals but does not eliminate them and can amplify absence-side contradictions. Finally, QCCV-Neg (Question-Conditioned Consistency Verifier for Negation) deterministically repairs the measured polarity-confused subset without retraining, raising MedGemma and Qwen2.5-VL to 96.60% and 95.32% accuracy on the direct presence probe. These results show that standard accuracy can hide a clinically meaningful inference-time polarity failure. Source code and benchmark construction scripts are available at https://github.com/fangzr/cxr-contrabench-code.

cs.CV↗

arXiv2026

SAC Flow: Sample-Efficient Reinforcement Learning of Flow-Based Policies via Velocity-Reparameterized Sequential Modeling

Training expressive flow-based policies with off-policy reinforcement learning is notoriously unstable due to gradient pathologies in the multi-step action sampling process. We trace this instability to a fundamental connection: the flow rollout is algebraically equivalent to a residual recurrent computation, making it susceptible to the same vanishing and exploding gradients as RNNs. To address this, we reparameterize the velocity network using principles from modern sequential models, introducing two stable architectures: Flow-G, which incorporates a gated velocity, and Flow-T, which utilizes a decoded velocity. We then develop a practical SAC-based algorithm, enabled by a noise-augmented rollout, that facilitates direct end-to-end training of these policies. Our approach supports both from-scratch and offline-to-online learning and achieves state-of-the-art performance on continuous control and robotic manipulation benchmarks, eliminating the need for common workarounds like policy distillation or surrogate objectives.

cs.RO↗

arXiv2026

SpaceWire-based Data Acquisition Network for the Solar Flare Sounding Rocket Experiment FOXSI-4 and FOXSI-5

We developed a SpaceWire-based data acquisition (DAQ) system for the FOXSI-4 and FOXSI-5 sounding rocket experiments, which aim to observe solar flares with high sensitivity and dynamic range using direct X-ray focusing optics. The FOXSI-4 mission, launched on April 17, 2024, achieved the first direct focusing observation of a GOES M1.6 class solar flare with imaging spectroscopy capabilities in the soft and hard X-ray energy ranges, using a suite of advanced detectors, including two CMOS sensors, four CdTe double-sided strip detectors (CdTe-DSDs), and a Quad-Timepix3 detector. To accommodate the high photon flux from a solar flare and these diverse detector types, a modular DAQ network architecture was implemented based on SpaceWire and the Remote Memory Access Protocol (RMAP). This modular architecture enabled fast, reliable, and scalable communication among various onboard components, including detectors, readout boards, onboard computers, and telemetry systems. In addition, by standardizing the communication interface and modularizing each detector unit and its associated electronics, the architecture also supported distributed development among collaborating institutions, simplifying integration and reducing overall complexity. To realize this architecture, we developed FPGA-based readout boards (SPMU-001 and SPMU-002) that support SpaceWire communication for high-speed data transfer and flexible instrument control. In addition, a real-time ground support system was developed to handle telemetry and command operations during flight, enabling live monitoring and adaptive configuration of onboard instruments in response to the properties of the observed solar flare. The same architecture is being adopted for the upcoming FOXSI-5 mission, scheduled for launch in 2026.

astro-ph.IM↗

arXiv2025

Policy Newton Algorithm in Reproducing Kernel Hilbert Space

Reinforcement learning (RL) policies represented in Reproducing Kernel Hilbert Spaces (RKHS) offer powerful representational capabilities. While second-order optimization methods like Newton's method demonstrate faster convergence than first-order approaches, current RKHS-based policy optimization remains constrained to first-order techniques. This limitation stems primarily from the intractability of explicitly computing and inverting the infinite-dimensional Hessian operator in RKHS. We introduce Policy Newton in RKHS, the first second-order optimization framework specifically designed for RL policies represented in RKHS. Our approach circumvents direct computation of the inverse Hessian operator by optimizing a cubic regularized auxiliary objective function. Crucially, we leverage the Representer Theorem to transform this infinite-dimensional optimization into an equivalent, computationally tractable finite-dimensional problem whose dimensionality scales with the trajectory data volume. We establish theoretical guarantees proving convergence to a local optimum with a local quadratic convergence rate. Empirical evaluations on a toy financial asset allocation problem validate these theoretical properties, while experiments on standard RL benchmarks demonstrate that Policy Newton in RKHS achieves superior convergence speed and higher episodic rewards compared to established first-order RKHS approaches and parametric second-order methods. Our work bridges a critical gap between non-parametric policy representations and second-order optimization methods in reinforcement learning.

