Searcharxiv⌕ Search

arXiv subjects

Jun Wang

Publications and source records attributed to Jun Wang.

At least 19 recordsLinked to original sources
arXiv2026

ScamWorld: An Interactive Environment for Evaluating Web Agents Against Online Fraud

Web agents increasingly browse, shop, and manage accounts on behalf of their users. Prior work has shown that they can be misled by adversarial content, including prompt injections and deceptive interfaces. Beyond such attacks, agents operating on the open web also encounter online fraud designed to deceive humans, including fraudulent online stores, fake investment platforms, and advance-fee schemes. How web agents respond to such fraud during task execution remains underexplored, in part because scam websites are short-lived, making systematic evaluation difficult. To address this gap, we introduce ScamWorld, an interactive environment that reconstructs reported scam websites across diverse scenarios, fraud mechanisms, and languages. Each reconstructed website reproduces documented fraud mechanisms, making relevant signs of fraud observable before the agent's final action and enabling controlled, repeatable evaluation. Through an extensive study of web agents powered by six frontier LLMs, we find that baseline deception rates range from 11% to 35% across models. Agents proceed more readily with payments than with requests for credentials or wallet information. A closer analysis of agent trajectories shows that some agents acknowledge suspicious conditions yet proceed, at times treating the user's instructions as authorisation to accept additional terms imposed by the website. ScamWorld provides a reusable testbed for systematically evaluating and stress-testing web agents against real-world online fraud.

cs.AI↗
arXiv2026

From topology-optimization brief to solver-checked release: separating intent, admissibility and acceptance knowledge for language agents

An engineering solver can converge even when a language agent has misread the user's problem. AgenticTO makes this distinction inspectable by keeping three records separate: the proposed problem specification, deterministic formulation checks, and numerical acceptance of the solved design. We evaluated the workflow on 50 author-built topology-optimization briefs, injected faults, five numerical problems, and complete agent sessions. The tests used linear-elastic models with one load case. Geometry tools raised the primary provider's correct specifications and refusals from 37 to 44 of 50, while also increasing wrong specifications delivered as finished. Input-format feedback alone recovered most of the accuracy gain. The wrong specifications passed the formulation rules, which also missed most injected faults that changed the intended problem. Clarification corrected the remaining errors when an oracle supplied reference facts, but also prompted questions on fully specified briefs. Perfect accuracy under that oracle therefore does not establish performance with engineers. Numerical trials reached 73.7 million elements, although released designs used finer grids than requested. The records expose how a proposed problem becomes an accepted numerical result; agreement with the brief still needs a separate judgment.

cs.CE↗
arXiv2026

Can Large Language Model Agents Learn by Watching Tutorial Videos?

Humans continuously acquire knowledge from carefully crafted materials such as books and tutorial videos, carrying that understanding into future tasks. Millions of such instructional videos are uploaded online every day, yet LLM agents remain unable to exploit this ever-expanding wealth of human teaching. Their capabilities are frozen after training, while parameter updates at scale remain prohibitively expensive. This raises the central question: can LLM agents learn from this vast instructional resource without parameter updates? We introduce Tutorial Agentic Learning (TAL), a nonparametric continual-learning paradigm that consolidates tutorial videos into a persistent, structured tutorialbook memory through a two-stage adversarial process that harvests hard, video-grounded questions and uses them to refine the memory. At inference time, TAL retrieves through a Table-of-Contents (TOC) tree, tying retrieval directly to the model's reasoning rather than to embedding similarity. To study this setting, we introduce VidLearn, a multimodal QA benchmark built from 72 chess tutorial videos spanning 12 openings. On VidLearn, TAL outperforms all evaluated baselines and improves monotonically as more videos are consumed, achieving up to 113% and 91% relative gains over the corresponding vanilla MLLMs on the training and test splits, respectively. In actual gameplay, TAL reaches a stronger opening-phase position than its vanilla counterpart in 69% of head-to-head games. Evaluation in a second domain further demonstrates the generality of TAL. Together, these results highlight instructional videos as a scalable resource for parameter-update-free agent continual learning.

