A M Zhang

Hy Vision Team, Huawen Shen, Zhengyang Tang et al.

As large multimodal models move from understanding content to operating on digital environments, mobile GUI has emerged as a challenging and consequential testbed for digital embodied intelligence. Mobile agents operate under three coupled constraints: precise perception of complex interfaces, scalable acquisition of high-quality interaction data, and robust long-horizon decision making under compounding execution errors. This report presents HyMobileAgent, a mobile GUI agent built on Hy3.0-VL-A3B, a vision-native foundation model featuring native any-resolution input, an A3B-scale deployment budget, and a 32K context window to model extended interaction histories. Rather than relying solely on model scaling, we develop a joint data and environment centric scaling framework to address the key bottlenecks of mobile interaction. Our framework integrates a GUI perception flywheel combining mock-interface synthesis, rejection sampling, and icon-specific augmentation; a knowledge pipeline that transforms tutorial videos into structured interaction data; a million-scale action data pipeline deployed across more than 2000 sandbox and real-device instances with automated failure attribution; the PhoneWorld Mock App Factory, providing a resettable training environment with 34 mock applications and over 34000 tasks; and a structured Planning-and-Reflection mechanism with explicit dead-loop detection for reliable long-horizon execution. We also introduce a progressive training recipe consisting of mid-training, supervised fine-tuning, and reinforcement learning with task-specific reward designs.

Bohan Liu, Wenqian Ye, Guangzhi Xiong et al.

Models trained via Contrastive Language-Image Pretraining (CLIP) serve as the foundational vision encoders for most modern Large Vision Language Models (LVLMs). Despite their widespread adoption, CLIP models exhibit a critical yet underexplored failure mode: irrelevant text appearing within images confounds visual representations, biasing them toward lexical meaning rather than true visual semantics. This robustness issue, commonly described as a Typographic Attack (TA), exposes a vulnerability that poses a significant risk to safety-critical applications such as autonomous driving. To achieve interpretable and effective robustness against TA, we propose a novel, training-free mechanistic interpretability method. Our method provides sampling-based interpretations of hidden state representations and quantitatively attributes semantic versus lexical focus to individual attention heads. Through probabilistic analysis and circuit mining, we isolate specific Vision Transformer (ViT) components that disproportionately encode lexical information, thereby identifying the mechanistic source of TA. We further show that simple interventions applied directly to the identified circuits, without any additional training, can substantially improve robustness against Typographic Attacks in object classification. These interventions, such as selective adjustment of attention weights, also outperform both supervised and training-free defense methods. Our experiments demonstrate that applying the proposed intervention to the vision encoders of several state-of-the-art LVLMs yields substantial gains in Visual Question Answering accuracy under Typographic Attack interference on RIO-Bench. These results confirm both the efficacy and the generalizability of our mechanistic approach. Code is released at https://github.com/Liu-524/SamplingTAR.

Fangzheng Li, Aimin Zhang, Chen Lv

Tool Calling and Structured Output are two core capabilities of modern Agent systems, yet their interaction under joint deployment conditions remains insufficiently understood. This paper reports a reproducible phenomenon observed in a production Agent system: when Tool Calling and JSON Schema constraints are simultaneously enabled, multiple open-weight models cease invoking tools despite maintaining high schema compliance. We refer to this behavior as Tool Suppression. Through controlled experiments across multiple model families and deployment settings, we consistently reproduce Tool Suppression under joint constraints, while tool execution and schema compliance remain functional when evaluated independently. Further analysis reveals that JSON Schema constraints are compiled into grammar-based token masks, causing tool-call tokens to become unreachable during decoding. This provides an implementation-level explanation for the observed behavior. To interpret the phenomenon, we formulate the Constraint Priority Inversion (CPI) hypothesis, which suggests that schema satisfaction may dominate action-selection behavior under multiple simultaneous constraints. We present CPI as a behavioral hypothesis consistent with the observed evidence rather than a verified internal mechanism. To mitigate the problem, we propose Transparent Two-Pass Execution, an inference-time strategy that decouples tool execution from schema-constrained response generation. Experimental results show that this approach restores tool invocation while preserving structured output guarantees without requiring model retraining. These findings suggest that evaluating tool use and structured output separately may overlook important reliability issues in production Agent systems. Code, data, and docs will be released at https://github.com/Fzsama/Constrain-Tax-26-06.git.

