arXiv subjects

Jian Wang

Publications and source records attributed to Jian Wang.

At least 55 records · Page 3Linked to original sources

arXiv2026

EHR-RobustGym: Benchmarking and Training Agents for Robust Clinical Reasoning

In hospital workflows, electronic health records (EHRs) are often noisy, and may not contain the evidence needed to confirm events or measurements referenced in a clinical query. Even when database retrieval succeeds, clinical agents can overlook such discrepancies and return plausible but unsupported answers. We introduce EHR-RobustGym, a scalable and interactive environment for evaluating and training robust clinical agents grounded in noisy EHRs. Built on MIMIC-IV hospital records (365K patients, 31 tables, and over 500M records), EHR-RobustGym comprises 5,486 Clean-Noise pairs spanning six clinical intents and both patient-level and population-level queries. The pairs test robustness to Record-level, Value-level, and Query-level noise, while interactive SQL/Python execution and outcome verification support trajectory collection and training. Evaluating multiple LLMs reveals substantial robustness gaps: average task success across proprietary and large-scale open-weight models drops from 62.2% on Clean questions to 37.9% on Noise questions. At k=4, pass^k consistency falls below 50% for most evaluated models, exposing instability in clinical task completion. Supervised fine-tuning and reinforcement learning in EHR-RobustGym improve performance, with gains generalizing to five external EHR benchmarks. Together, these results position EHR-RobustGym as a testbed for evaluating and improving the evidence-grounded robustness of clinical agents.

cs.AI↗

arXiv2026

From Imitation to Reward Discovery: On-Policy Warmup for Agentic RL

Reinforcement learning with a verifiable reward (RLVR) offers a scalable approach to training language-model agents, yet sparse outcome rewards can leave early training with little signal for policy improvement. We identify an On-Policy Acceleration Phenomenon: in our main comparisons, RLVR initialized with on-policy distillation reaches high performance earlier in training and achieves both higher average performance during subsequent RLVR and higher final performance than the alternative baselines. Motivated by this observation, we study On-Policy Warmup (OPW), a teacher-guided stage in which the student trains with teacher supervision on its own interaction trajectories before transitioning to RLVR. Unlike imitation on fixed teacher-generated trajectories, OPW targets states induced by the student's own decisions, including imperfect actions and recovery situations. We provide a theoretical explanation by connecting on-policy reverse-KL distillation to trajectory-level distribution matching. Under a competent teacher and sufficiently small population distillation loss, this connection yields a lower bound on initial verifier success and a corresponding bound on reward-discovery complexity. For group-relative RLVR, we further characterize when increased success probability produces more reward-informative groups. Together, our findings support on-policy distillation as an effective warmup for agentic RLVR and identify initial reward discovery as a mechanism that can contribute to the observed acceleration.

cs.LG↗

arXiv2026

Experimental demonstration of broadband-laser suppression of cross-beam energy transfer

In direct-drive inertial confinement fusion (ICF), cross-beam energy transfer (CBET) redirects a significant fraction of the incident laser energy out of the plasma. We report the experimental demonstration that broadband lasers reduce CBET-enhanced reflected-light return, performed at the low-coherence Kunwu laser facility with two crossed beams of 0.6% bandwidth and up to 550J at $\sim$2.6$\times$10$^{14}$ W/cm$^{2}$. A coupled ray-tracing model shows that the strong CBET amplification of narrowband reflected light is much weaker under broadband illumination. In symmetric incidence condition of two orthogonal beams, the total fractional scattered energy decreases from 7.57% to 4.64%; in asymmetric incidence condition, which separates stimulated Brillouin scattering (SBS) from specular reflection, the SBS fraction decreases from 3.75% to 0.46% and the specular-reflection fraction decreases from approximately 2.2% to 1.3%. These results establish broadband lasers as an effective, experimentally validated approach to reducing energy escape through CBET-enhanced reflection in direct-drive ICF.

physics.plasm-ph↗

arXiv2026

Can LLMs Value the Right Evidence? Evidence-Value Misalignment in Dynamic Medical Diagnosis