cs.LG↗
arXiv2026

A/B Agent: A Self-Evolving Agent for Strategy Iteration in Industrial A/B Testing

Industrial recommendation strategy iteration heavily relies on large-scale A/B experimentation. Traditional tuning requires experts to repeatedly design strategies, configure experiments, analyze results, and adjust parameters, making the process labor-intensive and time-consuming. Meanwhile, valuable knowledge from historical experiments is often fragmented, making systematic reuse difficult through manual expert effort alone. Existing RAG agents partially alleviate this burden by retrieving prior strategies, but typically organize experience in a flat manner, overlooking the hierarchical relationships among business scenarios, recommendation stages, optimization objectives, and experimental contexts. This often results in mismatched retrieval and limited cross-scenario transfer, while preventing agents from continuously refining strategies and parameters through sequential A/B feedback. To address these limitations, we propose \textbf{\ourname}, a closed-loop A/B agent for industrial recommendation strategy optimization. The framework comprises three tightly coupled core components: \textit{Historical Strategy Knowledge Organization}, \textit{Autonomous Target-Aware Strategy Generation}, and \textit{Experiment-Guided Strategy Self-Evolution}. It organizes historical strategies into a hierarchical experience tree, retrieves transferable evidence through multi-path Tree-RAG to generate executable strategies, and continuously analyzes online A/B feedback to guide autonomous tuning and update the experience tree for self-evolution. Extensive offline and online evaluations demonstrate its effectiveness, including a \textbf{4.829%} improvement in GMV in a real-world short-video e-commerce recommendation system while maintaining positive gains across all guardrail metrics.

cs.AI↗
arXiv2026

AVSG: Accelerated Vectorized Sparse Gather for Efficient KV Cache Offload in Sparse-Attention LLM Serving

Dynamic sparse attention reduces long-context attention computation by selecting only a subset of tokens, but still requires access to the full KV cache, leaving serving memory-bound. Offloading the KV cache to host memory reduces device memory pressure but places H2D transfers on the decoding critical path. In DSA, substantial overlap in selected KV entries across decoding steps creates an opportunity for device-resident reuse. Exploiting this reuse efficiently, however, presents three critical challenges: costly matching of selected tokens against entries retained in the HBM buffer, uneven distribution of the remaining H2D transfers across accelerator cores, and retaining frequently accessed KV entries within limited HBM capacity. We present AVSG, an operator that addresses these challenges for DSA while preserving its exact token selections. AVSG uses vectorized hash matching to identify reusable HBM-buffer slots and reserve slots for entries requiring transfer, then evenly partitions these H2D transfers across accelerator cores. Lifetime-based buffer management retains frequently accessed KV entries in the device buffer across decoding steps, and shared slot reservations extend reuse across tokens within a multi-token prediction iteration. On a single NPU, vectorized hash matching is 2.80x faster than scalar dual-pointer matching, miss-only transfer raises effective H2D bandwidth by up to 34.44x over request-level assignment, and an 8K-entry buffer reaches a 94.83% hit rate on real requests. These gains translate into end-to-end improvement: on a serving stack processing a real-world production dataset, AVSG reduces time per output token by 39% and increases output throughput by 1.27x relative to the same KV offload layout without HBM-buffer reuse. The code of AVSG is released at https://gitcode.com/cann/cann-recipes-infer.

eess.SP↗
arXiv2026

HiGenTrans: Hierarchical LLM Knowledge Transfer for Generative Recommendation

Generative recommendation promises to simplify conventional multi-stage recommender systems by generating item Semantic IDs (SIDs) autoregressively. However, two obstacles limit its effective- ness in industrial settings: interaction-only supervision underuses the semantic knowledge encoded in pretrained language models, while beam-search likelihood ranks token sequences rather than directly estimating item relevance. To address these limitations, we propose HiGenTrans, a three-level framework that introduces LLM knowledge at the input, representation, and decision levels. First, Global Semantic Priming (GSP) adds an LLM-derived user representation to the beginning-of-sequence embedding, provid- ing a semantic prior before SID generation. Second, Hierarchi- cal Representation Transfer (HRT) aligns intermediate LLM representations with depth-matched GenRec layers through mean- centered projections and gated residual connections. Finally, Multi- Signal Decision Calibration (MDC) uses a lightweight mixture- of-experts calibrator to combine semantic relevance, generation confidence, and candidate-set context when reranking beam-search outputs. All semantic features are computed only from the observ- able interaction prefix, preventing target leakage, and can be cached offline. Experiments on three public benchmarks show that HiGen- Trans improves Recall@10 by up to 12.9% and NDCG@10 by up to 33.7% over the strongest baselines. These results demonstrate the value of transferring LLM knowledge at complementary stages of the generative recommendation pipeline.