Nathan Senyard, Salem Hamdani, Astrid Zhang et al.

We propose Hedgementation: a new benchmark to evaluate machine learning models for hedgerow mapping from remote sensing data at country scale and 10m$^2$ spatial resolution. We combine and harmonize multiple remote sensing data products and ground truth labels sourced from a hedgerow inventory in France. We measure the ability of three baseline models to generalize across spatial distance, and across climatic zones, a more explicitly challenging task. Our benchmark tests both supervised and self-supervised learning approaches for remote sensing, applied to tracking fine-scale features of high agricultural importance. The code to reproduce the benchmark and baselines results is available at https://github.com/hedgementation/hedgementation.

Aoyu Zhang, Dongping Liu, Luyao Zhang

Quantum computing promises transformative advances across science and industry, yet the physical hardware that enables these computations remains invisible to the public: quantum processors operate inside sealed dilution refrigerators at temperatures near absolute zero, making direct observation impossible. This "imagination gap" between quantum computing's growing societal impact and the public's ability to visualize it represents a significant barrier to quantum literacy and workforce development. We present Quantum Cinema, an open-source, browser-based interactive application that closes this gap by transforming invisible quantum hardware into explorable, cinematic experiences using generative world models. Quantum Cinema guides users through a four-act narrative -- from the foundational Nobel Prize-winning science of quantum entanglement, through curated video introductions to three major quantum computing architectures (trapped-ion, neutral-atom, and superconducting systems), into immersive three-dimensional generative worlds that make invisible quantum phenomena observable, and finally to interactive radar-chart comparisons grounded in real quantum device specifications. All three-dimensional environments are generated using WorldLabs' generative world model platform and are scientifically grounded in curated metrics from Amazon Web Services (AWS) Braket quantum hardware. Quantum Cinema requires no installation, no specialized hardware, and no quantum computing background. It is designed to serve two distinct communities: scholars and developers seeking to replicate or extend the platform, and educators, researchers, and science communicators seeking an intuitive tool for explaining quantum hardware to diverse audiences. This paper describes the system architecture, the generative world model pipeline, use cases for both communities, and directions for future work.

Andrew Zhang, Chengzhan Li

Most agent evaluations collapse a multi-step trace into a final answer, a success flag, or a trajectory-level score. These aggregates obscure the diagnostic question developers need most: which action changed the state in a useful direction? We introduce Agent Step Value (ASV), a state-transition measurement framework that scores each observed action by the change it induces in a state-grounded evaluator's distribution over fixed candidate outcomes. ASV renders redacted before/after state projections, uses a stateless LLM evaluator to assign candidate log scores, and reports both gold-free belief diagnostics and offline oracle validation metrics. A label-free rationale pass separates evaluator deliberation from one-token option scoring, preserving candidate likelihoods while exposing leakage and floor-score events. On 100 reviewed open-QA evidence-seeking tasks with live PubMed retrieval, a partially live DeepSeek actor, and DeepSeek log-probability scoring, ASV evaluates 1,100 steps and 2,200 states. Under the fixed-layout rationale-conditioned protocol, mean gold-margin gain is -2.335 (trajectory-bootstrap 95\% CI [-3.395, -1.272]), entropy movement is 0.000, and mean Bayesian surprise is 2.693. ASV therefore localizes constructive and destructive belief pivots that final-answer scores and entropy-only step metrics miss. We release the standalone ASV Eval toolkit.