A correct diagnosis reached from insufficient or misleading evidence can pose a clinical hazard, yet outcome-based accuracy may reward such lucky guesses. We call this mismatch between diagnostic decisions and the value of available evidence Evidence-Value Misalignment (EVM). To disentangle evidential grounding independently from diagnostic accuracy, we introduce MedEVM, a dynamic benchmarking environment comprising 1,050 cases across 24 disease systems. Observations arrive turn by turn, requiring models to continuously calibrate its decision by deciding whether to wait for more evidence or submit a diagnosis. Across 9 LLMs, four interesting patterns are observed. (1) Miscalibrated evidence tracking. Making a diagnosis often fails to calibrate evidence sufficiency, even in more capable models, and even worsens in reasoning mode. (2) Misaligned diagnosis submission. Confidence in the correct diagnosis often fails to ensure timely submission despite sufficient evidence. (3) Evidence order matters. Reordering the same evidence changes diagnoses even when model confidence remains similar. (4) Misleading evidence remains influential. Added misleading evidence redirects diagnoses even after prior evidence becomes sufficient. We further verify that EVM predicts errors and that preventing premature submission improves accuracy. These findings motivate Evidence-Verified Diagnosis Harness (EVD-Harness). It decouples diagnosis generation from submission through an offline Contrastive Diagnostic Wiki and three online control stages, namely observation management, proposal and witness verification, and diagnosis submission control. Across five LLMs, EVD-Harness improves accuracy by 12.0--51.1 percentage points while mitigating EVM-related failures. Our results demonstrate that verifying evidential support before submission can make diagnostic decisions more reliable.

cs.CL↗

arXiv2026

LLM-Based Generative Retrieval for Snapchat Content Recommendation

Pretrained large language models (LLMs) are promising retrieval engines because they combine rich semantic priors, strong sequence modeling capabilities, and favorable scaling behavior. However, turning a pretrained LLM into a generative retriever in production deployment raises several challenges: the model must learn an internal item vocabulary that was absent from pretraining, and generate valid item identifiers under strict latency and cost constraints. We address these challenges through the design and launch of SnapLGR, an LLM-based generative retrieval system for short-video recommendation at Snapchat. The system is built around three main designs. First, we construct semantic identifiers (SIDs) from multimodal item embeddings and enhance them with Personalized PageRank (PPR)-based co-engagement contrastive learning, resulting in improved codebook utilization, reduced collisions, and infused collaborative signal. Second, we use continued pretraining (CPT) to ground the introduced SID tokens before supervised fine-tuning (SFT) on user interaction sequences. Third, we make SnapLGR serving practical through TensorRT-LLM CUDA-backed beam search and a decentralized worker-loop architecture. In a live A/B test, the launched system increased View Time by 0.37%, Time Spent by 0.09%, Deep Sessions by 0.18%, and Deep Sessions Unique User by 0.11% relative to the existing TIGER-style generative retrieval baseline. We then decompose this offline gap under a fixed tokenizer and quantify the gains due to model architecture, scaling, and pretraining. Overall, our deployment shows that successful production SnapLGR requires joint design across representation learning, vocabulary grounding, and efficient training and serving.

cs.IR↗

arXiv2026

Uniruledness and the sign of total scalar curvature

For every integer $n\ge3$, we construct a smooth projective manifold $X$ of complex dimension $n$ whose canonical bundle is not pseudoeffective, or equivalently, which is uniruled, but every Kähler metric has negative total scalar curvature. In particular, $X$ admits no Kähler metric of positive scalar curvature, while it admits a Riemannian metric of positive scalar curvature. Thus the equivalence between uniruledness and the existence of a Kähler metric with positive scalar curvature holds in complex dimensions one and two, but fails in higher dimensions.

