arXiv subjects

Yu Wang

Publications and source records attributed to Yu Wang.

At least 19 recordsLinked to original sources

arXiv2026

SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations

Large language models exhibit complementary strengths, motivating routing methods that dispatch each query to the most suitable model. Although existing routers are effective in single-turn settings, they do not directly transfer to multi-turn dialogue, where routing performance critically depends on how historical context is segmented, retained, and incorporated into the current prompt. This introduces two fundamental challenges: preventing information loss and information confusion during context construction, and evaluating routing quality without conflating model selection with prompt construction quality. In this paper, we propose SWRouter, a Similarity-Contractive Window Router for multi-turn large language model routing. SWRouter combines a similarity-based context segmentation mechanism for prompt construction with a dual-metric evaluation framework that decouples construction accuracy from router performance. Experiments on multi-turn dialogue benchmarks demonstrate that SWRouter consistently surpasses strong baselines, achieving a 16.26% improvement in evaluation accuracy over the best individual large language model and an additional 8.22% gain over the Conv-ID Context baseline. Our results highlight that multi-turn large language model routing requires a joint design of context construction and evaluation, rather than a direct extension of single-turn routing methods.

cs.CL↗

arXiv2026

$S^3$-Bench: Evaluating Speech Interaction Models as Scientific Voice Assistants

The advance of multimodal large language models (MLLMs) has fundamentally reshaped the paradigm of human-computer interaction, especially speech interaction models capable of seamless conversations. Despite remarkable performance as general voice assistants, their performance in specialized domains remains underexplored, particularly in scientific areas. Scientific interactions introduce formidable challenges, involving rare technical terminology, spoken norms of abbreviations, and the natural verbalization of symbolic special expressions. In this paper, we introduce S$^3$-Bench, a systematic evaluation framework covering 10 major disciplines, consisting of a Knowledge set for speech question-answering and a Dialogue set for multi-turn progressive interactions with simulated user agents. By decomposing a complete atomic turn into stages of speech recognition, perception, knowledge utilization with reasoning, and response pronunciation, we systematically characterize the common challenges and performance tradeoffs of existing approaches. Furthermore, experiments on multi-turn interactions reveal persistent limitations in user adaptation and the generation of accurate, comprehensive, and efficient responses.

cs.CL↗

arXiv2026

Structural Process Supervision for Latent Chain-of-Thought Reasoning

Latent reasoning approaches enhance token-level efficiency and robustness by replacing verbose, explicit chain-of-thought (CoT) tokens with compact continuous-space embeddings. However, existing methods lack direct process supervision over these latent embeddings, which often leads to representation collapse and uneven information distribution. To address this, we propose Prototype-Mediated Process Supervision (PMPS), which introduces learnable reasoning prototypes as semantic anchors to provide structural process-level supervision for latent reasoning. PMPS projects latent embeddings and explicit CoT embeddings into a shared prototype space, achieving many-to-many soft alignment between unequal-length representations through prototype assignment. Meanwhile, we introduce a Progressive Sequential Alignment (PSA) module to further guide training: positional priors initially encourage sequential alignment structure, then gradually relax to permit adaptive matching. Experimental results show that PMPS compresses output token length to under 50% of explicit CoT on GSM8K-Aug. Compared to leading baseline SIM-CoT, our method achieves average accuracy gains of 2.08% across different model families. On GPT-2, PMPS even surpasses CoT-SFT. On larger models and a more challenging task, PMPS consistently attains the highest accuracy among all latent reasoning methods with comparable output length.

cs.AI↗

arXiv2026

RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibility

Vision--language--action (VLA) models acquire broad generalization through large-scale pretraining, yet adapting them to a new task and robot embodiment still requires post-training on newly collected data. Unlike pretraining, post-training targets task- and embodiment-specific adaptation, making it particularly sensitive to data quality. In practice, collected robot datasets often contain heterogeneous errors, including execution mistakes, sensor drift, and timestamp misalignment, which can impair post-training and policy performance. Manual inspection is costly, while existing data-cleaning methods are typically tailored to particular corruption types. To address these challenges, we introduce \textsc{RoboDrop}, a data-curation framework that audits supervision using local gradient compatibility measured along the training trajectory as a proxy for its effect on post-training performance. During a one-epoch warm-up run, RoboDrop scores each candidate sample online by comparing its gradient with those of task-semantic and visually matched validation samples. The resulting sample scores are aggregated at the episode level, and a simple automatic post-processing rule converts them into filtering decisions. We evaluate RoboDrop on controlled observation--action corruptions, naturally suboptimal demonstrations in simulation, and real-robot datasets containing non-expert collection errors. Across these settings, RoboDrop more accurately distinguishes unreliable demonstrations than prior methods, while post-training on the curated data consistently yields stronger downstream policies, with average real-robot rollout success rising from $35.0\%$ to $67.5\%$. These results establish training-trajectory-aware, context-conditioned supervision auditing as an effective approach to robust VLA post-training.