cs.AI↗
arXiv2026

MAPLE: A Multimodal Mobile Agent with Playbook-Learned Experience

Autonomous agents powered by vision-language models (VLMs) show great potential in navigating mobile graphical user interfaces (GUIs) to complete complex tasks. However, existing paradigms face fundamental scalability bottlenecks: fine-tuning requires massive high-fidelity datasets or prohibitive training time and compute, while gradient-free approaches rely on custom static heuristics or costly trial-and-error trajectory exploration where feedback is deferred until the end of execution. To address these limitations, we introduce MAPLE (Multimodal Agent with Playbook-Learned Experience), a novel gradient-free multi-agent framework that enables dynamic, step-wise experience sharing and execution-time prompt evolution. Unlike traditional action-reflection architectures that depend on costly, post-hoc self-reflection, MAPLE enables agents to cross-reference and distribute runtime insights in real time at each operational step. This dynamic exchange eliminates blind trial-and-error looping and steers agents along direct, grounded interaction paths. Extensive evaluations on the AndroidWorld and MobileWorld benchmarks demonstrate that MAPLE achieves state-of-the-art performance on complex, long-horizon mobile tasks, while proving scalable and adaptable across novel environments.

cs.LG↗
arXiv2026

Bounds for lattice points in terms of volume and successive minima

We prove, in every dimension, the upper bound conjectured by Betke, Henk and Wills for the lattice-point enumerator in terms of volume and successive minima. For every convex body $K\subset R^n$ and full-rank lattice $Λ$, we establish $$ \#(K\capΛ)\le\frac{\mathrm{vol}_n(K)}{\detΛ}\prod_{i=1}^n\left(1+\frac{iλ_i}{2}\right), $$ where $λ_i$ are the successive minima of $(K-K)/2$ with respect to $Λ$. Equality holds precisely for lattice translates of positive integral dilates of unimodular simplices. We also prove $$ \#(\mathrm{int}K\capΛ)\ge\frac{\mathrm{vol}_n(K)}{\detΛ} \left(1-\frac{n}{4}\sum_{i=1}^{n-1}λ_i-\frac{n}{2}λ_n\right), $$ which implies the lower estimate with factor $1-\frac{1}{2}\sum_{i=1}^n iλ_i$. In dimension two we prove the lower bound with factor $(1-λ_1/2)(1-λ_2)$ when $λ_2\le1$. We also observe that a conjectured product lower bound for origin-symmetric bodies fails in dimension three. The proofs combine geometric compression, lattice-cell counts in coordinate sections, and an elementary coefficient comparison.

math.MG↗
arXiv2026

$π$-StepNFT: Wider Space Needs Finer Steps in Online RL for Flow-based VLAs

Flow-based vision-language-action (VLA) models have emerged as a powerful paradigm for embodied control, but multi-step denoising makes policy-gradient online reinforcement learning (RL) difficult because the final-action likelihood induced by the sampler is not directly tractable. We identify an exploration--supervision mismatch within the denoising trajectory, where stochastic differential equation (SDE) sampling explores a wider local stochastic transition space while supervision derived from the final denoised output does not directly use the adjacent transitions observed during sampling. We propose $π$-StepNFT (Step-wise Negative-aware Fine-Tuning), a flow-native online RL framework that couples wider stochastic exploration with finer step-wise supervision without computing the final-action likelihood. $π$-StepNFT treats observed SDE transitions as supervision units, using variance-normalized transition errors to rank mirrored positive and negative velocity branches and map rollout feedback to local velocity updates. On LIBERO, starting from shared few-shot SFT initializations, $π$-StepNFT raises average success from 57.6\% to 92.0\% for $π_0$ and from 77.1\% to 94.0\% for $π_{0.5}$. On ManiSkill, it achieves the highest reported mean success among the evaluated RL baselines on five test-only visual perturbations. Our project homepage is available at https://wangst0181.github.io/pi-StepNFT/.