math.AG↗

arXiv2026

1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation

Sparse on-policy distillation (OPD) allocates teacher supervision to a small subset of tokens in student-generated trajectories. However, useful teacher guidance can yield a noisy update when its gradient is estimated from a sampled next token. We study this estimation problem at a fixed prefix in information geometry and propose an information-efficiency ratio (IER) based on a signal-to-noise decomposition. IER characterizes relative gradient estimation error under an optimal scalar baseline. A candidate-set approximation enables token selection based on IER and its combination with existing usefulness scores, while retaining the sampled reverse-KL training objective. On mathematical and medical reasoning tasks, adding IER improves existing selectors in multiple settings, with sparse configurations matching or exceeding full OPD without token selection at small token budgets of 0.1\%--1\%. These results support accounting for both usefulness and gradient-estimation reliability when allocating sparse supervision. Our code is available at https://github.com/BruceSheng1202/IER-OPD.

cs.LG↗

arXiv2026

One-forms, spectral Einstein functionals and the noncommutative residue

For two one-forms and the Dirac operator, Dabrowski etc. recovered the spectral Einstein functionals by computing their noncommutative residue in Theorem 4.1 \cite{DL}. In this paper, we generalize the results of Dabrowski etc. to the cases of four dimensional spin manifolds with boundary.

math.DG↗

arXiv2026

CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation

Adapting large language models to individual users remains challenging due to the tension between fine-grained personalization and scalable deployment. We present CARD, a hierarchical framework that achieves effective personalization through progressive refinement. CARD first clusters users according to shared stylistic patterns and learns group-specific LoRA adapters, enabling robust generalization and strong low-resource performance. To capture individual differences within each cluster, we propose an implicit preference learning mechanism that contrasts user-authored text with cluster-level generations, allowing the model to infer user-specific style preferences without manual annotation. At inference time, CARD injects personalization exclusively at decoding via lightweight user preference vectors and low-rank logit corrections, while keeping the base model frozen. Experiments on the LaMP and LongLaMP benchmarks show that CARD achieves superior generation quality compared to baselines, while significantly improving efficiency and scalability for practical personalized text generation.

cs.AI↗

arXiv2026

Do Implicit Personalization and Explicit Styles Conflict? PsPLUG: A Lightweight Plug-in for Balancing Personalization and Style in Customized LLMs

Personalized large language models are often expected to follow explicit style instructions, yet we find that such instructions can undermine the user-specific characteristics that personalization methods aim to preserve. We call this failure mode personalization collapse: explicit style control can conflict with implicit user preferences. To address this challenge, we propose PsPLUG, a lightweight plug-in that learns a user-specific residual after accounting for the requested style. PsPLUG also allows us to tune personalization strength at inference time. Our experiments show that explicit style instructions can diminish personalization in existing methods, whereas PsPLUG better preserves user preferences while providing precise control over the balance between personalization and style adherence.

cs.AI↗

arXiv2026

Curvature-Flux Representation of Local Dirac-Node Topology from Non-Hermitian Zeeman Quantum Geometry

Local Dirac-node topology is usually described by a winding number or Berry phase, whereas curvature-flux formulations are commonly associated with global band topology. Here we show that the non-Hermitian Zeeman quantum geometric tensor (QGT) provides a curvature-flux representation of local $π_1$ topology. Unlike the conventional Hermitian QGT, whose metric and Berry curvature are fixed respectively by its real symmetric and imaginary antisymmetric parts, the Zeeman QGT must be decomposed according to symmetry under index exchange. This decomposition yields normal and anomalous metric-curvature sectors. The normal sector has the same metric-curvature structure as the conventional Hermitian QGT. The anomalous sector, by contrast, contains two components absent in the conventional QGT: an imaginary symmetric metric-like tensor and a real antisymmetric curvature-like tensor. In a two-dimensional Dirac system, the anomalous Zeeman curvature forms a radial flux field Hodge-dual to the tangential winding field of the Dirac node, thereby converting the local winding invariant into a Gauss-type flux invariant. We further show that the four Zeeman-geometric sectors map one-to-one onto frequency- and symmetry-resolved gyrotropic conductivity channels, with reciprocal kinetic magnetoelectric response providing a complementary probe. These results establish non-Hermitian Zeeman quantum geometry as a measurable framework connecting local Dirac-node topology and transport.

quant-ph↗

arXiv2026

OceanMoE: Structured Conditional Sparse Computation for Long-Horizon Multivariate Ocean Forecasting