cs.RO↗

arXiv2026

Negative contacts in genus one: a comparison of punctured and root stack Gromov-Witten theories

Let $D$ be a smooth divisor in a smooth projective complex variety $X$. For connected curves of arithmetic genus one with prescribed signed contact orders, we prove that the refined punctured cycle of Battistella--Nabijou--Ranganathan and the negative-contact cycle of Fan--Wu--You agree after pushforward to the common moduli space of stable maps with divisor evaluations. Thus the pushed-forward refined punctured cycle is the constant coefficient of the pushed-forward root-stack virtual class, normalized by one power of the root order for each negative contact. The key step is a comparison for the universal target. After restricting to finite-type open substacks determined by the fixed pair $(X,D)$ and numerical data $\Gamma$, we prove that the positive BNR space maps finitely and with generic degree one onto Crumplin's main component. Using Crumplin's genus-one component description and degree formulas, we identify this component's fundamental cycle with the constant coefficient of the universal orbifold virtual class under comparison of root orders. Refined zero-section pullback recovers the negative contacts, and compatible virtual pullbacks and root-forgetting pushforwards transfer the resulting identity to $(X,D)$.

math.AG↗

arXiv2026

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.

cs.CL↗

arXiv2026

VANTAGE-Bench: Evaluating the Infrastructure AI Gap in Vision-Language Models

As Vision-Language Models (VLMs) advance toward physical deployment, the focus has remained on action-oriented Embodied AI evaluated on subject-centric consumer video. This overlooks a pervasive class of Physical AI: Infrastructure AI, which relies on fixed cameras for open-loop insights like safety monitoring and operational logging. We introduce VANTAGE-Bench, a benchmark measuring this "Infrastructure AI Gap." It spans three operational domains (Logistics, Transportation, and Smart Spaces), unifies image and video evaluation across semantic, spatial, temporal, and spatio-temporal capabilities, and moves beyond multiple-choice to eight task formulations including dense captioning and spatio-temporal grounding. It adds a single-pass trajectory protocol for Single Object Tracking and, to our knowledge, the first such evaluation on fixed-camera infrastructure video, scored against specialist trackers. Annotation spans three regimes over 3,346 media assets: 3,342 video-task annotations, 4,281 image-grounding annotations, and 27,404 detection boxes. Evaluating 17 models zero-shot, we find the shortfall relative to consumer-centric benchmarks is concentrated, not general. Event verification, referring expressions, and temporal localization fall roughly 9 to 24 points at every model scale, while video question answering stays within 5.3 points of VideoMME and 2D spatial pointing shows no shortfall against BLINK. The temporal pillar is weakest in absolute terms: no system exceeds 55.7 mIoU on temporal localization or 37.3 SODA_c on dense video captioning. On tracking, frontier models come within roughly 5 points of specialist trackers over short horizons but separate as the horizon extends. Open-weight models lead 2D object localization outright, so neither scale nor proprietary access explains the pattern. Data, evaluation harness, and leaderboard: https://vantage-bench.org/

cs.CV↗

arXiv2026

A Single Fixed Shallow Circuit for Classical Shadows of Arbitrary n-Qubit States

Classical shadows extract quantum properties from reusable classical records, typically obtained through randomized measurement settings. Although individual settings may be shallow, switching among them introduces control, calibration, and reconfiguration costs beyond conventional metrics. Here we construct, for every $n$, a single fixed shallow quantum analyzer whose Born outcomes replace externally sampled settings as labels for reusable shadow snapshots. The analyzer combines a freshly prepared $n$-qubit fiducial register $A$ with parallel Bell readout of $A$ and the unknown system $S$, producing one $2n$-bit record per copy. The same circuit realizes a rank-one minimal informationally complete measurement: one fixed setting replaces the $3^n$ local-Pauli settings conventionally used for complete reconstruction while retaining the minimum $d^2$ outcomes with $d=2^n$. Its Pauli-diagonal frame admits an analytic inverse. For fixed Hermitian observables, the Haar-averaged conditional variance has a dimension-independent coefficient, whereas the state-uniform coefficient is dimension dependent. For $n\ge3$, worst-state Pauli variances remain bounded in the axial sector and scale as $\Theta(d)$ in the mixed sector. Fiducial preparation before system contact uses $n-1$ arbitrary two-qubit gates, no work qubits beyond $A$, and logarithmic depth under all-to-all connectivity. The unknown system undergoes one parallel system-ancilla entangling layer followed by local Hadamards and readout. Measurement-assisted preparation achieves $O(1)$ adaptive quantum depth using $n+1$ extra qubits in the heralded route, or $O(n\log n)$ auxiliaries for deterministic completion. These tradeoffs show that part of the measurement-setting randomness and control complexity normally supplied shot by shot can instead be compiled into a fixed reusable analyzer.

