arXiv subjects

Fan Wu

Publications and source records attributed to Fan Wu.

At least 19 recordsLinked to original sources

arXiv2026

PeV Diffuse Neutrinos from Magnetically Arrested Disks in Low-Luminosity Active Galactic Nuclei

Radiatively inefficient accretion flows (RIAFs) in low-luminosity active galactic nuclei (LLAGNs) are promising candidates for the PeV astrophysical diffuse neutrino background (DNB), but how the magnetically arrested disk (MAD) state affects their neutrino output and what fraction of LLAGNs are in this state remain unclear. We present a comprehensive study of diffuse neutrino emission from the inner accretion flows of LLAGNs in the MAD state, including all three flavors and both hadronuclear ($pp$) and photohadronic ($pγ$) production channels. Assuming that $\sim 5\%$ of the accretion power is transferred to nonthermal protons, current IceCube data allow at most $2.3\times10^{-3}$ of LLAGNs to have their inner accretion flows in a MAD state that can produce high-energy neutrinos. A lower fraction of the accretion power transferred to nonthermal protons would allow a proportionally larger fraction of such LLAGNs. Assuming that the comoving number density of low-excitation radio galaxies (LERGs) approximately matches that of LLAGNs in the MAD state, we also estimate a MAD fraction of $\simeq 1.6\times10^{-3}$. The corresponding diffuse intensity is compatible with current IceCube data and testable by IceCube-Gen2. For this MAD fraction, the all-flavor diffuse intensity peaks at $E_ν\simeq1.5\,{\rm PeV}$ with $E_ν^2Φ_ν\simeq2.6\times10^{-9}\,{\rm GeV\,cm^{-2}\,s^{-1}\,sr^{-1}}$. Current IceCube data are beginning to challenge several predictions from previous LLAGN RIAF models, and the MAD component can become the dominant contribution from LLAGN accretion flows beyond the cutoffs of most earlier RIAF spectra considered here.%%%This constraint suggests that only a small fraction of LLAGNs are supplied with sufficient coherent net poloidal magnetic flux.

astro-ph.HE↗

arXiv2026

Estimating Nonseparable Selection Models: A Functional Contraction Approach

We propose a novel method for estimating nonseparable selection models. We show that, for a given selection function and under suitable contraction conditions, the potential outcome distributions are nonparametrically identified from the selected outcome distributions and can be recovered using a simple iterative algorithm. This result enables a full-information approach to estimating selection models without imposing parametric or separability assumptions on the outcome equation. We propose a two-step estimation strategy for the potential outcome distributions and the parameters of the selection function and establish the consistency and asymptotic normality of the resulting estimators. Monte Carlo simulations demonstrate that our approach performs well in finite samples. The method is applicable to a wide range of empirical settings, including consumer demand models with only transaction prices, auctions with stochastic winning rules and incomplete bid data, and Roy-type models with data on accepted wages.

econ.EM↗

arXiv2026

Exploring Post-Training Alignment of Small Language Models for Biomedical Data-to-Text Generation: A Case Study of Medication Leaflet

Translating structured biomedical information into accurate, patient-friendly narratives remains a challenging biomedical data-to-text (D2T) task. While small language models (SLMs) are attractive for healthcare applications because of their lower computational cost, easier deployment, and improved data governance, the impact of post-training alignment on biomedical D2T generation has not been systematically studied. In this work, we compare four alignment methods: supervised fine-tuning (SFT), direct preference optimization (DPO), odds ratio preference optimization (ORPO), and group relative policy optimization (GRPO), using Qwen-2.5 models for medication leaflet generation. We further introduce FDALeaflets, a dataset derived from OpenFDA drug labels, to evaluate cross-regulatory generalization in a zero-shot setting. Selected aligned SLMs outperform GPT-5 on several reference-based metrics, whereas GPT-5 achieves better hallucination and adequacy scores on FDALeaflets under LLM-as-judge evaluation. Among RL-free alignment methods, ORPO achieves the best in-domain semantic performance, reaching a semantic textual similarity (STS) score of 0.9752. GRPO achieves the highest FDALeaflets STS (0.9407), exceeding GPT-5 (0.9295). These findings show that post-training alignment is a key factor in improving both semantic fidelity and generalization for biomedical D2T generation, while clinical safety and patient readability require human verification.