Ao Zhang, Bingyin Zhang, Hongfei Fu

In this paper, we present a linearly implicit, second-order block-centered finite difference (BCFD) prediction-then-projection scheme for the multi-species Keller-Segel chemotaxis system on non-uniform spatio-temporal grids. The proposed scheme integrates a standard Crank-Nicolson time-marching algorithm with an $L^2$ projection step to enforce positivity and mass conservation. The use of variable time stepsize and time-staggered discretization fully decouples the solutions of the multi-species cell density variables and the chemoattractant concentration variable while facilitating linearization, thereby greatly enhancing computational efficiency. Notably, the variable time-stepping algorithm and non-uniform grid BCFD discretization jointly enable adaptive resolution and local refinement near blow-up, thereby improving efficiency and accuracy without compromising the desired physical property-preserving in the simulation. Furthermore, using the mathematical induction method and the energy analysis approach, the unique solvability of the proposed scheme is rigorously proved, and we show that cell densities achieve second-order convergence in both time and space in the discrete $L^2$ norm, while the chemoattractant concentration achieves second-order convergence in the discrete $H^1$ norm. Representative numerical experiments are presented to validate the theoretical findings and demonstrate the reliability of the proposed scheme in simulating the blow-up phenomenon.

Jintao Xu, Yingzheng Ma, Jiong Dong et al.

Balance-oriented multi-warehouse inventory allocation is a recurring decision problem in large-scale e-commerce supply chains, in which a fixed replenishment quantity is distributed across warehouses to balance post-allocation inventory coverage while accounting for demand forecasts and heterogeneous allocation constraints. In practice, allocation requirements are often scenario-dependent and expressed in semi-structured or natural-language form rather than as ready-to-solve operations research (OR) formulations. We propose an OR-guided Large Language Model (LLM) for Allocation (ORLA) that uses solver feedback to generate, verify, and select OR formulations. ORLA integrates automatic "Problem-Model-Code (PMC)" generation, learning-based formulation selection, and feasibility restoration. We develop three complementary mixed-integer programming formulation families based on deviation minimization, soft band compliance, and knapsack-inspired allocation, together with solver-ready mixed-integer linear programming reformulations, modular constraint extensions, and a penalty-based relaxation mechanism for infeasible cases. The LLM component generates candidate formulations and executable solver code from textual or semi-structured specifications, while the solver provides verification signals for executability, feasibility, and solution quality. To address instance heterogeneity, ORLA estimates the expected quality of candidate formulations, selects promising candidates, and combines their outputs through score-aware aggregation. Experimental results on 29 production evaluation batches from JD.com show that the best single OR formulation improves allocation accuracy by 3.4 percentage points over the incumbent approach, while the full ORLA framework achieves a 4.5 percentage-point overall improvement and improves allocation accuracy in 26 of the 29 evaluation batches.

Alec Dewulf, Dhruv Pai, Li Yang et al.

We show that for tall matrix parameters, like projection matrices in the MLP layers, the Muon update can have row norms that are arbitrarily non-uniform. This can lead to a self-reinforcing feedback loop whereby neurons receive persistently small updates and eventually do not contribute meaningfully to network outputs. This problem is effectively mitigated by an additional row normalization step, but current methods do this in a way that moves the Muon update geometry away from the polar factor of the momentum matrix, which we find is undesirable. We propose Aurora, an optimizer that enforces row-uniformity of matrix parameter updates while respecting Muon's polar factor geometry. Aurora outperforms Muon in our pre-training experiments and, when combined with existing methods, achieves state-of-the-art performance among spectral optimizers on the optimizer track of the modded-nanoGPT speedrun. Additionally, we find that Aurora's empirical gains over Muon scale with the MLP expansion factor, suggesting that Aurora may allow for effective training of very wide MLP layers.

Hanzhi Chen, Anran Zhang, Simon Schaefer et al.