quant-ph↗

arXiv2026

Carleman Estimates for Wave Equations on the Half-Line: AI-Assisted Weights and Applications

We develop an AI-assisted search-and-certification workflow for constructing Carleman weights for wave equations on semi-infinite domains. Starting from a weighted cross-term identity used to derive the Carleman estimate, we derive explicit analytical screening conditions and use an AI system to propose symbolic candidate weights subject to those conditions. The search identifies a logarithmic weight structure with concrete parameter values; the subsequent human verification by the authors then retains this structure, derives the admissible parameter range, and rigorously certifies the selected weight by verifying the required quantitative conditions. Based on the identified and certified Carleman weight, the subsequent proofs, analysis, and applications are carried out entirely by the authors: we establish a global Carleman estimate for the wave operator on the half-line and derive a weighted conditional lateral Cauchy stability result. The proposed weight is further used in a finite-depth one-sided reconstruction problem for semilinear wave equations on the half-line, where the initial-time flux induces an endpoint graph stabilizer and a contractive frozen nonlinear reconstruction map. Numerical experiments support the predicted contraction behavior and show improved robustness to noisy data from the Carleman weighting, with the induced endpoint graph term providing further stabilization at higher noise levels.

math.AP↗

arXiv2026

An Exact Engine for Black-Hole Jets

For fifty years the Blandford--Znajek mechanism has been a mechanism and not an exact solution: force-free electrodynamics on a prescribed metric, with the jet's own field carrying no weight. Here that field gravitates. A split monopole weighs as much as a magnetic monopole, so its self-gravity is the magnetic Reissner--Nordstr\"om geometry; two copies glued across an equatorial current sheet put hole, disk and jet-driving flux into one spacetime, and force-free plasma makes a magnetosphere of it. Three things then follow that no test-field calculation can see. First, a steady jet is impossible. A stationary horizon cannot be heated but a slipping magnetosphere necessarily heats it, so the only stationary state is a dead one corotating with the hole, and a working jet is a black hole in decay at rates the field equations fix rather than assume. Second, flux enters the laws of black-hole mechanics as a charge, sharing one extremality budget with spin. That budget caps the jet power at $c^5/48G$ whatever the mass. Third, the lifetime output is finite: a maximally spinning hole delivers $1-e^{1/4}/\sqrt2=9.2\%$ of its mass and grows its horizon area by exactly $\sqrt e$. The horizon's own moment of inertia vanishes with the irrational exponent $(\sqrt{17}-1)/2$ set by the near-horizon throat, and what is left is a rotating hole whose angular momentum has passed to its own field.

astro-ph.HE↗

arXiv2026

Knowing When to Stop: Adaptive Action Chunking via Internal Cross-Attention Dynamics in VLAs

Action chunking is a standard execution strategy in modern Vision-Language-Action (VLA) frameworks, but fixed execution horizons impose a trade-off between efficiency and accuracy. Short chunks require frequent inference and may cause oscillatory behavior, whereas long chunks can become misaligned with newly observed states. We address this limitation with an adaptive action chunking approach based on internal cross-attention dynamics in the action expert. We observe that, as the prediction horizon extends, action-to-observation cross-attention becomes increasingly dispersed and its entropy rises toward a plateau. This pattern is associated with higher action prediction error and provides an online signal that the current observation offers limited grounding for further open-loop execution. Based on this observation, we introduce a training-free truncation mechanism that detects sustained high-entropy plateaus and dynamically selects the execution horizon during inference. The method uses attention weights already computed by the policy and introduces negligible additional overhead. Evaluations on $\pi_{0.5}$ and X-VLA across RoboTwin 2.0, LIBERO, and three real-world manipulation tasks show improved average task success over fixed-horizon and adaptive chunking baselines, while preserving efficient closed-loop control. These results show that cross-attention dynamics can provide a practical internal signal for adaptive action execution in VLAs.