cs.CL↗

arXiv2026

C$^2$U in RAG: Compositional Concept Unlearning in Retrieval-Augmented Generation

Retrieval-augmented generation can expose unwanted concepts through retrieved evidence even when the generator is moderated. We study query-side concept suppression with a fixed encoder and document index. Compositional Concept Unlearning (C$^2$U) routes a profile-derived direction through an augmented atom bank, adding only a bounded edited-minus-original delta to the raw query. Across five datasets, four encoders, and three seeds, the validation-selected configuration reduces pooled binary target Hit@5 by 4.25 percentage points with non-positive control degradation and target-exposure changes. Under the evaluated query transforms, it suppresses target density without increasing budgeted control-query exposure. Separately, ArXiv four-concept composition improves over an averaged-direction comparator after independent validation gates and first-use test evaluation.

cs.CL↗

arXiv2026

Risk Decomposition for Hallucination-Aware Video Sampling

Despite achieving high visual fidelity, video models still suffer from physical hallucinations, limiting their reliability as world simulators. Existing methods leverage model-internal or external signals to improve physical plausibility, but their roles and limitations require a unified explanation. We propose a two-dimensional physical-hallucination risk framework that defines low-density sampling risk and high-density drifting risk. The first risk captures generations assigned low likelihood by the video model, for example, a ball suddenly teleporting despite the model favoring continuous motion, while the second risk captures physical inconsistencies in high-likelihood generations, for example, a ball bouncing progressively higher without additional energy. Our analysis explains how model-internal signals steer sampling toward higher-density regions of the learned distribution to address low-density sampling risk. However, higher model likelihood does not guarantee physical plausibility, motivating external guidance to address high-density drifting risk. Based on this framework, we develop Hallucination-Aware Sampler (HA-Sampler), which applies external physical guidance within latent refinement to generate physically plausible videos, and HA-Scorer, which derives a likelihood-based signal from the generator itself for evaluating generated videos using continuous normalizing flow (CNF), with coarse ODE integration and finite-difference divergence estimates. Experiments on Physics-IQ and VideoPhy benchmarks support our theoretical analysis and demonstrate that our methods outperform the compared sampling and scoring SOTA baselines.

cs.CV↗

arXiv2026

RL-IF: Tracking Non-local Training Data Influence for Online Reinforcement Learning

Online reinforcement learning (RL) has achieved significant success in decision-making tasks, but modern online RL remains highly sensitive to the quality of training experience. Understanding the contribution of individual training instances is therefore central to data-centric RL, with applications ranging from experience selection to policy interpretation. Existing data attribution methods for online RL primarily focus on local influence within a single training round, overlooking the non-local data influence across multiple training rounds. In this work, we formalize data attribution in online RL through a trajectory-level leave-one-out quantity, RL-LOO, which measures how one training sample influences a downstream target after subsequent online updates have taken place. We then derive RL-IF, a first-order estimator that propagates data influence through both optimization effects and policy-induced sampling effects, and provide theoretical guarantees on its approximation accuracy under smoothness assumptions. Our analysis further disentangles these two effects and shows both theoretically and empirically that the full RL-IF estimator can often be well approximated using the optimization effect alone. Building on this observation, we develop RINSE, a practical data filtering method for online RL. Experiments on standard control benchmarks and an RLHF-style toxicity-mitigation setting show that RINSE improves training efficiency and final performance over standard PPO and local attribution baselines. These results suggest that RL-IF provides a practical and principled framework for understanding and improving online RL.

cs.LG↗

arXiv2026

Mechanism-Aligned Observability for Training Instability of Large Language Models

Frontier large language model training consumes massive accelerator fleets over long wall-clock times, making stability failures extremely costly. However, global symptoms such as loss divergences often emerge only after the underlying fault has already affected the training dynamics, further increasing the sunk compute. In which internal representation can a training fault become persistently and interpretably observable before it appears in global training metrics? We propose a mechanism-aligned principle for training observability: identify the internal state into which a fault is persistently written, derive the structure induced by its mechanism, and construct monitors in representations aligned with the computation of the affected module. For FlashAttention, we show that biased backward errors accumulate into a persistent low-dimensional structure in weight updates, producing spectral concentration. For MoE routers, we characterize routing dynamics through per-token routing behavior rather than parameter similarity, which may change without altering the routing function. Experiments show that the two monitors respond selectively to changes in their corresponding modules, before degradation becomes visible in global training metrics.