A central question in robot learning is how to acquire skills from the kinds of data that humans learn from: passive observation, embodied practice, and the experience of failure. Human videos provide the first of these in abundance, and prior work has shown they can initialize useful policies. Far less clear is whether they can support the second and third: whether priors extracted from human videos can ground a robot's own attempts well enough to evaluate them, correct them, and improve from them. In this work, we show that human videos can be used to learn embodiment-agnostic action, dynamics, and value representations that transfer across robot embodiments, providing the predictive foundation required for robots to autonomously improve from their own rollouts and failures. We introduce Dynamics-Guided Action Correction (DGAC), a training-free approach that leverages these adapted models to repair failed states: each failure becomes a query for which the learned models propose and rank corrective actions, turning failures into supervision for the next policy update. Across seven real-world manipulation tasks spanning both a mobile manipulator and a static manipulator arm, our approach improves success rates from 40% to 81% across multiple policy backbones, demonstrating cross-embodiment robot self-improvement from human-video priors. These results show that human priors and robot failures can be combined to enable scalable autonomous policy improvement. Project page: https://ethz-mrl.github.io/robot-self-improvement-website/.

Zhengyang Tang, Xin Lai, Pengyuan Lyu et al.

Phones are becoming an important execution surface for general-purpose agents, but training open models for reliable phone use remains difficult because the environment that matters at deployment, real devices running real apps, is slow, stateful, side-effectful, and hard to reset or verify, while scalable mock environments only approximate real behavior. We present PhoneBuddy, a training recipe and open-model line for agentic phone use that combines a real-app environment with a mock-app environment, PhoneWorld, which reconstructs runnable mock apps from real GUI usage structure. PhoneBuddy first builds a shared supervised fine-tuning stage from trajectories collected in both environments, then compares real-app RL against mixed RL across both environments. Across a 150-task human evaluation on real phones spanning apps, mini-apps, and cross-app workflows, task success rate improves from 36.67\% after supervised fine-tuning to 40.67\% after real-app RL and 45.33\% after mixed RL. On AndroidWorld, the same progression rises from 60.3\% to 77.2\% to 83.2\%. These results show that mock-app training is not a replacement for real-app RL, but a complementary source of scalable, resettable, and automatically checked interaction. The gains are strongest on app and mini-app tasks, while long-horizontal cross-app workflows remain an important open challenge.

Alex Zhang, Jackson Craig, Lara Waldrop

Spacecraft operations scheduling is a highly constrained, long-horizon combinatorial optimization problem that traditionally relies on heuristics, constraint programming, or manual planning. We present a scalable deep reinforcement learning framework developed and deployed for NASA's Carruthers Geocorona Observatory mission. Our framework introduces a macro-action abstraction known as activity blocks coupled with dynamic action-masking to navigate the intractably large search space and strictly enforce complex power, thermal, and instrument constraints. The resulting architecture generates globally feasible schedules with overwhelming probability, establishes operational trust, and executes a full training cycle in under six hours, circumventing the need for policy robustness by enabling rapid, on-demand retraining. Further, resulting schedules outperform baseline heuristics in scheduled science quality. The deep reinforcement learning framework was deployed as the default operational scheduler for the Carruthers Geocorona Observatory mission from the outset of the mission, demonstrating that deep reinforcement learning can be trusted for real spacecraft operations under complex, evolving constraints.

Anping Zhang, Yang Tan, Yuanbo Tang et al.