cs.RO↗

arXiv2026

DualDiff3D: Dual Structure-Appearance Diffusion Priors for Reliability-Enhanced 3D Gaussian Splatting

While 3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction and novel-view synthesis, scenarios with limited input views often lead to poor reconstruction quality and artifacts in rendered novel views. Recent efforts attempt to utilize powerful diffusion priors, yet they typically process rendered and reference views concatenated along an additional dimension in a single network. These methods overlook an inherent nature that different views should maintain appearance similarity but differ in structure due to view shifts, leading to blur caused by conflicts between the two properties. In this paper, we propose DualDiff, a novel pipeline that leverages dual diffusion priors with a Structure-Appearance Attention (SAA) module to introduce reference guidance for refining low-quality novel views rendered from flawed 3D representations. Specifically, we retain one diffusion branch to focus on extracting structural information from the low-quality novel views, while introducing another branch to ensure appearance consistency with reference views. Furthermore, we present a 3D reconstruction framework named DualDiff3D, which integrates a reliability-enhanced Render-Refine-Optimize (RRO) loop to progressively and robustly incorporate the refined novel views, yielding more accurate 3DGS. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods even in the inference-only setting, with further performance gains achievable through training. Our code and pre-trained weights are available at https://github.com/Akaneqwq/DualDiff3D.

cs.CV↗

arXiv2026

Human-Anchored Factuality Evaluation with Strategic Annotation

LLM-based factuality judges provide scalable evaluation signals, but their metrics are often systematically biased relative to human judgments. We study human-anchored factuality evaluation under limited annotation budgets, where judge predictions on the full dataset are combined with human labels on a small selectively sampled subset to obtain statistically valid estimates. The efficiency of this approach depends critically on which examples receive human annotation: in factuality evaluation, judge-human misalignment is not driven solely by low confidence, but also by structured failure modes such as incomplete evidence, temporal mismatch, unverifiable claims, and rubric misalignment. To exploit this structure, we introduce a factuality-specific annotation policy design pipeline that uses failure-space analysis (FSA) to derive diverse predictive signals for modeling human-judge misalignment. On an internal reference-based factuality evaluation system (AutoFA) and RAGTruth, where judge-predicted estimates substantially underestimate human-annotated factual accuracy, our FSA-guided policy improves annotation efficiency over uniform sampling and uncertainty-driven baselines, achieving effective-sample-size gains of 40.3% on AutoFA and 27.1% on RAGTruth.

cs.CL↗

arXiv2026

Target-Aware State-Adaptive $p$-Dirichlet Graph Neural Regression for Non-Invasive Body-Composition Estimation

Accurate estimation of body-composition outcomes, including body fat percentage (BFP), bone mineral density (BMD), and appendicular lean mass (ALM), is important for evaluating metabolic, skeletal, and muscular health. Direct assessment using dual-energy X-ray absorptiometry (DXA), however, requires specialized equipment and involves ionizing radiation. We propose a target-aware, state-adaptive $p$-Dirichlet energy-flow graph neural regression ($p$SADE-GNR) framework for estimating these outcomes from non-invasive anthropometric measurements. A neural encoder maps participant representations to hidden states that are propagated over an outcome-specific participant-similarity graph by a state-adaptive forward-Euler discretization of the graph $p$-Dirichlet energy flow. Graph distances weight each original or latent coordinate by its normalized absolute training-fold correlation with the outcome. Using clinical data from the Pennington Biomedical Research Center and five-fold cross-validation, the correlation-weighted model using the original standardized measurements achieved the lowest root mean squared error in all nine primary outcome-cohort combinations and outperformed previously reported support vector regression or least-squares support vector regression reference values in eight of nine comparisons. Autoencoder, variational-autoencoder, and Gaussian-mixture variational-autoencoder representations generally did not improve primary-outcome prediction or reduce computational cost. In an exploratory age-prediction analysis including ALM, BMD, and BFP as predictors, the correlation-weighted GMVAE model achieved the lowest mean error in all three cohorts. These results support target-aware, state-adaptive $p$-Dirichlet graph neural regression for non-invasive body-composition estimation.

cs.LG↗

arXiv2026

When Tool Outputs Become Commands: Separating Action Induction from Runtime Authorization in Tool-Augmented LLM Agents