cs.LG↗

arXiv2025

Bidirectional Soft Actor-Critic: Leveraging Forward and Reverse KL Divergence for Efficient Reinforcement Learning

The Soft Actor-Critic (SAC) algorithm, a state-of-the-art method in maximum entropy reinforcement learning, traditionally relies on minimizing reverse Kullback-Leibler (KL) divergence for policy updates. However, this approach leads to an intractable optimal projection policy, necessitating gradient-based approximations that can suffer from instability and poor sample efficiency. This paper investigates the alternative use of forward KL divergence within SAC. We demonstrate that for Gaussian policies, forward KL divergence yields an explicit optimal projection policy -- corresponding to the mean and variance of the target Boltzmann distribution's action marginals. Building on the distinct advantages of both KL directions, we propose Bidirectional SAC, an algorithm that first initializes the policy using the explicit forward KL projection and then refines it by optimizing the reverse KL divergence. Comprehensive experiments on continuous control benchmarks show that Bidirectional SAC significantly outperforms standard SAC and other baselines, achieving up to a $30\%$ increase in episodic rewards, alongside enhanced sample efficiency.

cs.LG↗

arXiv2024

Energetic Electrons Accelerated and Trapped in a Magnetic Bottle above a Solar Flare Arcade

Where and how flares efficiently accelerate charged particles remains an unresolved question. Recent studies revealed that a "magnetic bottle" structure, which forms near the bottom of a large-scale reconnection current sheet above the flare arcade, is an excellent candidate for confining and accelerating charged particles. However, further understanding its role requires linking the various observational signatures to the underlying coupled plasma and particle processes. Here we present the first study combining multiwavelength observations with data-informed macroscopic magnetohydrodynamics and particle modeling in a realistic eruptive flare geometry. The presence of an above-the-loop-top magnetic bottle structure is strongly supported by the observations, which feature not only a local minimum of magnetic field strength but also abruptly slowing down plasma downflows. It also coincides with a compact hard X-ray source and an extended microwave source that bestrides above the flare arcade. Spatially resolved spectral analysis suggests that nonthermal electrons are highly concentrated in this region. Our model returns synthetic emission signatures that are well matched to the observations. The results suggest that the energetic electrons are strongly trapped in the magnetic bottle region due to turbulence, with only a small fraction managing to escape. The electrons are primarily accelerated by plasma compression and facilitated by a fast-mode termination shock via the Fermi mechanism. Our results provide concrete support for the magnetic bottle as the primary electron acceleration site in eruptive solar flares. They also offer new insights into understanding the previously reported small population of flare-accelerated electrons entering interplanetary space.

astro-ph.SR↗

arXiv2024

Episodic energy release during the main- and post-impulsive phase of a solar flare

When and where the magnetic field energy is released and converted in eruptive solar flares remains an outstanding topic in solar physics. To shed light on this question, here we report multi-wavelength observations of a C9.4-class eruptive limb flare that occurred on 2017 August 20. The flare, accompanied by a magnetic flux rope eruption and a white light coronal mass ejection, features three post-impulsive X-ray and microwave bursts immediately following its main impulsive phase. For each burst, both microwave and X-ray imaging suggest that the non-thermal electrons are located in the above-the-loop-top region. Interestingly, contrary to many other flares, the peak flux of the three post-impulsive microwave and X-ray bursts shows an increase for later bursts. Spectral analysis reveals that the sources have a hardening spectral index, suggesting a more efficient electron acceleration into the later post-impulsive bursts. We observe a positive correlation between the acceleration of the magnetic flux rope and the non-thermal energy release during the post-impulsive bursts in the same event. Intriguingly, different from some other eruptive events, this correlation does not hold for the main impulse phase of this event, which we interpret as energy release due to the tether-cutting reconnection before the primary flux rope acceleration occurs. In addition, using footpoint brightenings at conjugate flare ribbons, a weakening reconnection guide field is inferred, which may also contribute to the hardening of the non-thermal electrons during the post-impulsive phase.

astro-ph.SR↗

arXiv2024

GeoLocator: a location-integrated large multimodal model for inferring geo-privacy