Human collective participation is rarely steady in time: it is bursty, with short episodes of intense activity separated by long quiet intervals. In crisis response and community mobilization, predicting when people act matters as much as predicting whether they act. Such settings are increasingly modeled with LLM-based social simulators, yet these simulators are validated on whether each action is individually plausible, not on whether actions are timed as in reality. Their temporal realism, the degree to which simulated activity reproduces the bursty, heavy-tailed timing of real human systems, thus remains untested. We examine this gap using a multi-year, city-scale log of offline volunteering in Shenzhen that spans the COVID-19 pandemic. Empirically, we establish that bursty timing is common at individual and tracked-group levels, that it is largely endogenous and self-exciting, and that it is amplified by the pandemic rather than produced by daily activity cycles. A standard LLM-only simulator reproduces almost none of this timing: its synchronous schedule has no self-excitation channel, so agents act on a near-regular clock. Guided by these findings, we build a simulator in which a data-calibrated self-excitation channel and a crisis-period regime decide when each agent acts and query the LLM only at those moments, leaving it to decide which task to join and whether to commit. The LLM-only baseline yields no bursty agents (median burstiness $B=-0.14$); a single data-calibrated gate is then sufficient to lift per-agent timing above the burst threshold (median $B\approx0.37$) without degrading LLM content decisions. These results indicate that temporal realism in LLM-based crisis-response simulation is best achieved by decoupling when agents act, governed by an explicit self-excitation and crisis-activation mechanism, from what they do, governed by the LLM.

Anqi Zou, Han Deng, Chengyu Zhang et al.

Current computer-use benchmarks primarily focus on software operation tasks in virtualized systems, whereas scientific instrumentation scenarios require coordinated control over complex interfaces, and feedback-driven parameter adjustment. However, directly evaluating agents on physical high-precision instruments is impractical due to high cost, safety risks, limited accessibility, and difficulty in ensuring reproducible evaluation. This motivates the need for a simulated yet realistic testbed that preserves the operational challenges of scientific instruments while enabling scalable and safe benchmarking. To this end, we introduce LabOSBench, a challenging benchmark for multimodal GUI agents built on a suite of web-based scientific-instrument simulators. Operating directly via a browser, LabOSBench avoids resource-heavy OS virtualization while supporting flexible task configuration and execution-based evaluation. Specifically, LabOSBench constructs 96 subtasks across eight instrument simulators, covering workflows from sample loading, alignment, parameter tuning, and data acquisition to result inspection. We evaluate general-purpose vision-language models, specialized GUI agent models, and advanced agentic frameworks at both subtask and end-to-end levels. Our experiments reveal that while existing agents can complete many structured GUI subtasks, they still struggle with feedback-driven operations and long-horizon workflow execution. Overall, LabOSBench provides a reproducible, low-cost testbed for advancing computer-using agents toward scientific-instrument control.

Haoyang Fang, Wei Zhu, Boran Han et al.

RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction matters because fixed schedules commit all parameters to fixed trajectories and therefore cannot express the non-stationary exploration-exploitation tradeoffs that regularization must track; the principle provides actionable design rules for multi-stage training. We discover this through LLMZero, a system where LLM agents search over training trajectories via tree search, diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across 4 diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9% to 140% relative and over grid search by 6% to 15% relative, consistently outperforming random search and the skill-based agent. The structural principle transfers across tasks, providing an explanation for why discovered strategies take qualitatively different forms yet share similar parameter dynamics.

Haoyang Zhong, Yifei Sun, Antong Zhang et al.

Retrieval-Augmented Generation (RAG) has emerged as a paradigm for enhancing large language models (LLMs) with external knowledge, yet existing graph-based methods face a fundamental limitation: entity-centric and chunk-centric approaches operate on representations anchored to original text without true knowledge fusion. While entity-centric methods connect logically related content and chunk-centric methods preserve context, both retrieve information separately through similarity search, missing emergent understanding from their synthesis. In this paper, we propose HyGRAG, a hierarchical graph RAG framework that transcends source documents by addressing three core challenges: constructing summaries that genuinely integrate contextual and relational information, leveraging these synthesized representations to access emergent knowledge during retrieval, and efficiently updating hierarchical structures for dynamic corpora. Specifically, we design hierarchical index structures over hybrid graphs with both chunk and entity nodes, then iteratively cluster them and generate LLM-based summaries. Then, we design context and relation-aware retrieval that searches across all abstraction levels while expanding through community membership. Moreover, we enable dynamic knowledge update through attachment-based algorithms with only local re-summarization. Experimental results show that HyGRAG improves the average accuracy of multi-hop reasoning tasks by 9.7%, while maintaining reasonable efficiency.