Multivariate ocean forecasting must exploit shared evolution in a coupled ocean system while adapting to the heterogeneous statistical and dynamical characteristics of different prediction variables and locations. Fully shared models may lack the flexibility to handle this heterogeneity, whereas fully independent models discard the common ocean context shared across variables. The key question is how to retain shared context in a unified model while allowing computation to specialize according to the prediction target and local state. We propose OceanMoE, a structured conditional sparse Mixture-of-Experts framework that combines sharing and specialization for multivariate ocean forecasting. OceanMoE fuses cross-variable information to construct target-specific local representations and uses them to perform content-conditioned sparse routing at each spatial location, with the number of active experts adapted to router confidence. In the decoder, routing is augmented with a learned geographic bias parameterized by spherical-harmonic spatial bases, while shared residual and seasonal pathways provide common cross-variable and month-dependent context. Experiments on long-horizon autoregressive ORAS5 forecasting show that OceanMoE lowers aggregate forecasting error in both evaluated settings and maintains lower geometric-mean normalized RMSE than the corresponding baselines over most later rollout months. Routing analyses further show that expert allocation varies with prediction targets and spatial locations. These results support structured conditional computation as a modeling strategy for balancing shared ocean context with adaptive specialization.

cs.LG↗

arXiv2026

Observation of multiphoton entanglement in resonance fluoresce

Resonance fluorescence, the process in which a single two-level quantum emitter interacts with a near-resonant coherent light, is a cornerstone of quantum optics as a paradigmatic light-matter interaction. Quantum phenomena in resonance fluorescence have been accounted for in terms of one- and two-photon processes; however, the physical processes arising from the interaction of a single two-level atom with higher photon numbers remain unexplored experimentally. Here we experimentally reveal the multiphoton scattering nature of resonance fluorescence, which intrinsically gives rise to multiphoton entanglement. By accessing quantum fluctuations in the field emitted from a cavity-quantum electrodynamics system, we resolve third- and fourth-order photon scattering events from a single two-level atom. In the three-photon component, the scattered photons are shown to be genuinely energy-time entangled, as verified by a violation of Svetlichny's inequality. As a demonstration of an application, a quantum secret sharing protocol (QSS) is implemented using this entangled photon source. Our study refines the descriptive framework of resonance fluorescence and broadens the scope of its investigation. It also establishes a fundamentally simple route for generating multiphoton entangled states for quantum information processing and quantum metrology.

physics.atom-ph↗

arXiv2026

Re2A: Situated Conversational Recommendation via Rubric-based Preference Reasoning and Alignment

Real-world recommendation scenarios are commonly grounded in shared physical environments during user-recommender interactions. This motivates situated conversational recommendation (SCR), a complex task requiring recommender assistants to jointly reason over dialogue history, co-observed scenes, and in-scene item attributes. However, current approaches struggle with this setting due to two intertwined challenges: accurately understanding situated user preferences throughout the conversation and generating responses that simultaneously satisfy user needs and grounded situations. To this end, we propose Re2A, a framework that formulates SCR as a structured reason-then-align process. We introduce rubric-based preference reasoning, which uses automated rubrics to guide the model toward producing explicit preference states. Based on these states, we propose a preference-conditioned optimization to align response generation with dual objectives: user preference satisfaction and situation consistency. Extensive experiments on two SCR datasets demonstrate that Re2A consistently outperforms state-of-the-art methods, delivering more precise, context-aware conversational recommendations. Our code is available at https://github.com/DongdingLin/Re2A.

cs.AI↗

arXiv2026

Mind the Gap: Rethinking I/O Design for Contact-Rich Visuomotor Policy Learning

Contact-rich teleoperation logs expose a policy I/O design choice: demonstrations may contain the robot execution (E), leader command (C), or both. These signals are not interchangeable: E2E may discard contact-generating command offsets, whereas E2C preserves these offsets but omits the robot's execution response. We propose Dual-State Conditioning (EC2C), which conditions on both E and C while predicting future C, exposing command-execution mismatch as a cue for contact, latency, payload, and operator compensation; in quasi-static contact, this cue is often force-correlated. On a low-cost setup without force, tactile, or motor-current policy input, EC2C outperforms E2E and a strong E2C baseline across several real-world contact-rich, force-sensitive, and dynamic tasks. These results support EC2C as a practical default I/O setting for contact-rich imitation learning. We further formulate latency-adaptive inpainting as a temporal extension of this I/O choice for action-chunking policies, and discuss when long histories help dynamic inference or introduce causal confounding.