Geographic privacy or geo-privacy refers to the keeping private of one's geographic location, especially the restriction of geographical data maintained by personal electronic devices. Geo-privacy is a crucial aspect of personal security; however, it often goes unnoticed in daily activities. With the surge in the use of Large Multimodal Models (LMMs), such as GPT-4, for Open Source Intelligence (OSINT), the potential risks associated with geo-privacy breaches have intensified. This study develops a location-integrated GPT-4 based model named GeoLocator and designs four-dimensional experiments to demonstrate its capability in inferring the locational information of input imageries and/or social media contents. Our experiments reveal that GeoLocator generates specific geographic details with high accuracy and consequently embeds the risk of the model users exposing geospatial information to the public unintentionally, highlighting the thread of online data sharing, information gathering technologies and LLMs on geo-privacy. We conclude with the broader implications of GeoLocator and our findings for individuals and the community at large, by emphasizing the urgency for enhanced awareness and protective measures against geo-privacy leakage in the era of advanced AI and widespread social media usage.

cs.CY↗

arXiv2023

Thermal Evolution of an Active Region through Quiet and Flaring Phases as Observed by NuSTAR XRT, and AIA

Solar active regions contain a broad range of temperatures, with the thermal plasma distribution often observed to peak in the few millions of kelvin. Differential emission measure (DEM) analysis can allow instruments with diverse temperature responses to be used in concert to estimate this distribution. NuSTAR HXR observations are uniquely sensitive to the highest-temperature components of the corona, and thus extremely powerful for examining signatures of reconnection-driven heating. Here, we use NuSTAR diagnostics in combination with EUV and SXR observations (from SDO/AIA and Hinode/XRT) to construct DEMs over 170 distinct time intervals during a five-hour observation of an alternately flaring and quiet active region (NOAA designation AR 12712). This represents the first HXR study to examine the time evolution of the distribution of thermal plasma in an active region. During microflares, we find that the initial microflare-associated plasma heating is dominantly heating of material that is already relatively hot, followed later on by broader heating of initially-cooler material. During quiescent times, we show that the amount of extremely hot (>10 MK) material in this region is significantly (~3 orders of magnitude) less than that found in the quiescent active region observed in HXRs by FOXSI-2 (Ishikawa et al. 2017). This result implies there can be radically different high-temperature thermal distributions in different active regions, and strongly motivates future HXR DEM studies covering a large number of these regions.

astro-ph.SR↗

arXiv2023

The need for focused, hard X-ray investigations of the Sun

Understanding the nature of energetic particles in the solar atmosphere is one of the most important outstanding problems in heliophysics. Flare-accelerated particles compose a huge fraction of the flare energy budget; they have large influences on how events develop; they are an important source of high-energy particles found in the heliosphere; and they are the single most important corollary to other areas of high-energy astrophysics. Despite the importance of this area of study, this topic has in the past decade received only a small fraction of the resources necessary for a full investigation. For example, NASA has selected no new Explorer-class instrument in the past two decades that is capable of examining this topic. The advances that are currently being made in understanding flare-accelerated electrons are largely undertaken with data from EOVSA (NSF), STIX (ESA), and NuSTAR (NASA Astrophysics). This is despite the inclusion in the previous Heliophysics decadal survey of the FOXSI concept as part of the SEE2020 mission, and also despite NASA's having invested heavily in readying the technology for such an instrument via four flights of the FOXSI sounding rocket experiment. Due to that investment, the instrumentation stands ready to implement a hard X-ray mission to investigate flare-accelerated electrons. This white paper describes the scientific motivation for why this venture should be undertaken soon.

astro-ph.IM↗

arXiv2022

Observations of magnetic reconnection and particle acceleration locations in solar coronal jets

We present a multi-wavelength analysis of two flare-related jets on November 13, 2014, using data from SDO/AIA, RHESSI, Hinode/XRT, and IRIS. Unlike most coronal jets where hard X-ray (HXR) emissions are usually observed near the jet base, in these events HXR emissions are found at several locations, including in the corona. We carry out the first differential emission measure (DEM) analysis that combines both AIA (and XRT when available) bandpass filter data and RHESSI HXR measurements for coronal jets, and obtain self-consistent results across a wide temperature range and into non-thermal energies. In both events, hot plasma first appeared at the jet base, but as the base plasma gradually cooled, hot plasma also appeared near the jet top. Moreover, non-thermal electrons, while only mildly energetic, are found in multiple HXR locations and contain a large amount of total energy. Particularly, the energetic electrons that produced the HXR sources at the jet top were accelerated near the top location, rather than traveling from a reconnection site at the jet base. This means that there was more than one particle acceleration site in each event. Jet velocities are consistent with previous studies, including upward and downward velocities around ~200 km/s and ~100 km/s respectively, and fast outflows of 400-700 km/s. We also examine the energy partition in the later event, and find that the non-thermal energy in accelerated electrons is most significant compared to other energy forms considered. We discuss the interpretations and provide constraints on mechanisms for coronal jet formation.