cs.CL↗

arXiv2026

Toward Collective Model Intelligence: Characterizing and Coordinating Models

Independently developed AI models accumulate diverse and complementary capabilities, but a growing model collection does not automatically realize Collective AI. Addressing this gap requires understanding model capabilities with limited evaluation, organizing complementary models to collaborate on diverse tasks, and executing model calls efficiently under resource constraints. We present a systematic approach to pursuing Collective Model Intelligence through three connected components. First, Model Eval constructs evidence-based capability profiles by combining query-level performance estimation with cross-domain prediction, reducing the need for exhaustive evaluation. Second, Mio (Models in One) organizes complementary model candidates and supports task-conditioned, cost-aware selection and collaboration as task requirements evolve. Third, Mos (Model Serving OS) separates model selection from execution and provides endpoint scheduling and shared model deployment and distribution. Experiments show lower capability-estimation errors under limited evaluation. Across nine public benchmarks, Mio 1.0 Lite achieves the highest average score and lowest average cost among compared methods, with per-request costs 73.7-97.5% lower than those of the best-scoring alternatives. A prewarmed shared cache accelerates model-file retrieval by up to 23.3 times. This work lays a foundation for Collective Model Intelligence by characterizing and coordinating the diverse capabilities of an expanding model universe.

cs.AI↗

arXiv2026

Restarted Fixed-Point Acceleration of Proximal ADMM for Capped-$\ell_1$ Penalized Optimization

We consider box-constrained optimization with a convex, Lipschitz continuous loss function that may be nonsmooth and a capped-$\ell_1$ penalty. Exploiting the convex branch representation of the capped-$\ell_1$ penalty, we construct an associated nonexpansive fixed-point mapping for each branch and, based on these mappings, propose a restarted accelerated proximal alternating direction method of multipliers (RA-pADMM). The resulting algorithm consists of a preliminary branch identification phase and a subsequent accelerated phase based on either the Halpern or fast Krasnosel'skiĭ--Mann iteration, together with a sufficient decrease restart mechanism. We prove that the preliminary phase terminates after finitely many iterations and that only finitely many branch updates are accepted during the subsequent phase. The primal iterates and proximal outputs converge to a common lifted stationary point, which is also a $ν$-strong local minimizer of the corresponding cardinality penalized problem. Moreover, the Halpern and fast Krasnosel'skiĭ--Mann variants achieve nonergodic $O(k^{-1})$ and asymptotic $o(k^{-1})$ convergence rates, respectively, for both the objective value gap and the Karush--Kuhn--Tucker residual associated with the final branch. Numerical experiments on sparse recovery and real data classification problems demonstrate the effectiveness of the proposed RA-pADMM algorithm.

math.OC↗

arXiv2026

Who Keeps a Bubble Alive? Skin in the Game and the Survival of Rational Bubbles

We study how retained ownership supports the value of an asset that pays no dividends and is held only for resale. A sponsor chooses hidden maintenance that determines market survival, bears its full cost, and captures only the benefit accruing to its retained tokens. A positive stationary equilibrium exists if and only if retention meets an endogenous minimum. Under additional curvature restrictions, retention above this threshold supports two positive stationary equilibria. Conditional on survival, prices below the upper equilibrium converge to the lower one while the upper price is the largest feasible initial valuation. Greater retention expands the feasible price interval. Credible news that future retention will be below the minimum causes an immediate collapse through backward unraveling. The model identifies a governance channel through which insider unlocks can affect valuation without changing total token supply or fundamental cash flows.

econ.TH↗

arXiv2026

Building a Sovereign Benchmark The Limits of Private Coordination

Public debt supplies liquidity and collateral services only if institutions coordinate on using the same claim. We study this coordination when use carries a recurring cost and a benefit that rises with other users. Above a quality threshold, two positive-adoption equilibria exist, and under best responses the initial user base decides between collapse and high adoption. Under logit choice, as the population grows and noise vanishes, the long-run distribution selects high adoption only above a strictly higher threshold. Debt issuance improves market depth but eventually dilutes backing per claim. This trade-off creates a range of debt levels supporting the liquidity premium of public debt in macroeconomic models and a narrower range supporting its long-run selection. At the boundary of the wider range, the premium is discontinuous. A matching transfer that corrects the adoption externality makes stationary selection maximize network surplus. Transfers can also help launch a network without permanent support.