cs.RO↗

arXiv2026

One- and two-dimensional cluster states for topological phase simulation and measurement-based quantum computation

Quantum entanglement is a fundamental resource for quantum information processing and serves as a critical benchmark for quantum hardware performance. Cluster states are a special class of entangled states that serve as universal resources for measurement-based quantum computation and possess an intrinsic symmetry-protected topological order, which confers robustness against symmetry-respecting noise. Here we report the scalable preparation and verification of genuine multipartite cluster states on the 105-qubit Zuchongzhi 3.1 superconducting processor. We achieve one-dimensional cluster states of up to 95 qubits and two-dimensional cluster states of up to 72 qubits. The symmetry-protected topological cluster states exhibit input-state-dependent robustness under symmetry-breaking perturbations due to an operational parity structure that enhances the performance of measurement-based quantum computation. Furthermore, we use our two-dimensional cluster states to implement the Deutsch-Jozsa algorithm within the measurement-based quantum computation framework, achieving higher output-state fidelity compared with traditional circuit-based models and a query efficiency advantage over classical approaches. Our work establishes a scalable platform that combines large-scale entanglement generation, symmetry-protected topological order and practical quantum algorithms to enable robust, fault-tolerant measurement-based quantum computation.

quant-ph↗

arXiv2026

Quantitative Rigidity of pseudo-rotations on the two-torus and Sarnak's conjecture

We establish quantitative rigidity results for pseudo-rotations of the two-torus under a $(C,δ)$-deviation condition relative to their rotation vectors. The main ingredient is a quantitative free-disk estimate that converts bounds on orbit deviation into explicit control of the distance between the iterates and the identity map. Under such a $(C,δ)$-deviation condition, we show that Hölder continuous super-Liouvillean irrational pseudo-rotations are $C^0$-rigid with an exponential decay rate and that $C^k$ semi-irrational pseudo-rotations of strong non-Brjuno type exhibit $C^{k-1}$-rigidity with a superpolynomial decay rate. Moreover, under this deviation condition and sufficiently large irrationality measure, we show that Hölder continuous skew products on $\mathbb{T}^2$ over circle rotations are $C^0$-rigid with a polynomial decay rate. As a consequence, all these classes satisfy Sarnak's conjecture.

math.DS↗

arXiv2026

The maximum $t$-diversity of $t$-intersecting families of permutations

The study of $t$-intersecting families in symmetric groups received lots of attention in the last two decades. In this paper, we study two different kinds of stability results for $t$-intersecting families in symmetric groups. Let $Σ_n$ denote the symmetric group on $\{1,2,\ldots,n\}$. The $t$-diversity $γ_t(\mathcal{F})$ of $\mathcal{F}\subseteqΣ_n$ is defined as the minimum number of members of $\mathcal{F}$ whose deletion results in a family with transversal number $t$. The star $t$-diversity $γ_t^{\star}(\mathcal{F})$ is defined as the minimum number of members of $\mathcal{F}$ whose deletion results in a $t$-star. For $n$ relatively large with respect to $t$, we determine the best possible bounds for both $γ_t(\mathcal{F})$ and $γ_t^{\star}(\mathcal{F})$ over all $t$-intersecting families $\mathcal{F}\subseteqΣ_n$. The equality holding conditions are characterized. For $γ_t(\mathcal{F})$, the extremal case is generated by all $2t$-subsets of a $3t$-partial permutation or, when $t=2$, the family complements of the lines of the Fano plane. For $γ_t^{\star}(\mathcal{F})$, the extremal family is generated by all $(t+1)$-subsets of a $(t+2)$-partial permutation.

math.CO↗