astro-ph.SR↗

arXiv2022

On the faintest solar coronal hard X-rays observed with FOXSI

Solar nanoflares are small eruptive events releasing magnetic energy in the quiet corona. If nanoflares follow the same physics as their larger counterparts, they should emit hard X-rays (HXRs) but with a rather faint intensity. A copious and continuous presence of nanoflares would deliver enormous amounts of energy into the solar corona, possibly accounting for its high temperatures. To date, there has not been any direct observation of such sustained and persistent HXRs from the quiescent Sun. However, Hannah et al. in 2010 constrained the quiet Sun HXR emission using almost 12 days of quiescent solar-off-pointing observations by RHESSI. These observations set upper limits at $3.4\times 10^{-2}$ photons$^{-1}$ s$^{-1}$ cm$^{-2}$ keV$^{-1}$ and $9.5\times 10^{-4}$ photons$^{-1}$ s$^{-1}$ cm$^{-2}$ keV$^{-1}$ for the 3-6 keV and 6-12 keV energy ranges, respectively. Observing feeble HXRs is challenging because it demands high sensitivity and dynamic range instruments in HXRs. The Focusing Optics X-ray Solar Imager (FOXSI) sounding rocket experiment excels in these two attributes. Particularly, FOXSI completed its third successful flight (FOXSI-3) on September 7th, 2018. During FOXSI-3's flight, the Sun exhibited a fairly quiet configuration, displaying only one aged non-flaring active region. Using the entire $\sim$6.5 minutes of FOXSI-3 data, we constrained the quiet Sun emission in HXRs. We found $2σ$ upper limits in the order of $\sim 10^{-3}$ photons$^{-1}$ s$^{-1}$ cm$^{-2}$ keV$^{-1}$ for the 5-10 keV energy range. FOXSI-3's upper limit is consistent with what was reported by Hannah et al., 2010, but FOXSI-3 achieved this result using $\sim$1/2640 less time than RHESSI. A possible future spacecraft using FOXSI's concept would allow enough observation time to constrain the current HXR quiet Sun limits further or perhaps even make direct detections.

astro-ph.SR↗

arXiv2020

Experiments of Federated Learning for COVID-19 Chest X-ray Images

AI plays an important role in COVID-19 identification. Computer vision and deep learning techniques can assist in determining COVID-19 infection with Chest X-ray Images. However, for the protection and respect of the privacy of patients, the hospital's specific medical-related data did not allow leakage and sharing without permission. Collecting such training data was a major challenge. To a certain extent, this has caused a lack of sufficient data samples when performing deep learning approaches to detect COVID-19. Federated Learning is an available way to address this issue. It can effectively address the issue of data silos and get a shared model without obtaining local data. In the work, we propose the use of federated learning for COVID-19 data training and deploy experiments to verify the effectiveness. And we also compare performances of four popular models (MobileNet, ResNet18, MoblieNet, and COVID-Net) with the federated learning framework and without the framework. This work aims to inspire more researches on federated learning about COVID-19.

eess.IV↗

arXiv2020

COVID-19 Public Opinion and Emotion Monitoring System Based on Time Series Thermal New Word Mining

With the spread and development of new epidemics, it is of great reference value to identify the changing trends of epidemics in public emotions. We designed and implemented the COVID-19 public opinion monitoring system based on time series thermal new word mining. A new word structure discovery scheme based on the timing explosion of network topics and a Chinese sentiment analysis method for the COVID-19 public opinion environment is proposed. Establish a "Scrapy-Redis-Bloomfilter" distributed crawler framework to collect data. The system can judge the positive and negative emotions of the reviewer based on the comments, and can also reflect the depth of the seven emotions such as Hopeful, Happy, and Depressed. Finally, we improved the sentiment discriminant model of this system and compared the sentiment discriminant error of COVID-19 related comments with the Jiagu deep learning model. The results show that our model has better generalization ability and smaller discriminant error. We designed a large data visualization screen, which can clearly show the trend of public emotions, the proportion of various emotion categories, keywords, hot topics, etc., and fully and intuitively reflect the development of public opinion.

cs.IR↗