Tool-augmented LLM agents must rely on untrusted runtime Observations to complete open-ended tasks; however, when tool outputs no longer merely provide data but begin to specify concrete actions, they effectively become ``commands'' that can drive real-world side effects beyond user intent. We argue that this risk arises from conflating action induction with execution authorization. To address this distinction, we propose SARA, which treats action induction and execution authorization as distinct runtime roles and separates action provenance from execution authority. On the Observation side, a context-isolated Action Probe exposes action-inducing semantics and persistently records action-origin provenance across steps as a review signal; on the execution side, actual tool calls are authorized only against the user objective and audited evidence from authorized successful executions, while satisfying goal, execution-chain, and argument-level support. To preserve this separation across multi-step execution, SARA applies No-History-Promotion to prevent historical recurrence from laundering action origins into execution authority. Across AgentDojo and AgentDyn, SARA limits ASR to no more than $0.63\%$ across four primary evaluation settings while maintaining competitive task utility, and consistently reduces ASR across additional Agent backbones.

cs.AI↗

arXiv2026

PolyMemDB: A Polyglot Database System for AI Memory Management

With the widespread adoption of personal intelligent agents, users generate massive, heterogeneous data during long-term interactions. Leveraging this data as long-term memory helps reduce token overhead and deliver personalized experiences. However, existing memory systems face two primary limitations: they rely on single-storage paradigms that fragment multi-dimensional data, and they lack fine-grained data provenance to resolve long-term factual conflicts, thereby worsening LLM hallucinations. In this demonstration, we introduce PolyMemDB, a novel system tailored for managing agent memory. PolyMemDB has a polyglot storage architecture designed to track and manage various memory types, including graph, vector, probability and spatial-temporal data. To ensure factual consistency and reduce hallucinations, it features a probabilistic inference engine that integrates temporal decay with semiring aggregation, resolving long-term factual conflicts, providing detailed data provenance, and enabling users to trace reasoning chains transparently.

cs.DB↗

arXiv2026

Upper H\"olderian with Explicit Exponent of Solution Mapping with Applications to Ball Constrained Least Squares Problems

In this paper, we propose an extension of the well-known Robinson implicit function theorem for generalized equations from the upper Lipschitzian case to the upper H\"olderian case. Explicit exponents dependence between the generalized equation and its linearization is determined. Applications to ball constrained least squares problems, including linear least squares and separable nonlinear least squares, are studied. In particular, we establish that the solution mapping of ball constrained linear least squares under linear perturbation is locally upper H\"older continuous with exponent $1/3$, which is of independent interest. Ultilizing the upper H\"olderian version of the implicit function theorem, we show the local upper H\"olderian of the solution mapping of parametric ball constrained linear least squares.

math.OC↗

arXiv2026

GRB 220101A: a most energetic $10^{54}$ erg long GRB triggered by two supernovae 3.5 seconds apart

GRB 220101A is a long GRB, with a total energy exceeding $10^{54}$ erg with a redshift $z = 4.61$ and one of the largest ever high-quality multi-wavelength observational coverage, from a large number of space-based and ground-based telescopes. We interpret this source in a doubly Binary driven peta nova (BdP-N) model. The progenitor is composed of a massive CO core of $\sim 10\,M_\odot$, highly magnetized with $B \sim 10^{6}$ G, associated to a neutron star (NS) and a white dwarf (WD) with orbital periods from minutes to hours. The large GRB luminosity is explained by a sequence of 7 episodes: episode 1 is triggered by a new kind of pair supernova (HB) which originates from the collapse of the strongly magnetized CO core. Accretion of the HB supernova ejecta (the ejecta) onto the white dwarf companion triggers after 3.5 sec the episode 2: the second supernova emitting neutrinos and creating a new neutron star ($\nu$NS). The ejecta, interacting with the magnetosphere of the binary NS companion originate the episode 3: the Ultra relativistic Prompt Emission (UPE) emission by far the most energetic episode of this GRB, with the formation of a powerful jet normal to the plane of the GRB. Following the UPE energy loss, the accretion of the ejecta on the NS companion leads to the episode 4: the formation of a black hole (BH) of $2.3 \ M_\odot$ leading to the observed GeV afterglow emission. Further accretion of the ejecta spin up the $\nu$NS to a period of $1.3$ ms which gives origin to the episode 5: the birth of a pulsar. The interaction of this milli-second pulsar with the remnants lead to the Episode 6: the synchrotron emission observed in the X-ray, optical and radio, The episode 7 is a 56.7 ms pulsar, as observed $10^{10}$ s after the first burst in the crab nebula.

astro-ph.HE↗