econ.TH↗

arXiv2026

Local-First Intelligence via Proactive SLM Collaborating with Feedback-Providing LLM

Personal intelligence requires strong capabilities under strict privacy and efficiency constraints. Unlike cloud-first intelligence, which relies on powerful but costly and privacy-vulnerable large language models (LLMs), we focus on local-first intelligence, emphasizing on-device small language models (SLMs), which enable efficient, private inference but have limited capacity. To reconcile the capability gap with privacy and efficiency, we propose a dynamic collaboration framework, where an SLM learns to proactively decide how to consult an LLM during multi-step reasoning, while the LLM provides adaptive feedback instead of acting as a passive tool, modeling them as heterogeneous agents under information asymmetry. We further systematically investigate how collaboration strategies are shaped by SLM and LLM capabilities as well as efficiency and privacy constraints. Evaluation results reveal a distinct scaling effect: stronger SLMs become more self-reliant, while stronger LLMs enable fewer and more informative interactions. In addition, the learned dynamic collaboration strategies significantly outperform various baselines and transfer robustly to unseen LLMs as well as to real-world personal tasks without additional training.

cs.CL↗

arXiv2026

A Safeguarded Accelerated Proximal ADMM Algorithm for Solving Structured Cardinality Penalized Optimization Problems

This paper considers a class of structured cardinality penalized optimization problems with nonsmooth convex loss functions, for which the proximal mapping associated with the structured $\ell_0$ penalty is generally unavailable. To overcome this difficulty, we develop a chain of exact reformulations based on variable splitting, exact penalization, and continuous relaxation, and establish rigorous equivalence relationships among these resulting models. In particular, we define a class of strong local minimizers for the original problem and show that these minimizers are equivalent to weak $d$-stationary points of the relaxation problem. Further, we propose a safeguarded accelerated proximal alternating direction method of multipliers (SAP-ADMM). We prove that the sequence generated by SAP-ADMM converges to a weak $d$-stationary point of the relaxation problem. Depending on the parameter settings and algorithmic regime, we establish $O(1/k)$ or $o(1/k)$ convergence rates for both the objective function value gap and the KKT residual. Numerical experiments on signal and image denoising demonstrate the effectiveness of both the proposed models and SAP-ADMM in terms of reconstruction quality and computational efficiency.

math.OC↗

arXiv2026

PixARG: Pixel-Asynchronous Retrieval-Augmented Diffusion for Fast and Realistic Image Generation

While diffusion models achieve impressive image generation capabilities, they suffer from high computational costs, implausible artifacts, and limited adaptability to dynamic or private data due to static, centralized training. To address these, we propose PixARG, a novel retrieval-augmented generative framework that integrates dynamic real-world image retrieval into the diffusion pipeline to simultaneously enhance efficiency and visual realism. Specifically, PixARG constructs a pixel-level affinity map between the user's text prompt and the retrieved image to guide a heterogeneous noise initialization. By modulating noise intensity across spatial regions, this mechanism preserves semantically relevant structural priors while perturbing less relevant areas, thereby significantly accelerating convergence. To accommodate this spatially non-uniform noise and ensure synchronized generation, we introduce a pixel-wise asynchronous denoising scheduler and a spatiotemporal denoiser conditioned on the affinity map. Furthermore, we derive an affinity-adaptive training objective that elegantly balances deterministic restoration with stochastic regeneration across pixels. Experiments on LAION and COYO datasets across offline, online, and user-specific settings show PixARG enhances realism and semantic alignment while drastically reducing diffusion steps.