cs.RO↗
arXiv2026

The inverse problem for Julia limiting directions in the Eremenko--Lyubich class

We study the inverse problem for Julia limiting directions in the Eremenko-Lyubich class $\mathcal B$, which is the class of transcendental entire functions with a bounded set of singular values, and its meromorphic generalization $\mathcal B_{mer}$. Every nonempty compact subset of the circle can be the set of Julia limiting directions of an entire $f\in\mathcal B$. For finite unions of nondegenerate closed arcs, the corresponding $f\in\mathcal B$, even with every finite order at or above an explicit bound determined by the arc lengths, can be attained. We also refine the series construction of the latter two authors to obtain $f\in \mathcal B_{mer}$ with any prescribed order and the set of Julia limiting directions.

math.CV↗
arXiv2026

Heterogeneous Co-Evolution Learning for Tool-Efficient Search Agents with Zero Labels

Deep search agents extend language models by interleaving reasoning with external retrieval. However, effective search-agent training requires addressing two challenges beyond final-answer optimization. First, the capability frontier of agents is largely determined by training questions, whereas fixed or labeled task collections cannot adapt as agents improve. Second, outcome-only reinforcement learning treats all successful trajectories equally, even when some correct answers rely on redundant searches, creating an inherent tension between answer accuracy and tool efficiency. To address these issues, we introduce \textsc{ZEST-Search}, a zero-label and tool-efficient framework for search-agent training via heterogeneous co-evolution. Starting from unlabeled reference questions, \textsc{ZEST-Search} constructs validated search tasks online through an answer-first and evidence-path-guided generation process. A heterogeneous proposer adaptively constructs training tasks around the solver's evolving capability boundary, while the solver learns search-augmented reasoning through reinforcement learning. To jointly optimize task difficulty and search efficiency, \textsc{ZEST-Search} employs an intermediate-difficulty reward for the proposer and a task-local tool-aware reward for the solver, encouraging informative tasks and correct answers with fewer redundant searches. Experiments on seven general-domain and multi-hop question answering benchmarks demonstrate that \textsc{ZEST-Search} consistently improves search-agent performance while achieving higher tool efficiency without labeled search-training data or expert search trajectories. Further analyses show that heterogeneous generation, adaptive difficulty control, and tool-aware optimization enable more effective search-agent training.

cs.LG↗
arXiv2026

EpiAgent: Epistemic Self-Play for LLM Agents

Epistemic game theory explains an agent's strategic behavior via the beliefs it holds about others and how these beliefs are updated as the interactions unfold. Modern LLM-based agents, however, often lack an explicit mechanism for learning and updating such beliefs to guide decisions under incomplete information. To bridge this gap, we propose EpiAgent, a framework that equips LLM-based agents with an explicit, learnable epistemic layer, which maintains beliefs over others' private information and combines these beliefs with the agent's available information into an epistemic state for action selection. Concretely, the agent's policy is decomposed into two coupled stages: Epistemic Update, which updates these beliefs from interaction histories, and Action Planning, which selects actions conditioned on the updated epistemic state. By formulating their dependency as a bi-level optimization problem, we introduce Epistemic Self-Play, an algorithm that alternately optimizes the two stages through self-play, utilizing an intrinsic prediction reward to evaluate how well inferred beliefs predict subsequent observations. Empirical results across both mixed and competitive scenarios show that EpiAgent consistently outperforms prompting-based and RL-based baselines. Further analyses reveal the benefits of dedicated epistemic training, improvements in belief inference accuracy, and the contribution of learned beliefs to action planning. These results demonstrate the value of an explicit, learnable epistemic layer for effective multi-agent interactions.

cs.AI↗
arXiv2026

Calculation and rejection diverge in language models tested at known mechanical limits