Chenxin Li, Zhengyao Fang, Zhengyang Tang et al.

Phone agents are increasingly expected to complete real mobile workflows rather than merely predict the next screen action. However, much of the current mobile-agent literature still evaluates agents primarily as GUI controllers that observe a screen, emit taps and swipes, and are scored by target app state. Real phone-use tasks are broader: they require deciding when to use app GUIs, device-side commands, or structured tools, while leaving evidence that the intended side effect actually occurred. We introduce PhoneHarness, a mixed-action benchmark and execution harness for studying phone-use agents on verifiable mobile workflows. PhoneHarness runs a device-side agent loop over GUI, CLI, and host-side tool actions, combining deterministic action routing with bounded GUI delegation and auditable execution traces. Its benchmark, PhoneHarness Bench, evaluates whether agents complete tasks with observable side effects, not only whether they produce plausible final answers. On the annotated evaluation split, PhoneHarness reaches a 75.0% pass rate, outperforming the strongest non-PhoneHarness settings by 12.9 percentage points. PhoneHarness and PhoneHarness Bench therefore play distinct but mutually dependent roles: the harness makes mixed phone workflows executable, while the benchmark measures whether agents can use that harness reliably and safely. Our findings suggest that reliable phone automation depends on action-surface routing and verifiable execution, not only visual GUI control.

Arthur Zhang, Carl Qi, Donne Su et al.

Open-world mapless navigation from sparse language instructions requires resolving underspecified goals and inferring which environmental cues are relevant for reaching the goal. For instance, reaching an out-of-view destination may require interpreting ramps, signs, or detours that reveal where to go or which route to take. Prior works are limited by their reliance on known navigation factors and closed-set factor categories, or identify cues before motion planning and miss plan-dependent cues. We argue that pretrained Vision-Language Models (VLMs) can discover novel instruction-relevant cues, but require adaptation to focus on which cues matter and how they should influence motion planning. We realize these ideas in Foresight, a test-time framework in which a finetuned VLM alternates between proposing image-space motion plans and critiquing them using the language goal and visual context. Subsequent plans are conditioned on prior critiques, enabling iterative motion refinement before execution. To align plan critiques and refinements with open-set behavior preferences, we learn a reward model from human feedback and use it to post-train the VLM with reinforcement learning in the plan-critique loop. In offline evaluations and 6 real-world environments, Foresight improves average task success by 37% and reduces interventions per mission by 52% relative to state-of-the-art test-time reasoning and foundation-model baselines, while running in real-time on a Jetson AGX Orin. We will release code, data, and training details to support future work on test-time reasoning for robot motion refinement. Additional videos at: https://amrl.cs.utexas.edu/foresight

Zheyuan Zhang, Zehao Wen, Alvin Zhang et al.

For agents to learn continuously from interaction with the world at test time, they must be able to explore effectively, acquire new world knowledge and skills, retain relevant episodic experiences, and plan over long horizons. To evaluate these key abilities of test-time continual learning agents, we introduce AgentOdyssey, a novel evaluation framework that procedurally generates open-ended text games with rich entities, world dynamics, and long-horizon tasks. Critically, AgentOdyssey goes beyond the conventional machine learning assumption that learning does not occur at test time by placing agents in a continuous, long-horizon setting that interleaves learning and inference throughout deployment. We further propose a multifaceted evaluation methodology that measures not only game progress but also offers diagnostic tests on world knowledge acquisition, episodic memory, object and action exploration, action diversity, and model cost. We evaluate diverse agent paradigms in the generated games. Our experimental results reveal critical limits in agents' key abilities, as well as factors that influence their meaningful horizon. Although performance scales with stronger base models, even the top agent remains far below human performance, leaving substantial headroom for improvement. Among agent mechanisms, we find that short-term memory benefits multiple agent paradigms and is an important component of agent test-time training.