cs.CV↗

arXiv2026

Demonstration Calibrated Port-Hamiltonian Retuning for Manipulation Policies

Diffusion and VLA policies for manipulation are often deployed through downstream impedance controllers. The stiffness and damping gains of these controllers affect task success, yet are commonly inherited from data collection rather than selected for the deployed policy. Although empirical gain sweeps can improve performance, they require repeated evaluation rollouts. We introduce PHRetune, an offline method that derives controller gains for a frozen policy without evaluation rollouts or gain search. Our approach learns a port-Hamiltonian model from demonstrations to estimate the effort and energy associated with the policy's predicted actions. The policy is applied to recorded demonstration observations, and its predictions are assessed against demonstration-derived effort and energy budgets. From this comparison, we derive a single gain scale in closed form, adjusting the downstream controller while preserving the policy and its action representation. The gains are fixed before evaluation, without requiring a prior manipulator model, task rewards, policy retraining, or additional runtime computation. Across LIBERO suites, PHRetune improves Diffusion Policy success by up to 9.4 percentage points, with the derived gains achieving the highest observed success rates in empirical gain sweeps. On all four real-world manipulation tasks, PHRetuned Diffusion Policy outperforms the nominal policy, alternative gain-tuning methods, and a policy-retraining baseline. The same procedure improves success with SmolVLA and OpenVLA-OFT on every task, while reducing acceleration and jerk for both VLA backbones.

cs.RO↗

arXiv2026

Sample-Efficiency of Kolmogorov-Arnold Networks

Deep reinforcement learning has achieved substantial performance gains over classical control approaches. Yet, a central challenge to learning in real-world applications is acquiring costly samples. Kolmogorov-Arnold Networks are a recently proposed architecture that can learn physical relationships in control problems effectively, with significantly higher parameter efficiency and interpretability when compared to Multi-Layer-Perceptron architectures. In this work, we systematically study sample-efficiency using computational experiments, covering the Feynman dataset and the Gymnasium RL benchmark. The results show that similar performance can be achieved with 40% fewer samples using the Kolmogorov-Arnold architecture, and that relative performance improvements up to 50% occur during the training process. The observed gains are robust to varying levels of noise in rewards. These results highlight the potential of the Kolmogorov-Arnold architectures for more sample-efficient reinforcement learning. Code: https://github.com/DerKevinRiehl/neurips26_kan_training

cs.LG↗

arXiv2026

F$^3$NO: Frequency-Decomposed Finite-Time Flow-map Neural Operators with Cross-Scale Conditioning

Neural operators enable fast PDE forecasting, but repeated predictions accumulate errors and fine-scale structures remain difficult to resolve. We introduce a frequency-decomposed finite-time flow-map neural operator (F$^3$NO) that leverages updated low-frequency features to guide nonlinear refinement of high-frequency information. Within each layer, this cross-scale conditioning connects global spectral processing with local detail refinement. The model directly predicts states at specified future times and adjusts the contributions of the two branches according to the prediction interval. For longer trajectories, it combines parallel predictions within short temporal segments with recursive propagation between segments. Experiments on five PDE benchmarks demonstrate improved forecasting accuracy over autoregressive and direct-prediction baselines. Ablations show that frequency-decomposed refinement can improve accuracy with fewer parameters, while the benefits of segmentation depend on spatial resolution and dynamical regime.

cs.LG↗

arXiv2026

Disk-Resident Graph ANN Search: An Experimental Evaluation

As data volumes grow while memory capacity remains limited, disk-resident graph-based approximate nearest neighbor (ANN) methods have become a practical alternative to memory-resident designs, shifting the bottleneck from computation to disk I/O. However, since their technical designs diverge widely across storage, layout, and execution paradigms, a systematic understanding of their fundamental performance trade-offs remains elusive. This paper presents a comprehensive experimental study of disk-resident graph-based ANN methods. First, we decompose such systems into five key technical components, i.e., storage strategy, disk layout, cache management, query execution, and update mechanism, and build a unified taxonomy of existing designs across these components. Second, we conduct fine-grained evaluations of representative strategies for each technical component to analyze the trade-offs in throughput, recall, and resource utilization. Third, we perform comprehensive end-to-end experiments and parameter-sensitivity analyses to evaluate overall system performance under diverse configurations. Fourth, our study reveals several non-obvious findings: (1) vector dimensionality fundamentally reshapes component effectiveness, necessitating dimension-aware design; (2) existing layout strategies exhibit surprisingly low I/O utilization (less than or equal to 15%); (3) page size critically affects feasibility and efficiency, with smaller pages preferred when layouts are carefully optimized; and (4) update strategies present clear workload-dependent trade-offs between in-place and out-of-place designs. Based on these findings, we derive practical guidelines for system design and configuration, and outline promising directions for future research.

cs.DB↗