When a language model returns engineering quantities with the status "solved", software that trusts it passes them on, even when the requested design is physically impossible. Such failures could reflect poor calculation, how the answer is requested, or a failure to judge feasibility. We set beam loads and disk speeds at known percentages above or below plastic collapse or burst, measured each model's limit calculation separately and tested 12 models in 4,860 calls, including follow-up experiments on new designs. We find that a model can recognize an impossible design when asked only whether it is feasible, yet accept the same design, reporting "solved", when asked for a full answer. Three models accepted beams 10 to 30% beyond collapse despite precise limit calculations in separate questions, and many accepted replies stated the violation. Asked only about feasibility, Opus 5 and Gemini 3.8 Flash rejected every impossible new design they were given, including those accepted in full answers, without false alarms. Rewording when to decline or changing the answer format raised rejection on the original beams, but not consistently on new designs. Software and benchmarks using a model's status should test feasibility judgment and full-answer rejection separately, and report the exact prompt.

cs.CE↗
arXiv2026

LinearARD: Linear-Memory Attention Distillation for RoPE Restoration

The extension of context windows in Large Language Models is typically facilitated by scaling positional encodings followed by Continued Pre-Training (CPT). While effective, this paradigm is notoriously data-hungry and computationally expensive, requiring massive long-text corpora to recalibrate the model to the shifted positional distribution. We propose LinearARD, a restoration distillation method that restores Rotary Position Embedding (RoPE)-scaled students through attention-structure consistency with a frozen native-RoPE teacher. Rather than next-token prediction or opaque hidden-state matching, LinearARD aligns row-wise distributions of dense $Q/Q$, $K/K$, and $V/V$ self-relation matrices from the final attention layer to directly supervise attention dynamics. To remove the quadratic memory bottleneck of $n \times n$ relation maps, we introduce a linear-activation-memory kernel that stores only per-token log-sum-exp statistics and recomputes logits in the backward pass to obtain exact Kullback-Leibler divergence gradients. Across LLaMA2-7B, LLaMA3-8B, and Mistral-7B-v0.1 extended to 32K context, LinearARD recovers 93.1%/94.2%/94.3% of native short-context performance and achieves strong long-context robustness on RULER using only 4.25M tokens---amounting to just 1.6% (a $\sim$60$\times$ reduction) of the 256M-token budget required by state-of-the-art baselines. Under this severely constrained budget, CPT and LongReD remain near zero on RULER, demonstrating that relation-level restoration provides a substantially better efficiency-quality tradeoff. Our code is available at https://github.com/gracefulning/LinearARD.

cs.CL↗
arXiv2026

Consequence-guided clarification: releasing minimum-compliance topology-optimized designs from incomplete specifications

A converged topology optimization can solve the wrong problem when loads, supports, or the domain are underspecified. Assuming defaults, or asking the problem's author about every detail, ignores how much each answer changes the design. We propose consequence-guided clarification (CGC), which bases acceptance on computed mechanical consequences. CGC computes a reference design for each of a finite list of complete specifications, called readings. A design is accepted only if its compliance is at most 5% above the reference at every reading consistent with the answers. Otherwise, CGC asks the question whose worst answer leaves the smallest worst-case excess, and withholds acceptance at its question limit. To save optimizations, the screening variant CGC-S postpones some references until a candidate must be checked. We developed and evaluated both on 36 arXiv-derived tasks, treating each of 134 readings in turn as intended and simulating correct answers. In 26 tasks, the default-reading design failed another reading. Apart from always-asked questions on load direction, sign, and material budget, asking everything took 188 questions; CGC and CGC-S each took 105. The acceptance test supplied most of this saving: applying it to the default design alone took 117. Relative to asking everything, CGC-S requested 4.6 extra reference designs per avoided question, each from two optimization starts, counting optimization time only. References are not certified optima: better ones pushed three accepted cases above tolerance, and finer meshes changed some decisions. Within these limits, the test shows which readings a design covers and which details block acceptance.

cs.CE↗
arXiv2026

Towards operational data-driven regional weather forecasting: extreme-event skill and near-real-time verification with HRRRCast

Improving convective-scale skill in data-driven regional weather models requires better representing rare high-impact events, training at native resolution, broadening variable output, and demonstrating performance in real-time settings. Here, we evaluate these challenges using HRRRCast, a data-driven convection-allowing emulator trained on the High-Resolution Rapid Refresh (HRRR) analysis data. HRRRCast combines native 3-km training, moderate sigmoid-weighted loss reweighting for reflectivity and precipitation, and broader system improvements including expansion of the training archive from 3 to 9 years (2016--2025), increased vertical resolution, expanded outputs, and architectural upgrades. The weighted loss improves high-threshold skill while preserving common-event behavior. Native-resolution training avoids aliasing from 6-km subsampling and substantially improves fractions skill score at convective scales, with gains that persist across pooling windows. HRRRCast has also been running in near-real-time mode for the past year. In seven months of independent observation-based verification, HRRRCast ensemble means improve several near-surface metrics relative to operational HRRR and maintain strong precipitation and reflectivity performance. Single-member reflectivity skill is strongest at 20--30 dBZ; at 40 dBZ, skill is near parity at 6-km pooling and better at pooling radii $\geq$30 km. Precipitation skill remains competitive through 1 in/h and extends to about 3 in/h in the best recent configuration. A focused comparison with HREF further shows useful low-cost probabilistic guidance for 3-h precipitation. At storm scale, HRRRCast exhibits a displacement--structure trade-off, with generally better storm placement than storm morphology.

physics.ao-ph↗
arXiv2026

Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents

Continual learning is the ability of LLM agents to improve over time by acquiring, refining, and applying knowledge from experience while retaining useful prior knowledge. It is essential for agents learning over long time horizons, yet current evaluations largely assess individual task performance rather than improvement through accumulated experience. We introduce Continual Learning Exam (CLE), a large-scale benchmark developed by 50+ domain experts to measure experience-driven improvement in realistic professional workflows. CLE comprises 120 human-authored tasks, each organised as a five-level ladder, yielding 600 levels spanning 89 professional roles across 12 occupational domains. Agents must pass each level to unlock the next, with later levels building on experience accumulated during earlier assignments. An evaluation agent assesses each submission against human-authored rubrics. We compare a single-submission setting with one that provides diagnostic feedback after failure and allows one revision. Across 17 harness-model configurations using frontier LLMs, feedback and revision yield a median completion gain of 5.3 percentage points. Yet the best-performing system clears only 47.7% of levels and completes all five levels in just 23.3% of tasks, highlighting the difficulty of sustained learning from experience for existing frontier LLMs. As such, the dataset can also be used to post-train LLMs to strengthen their continual learning capabilities.

cs.LG↗
arXiv2026

Probing Collision Grounding in Vision-Language Models for Safe Human-Robot Collaboration

Safe human--robot collaboration requires more than visual description: a monitor must determine whether the robot body is safely separated, already colliding with the scene or a person, or about to collide. We call this capability collision grounding: binding visual observations to robot body geometry, camera viewpoint, scene layout, human proximity, and temporal motion in order to infer present and imminent contact. We introduce TouchSafeBench, a physics-grounded benchmark for evaluating collision grounding in vision-language models (VLMs). Built in Habitat~3.0, TouchSafeBench contains 2,940 simulated indoor co-presence episodes across social navigation and social rearrangement, with synchronized multi-view RGB-D observations, top-down trajectory maps, calibrated camera metadata, and simulator-derived contact labels. We study two deployment-facing tasks: classifying the current safety state and warning about imminent collision before contact. Across three frontier or robotics-oriented VLMs and nine visual representations, current models remain far from reliable: the best average Macro-F1 stays below 50\%, explicit depth is not automatically transformed into robot-body collision evidence, and robot--scene contact is consistently harder than human-contact risk. TouchSafeBench reveals a central limitation of embodied VLMs: visual fluency does not imply physical accountability. Reliable robot safety monitors will need representations that explicitly bind viewpoint, robot morphology, metric geometry, and future collision. The benchmark is available at https://jun25336-bit.github.io/TouchSafeBench/ .

cs.CV↗