Searcharxiv⌕ Search

arXiv subjects

Qi Xu

Publications and source records attributed to Qi Xu.

At least 19 recordsLinked to original sources
arXiv2026

VideoZeroBench: Probing the Limits of Video MLLMs with Spatio-Temporal Evidence Verification

Video multimodal large language models achieve strong results on existing benchmarks, but answer accuracy alone does not establish whether they can locate the evidence needed to answer a question. We introduce VideoZeroBench, a challenging long-video benchmark with manually annotated question-answer pairs spanning 13 video domains. Questions target fine-grained cues, fleeting events, and evidence distributed across multiple segments. Temporal intervals and key-frame boxes are annotated where applicable. All questions undergo two rounds of cross-verification for answer validity and evidence quality. Our five-level diagnostic protocol compares answering with and without evidence hints, then combines answer correctness with independently evaluated temporal and spatial grounding. Across 19 evaluated models, the best standard QA accuracy is 24.8% (Level-3), achieved by Gemini-3.7-Flash. No model exceeds 1.8% when correct answers and accurate spatio-temporal localization are jointly required (Level-5). Analyses of atomic abilities, evidence spans, input modalities, and thinking-with-videos inference further characterize where the evaluated systems struggle. These findings motivate more precise evidence search and localization for long-video question answering. Our code and data are publicly released.

cs.CV↗
arXiv2026

An $L^{5/6}$ bound for the periodic Kuramoto-Sivashinsky equation

We consider the Kuramoto-Sivashinsky equation $u_t+uu_x+u_{xx}+u_{xxxx}=0$ with $L$-periodic boundary conditions and zero spatial mean. We prove that every smooth solution satisfies $\|u(t)\|_{L^2}\leq CL^{5/6}+CL^{3/2}/t$ for $t>0$, where $C$ is a universal constant. In particular, $\|u(t)\|_{L^2}\leq CL^{5/6}$ for all $t\geq L^{2/3}$, independently of the initial norm.

math.AP↗
arXiv2026

Strategist: Evolving Reasoning with Agents

A model is only half of an AI agent system. The other half is the inference strategy: the reasoning procedure that turns model calls into a response. Today that second half is written by hand, and the experience of building one strategy never carries to the next. We introduce Strategist, a meta-agent that evolves inference strategies across tasks. Each strategy is written as a blueprint,executable code that other strategies can reuse, and together the blueprints form a library. Given a task and a small development set, Strategist retrieves related blueprints, writes candidate strategies, tests them on real samples, and revises them from the results. Every candidate then enters the library, so later tasks build on earlier discoveries. On 30 benchmarks, Strategist beats the strongest baseline on 24 and matches its average accuracy (74.7% vs.74.0%) at 39% lower inference cost. Its strategies also transfer to unseen tasks and other models without further search. The part of an AI system that is not improved on its own now does.

cs.AI↗
arXiv2026

A Hierarchical Bayesian Differential Calibration of the Gaia DR3 Parallax Zero Point Using Galactic Globular Clusters

Globular clusters (GCs) occupy a broad range of stellar colors and magnitudes while sharing nearly the same distance. Gaia Data Release 3 parallaxes retain systematic zero-point residual variations for GCs, even after the correction by the widely used standard correction model (L21). We derive an empirical residual correction by the Hierarchical Bayesian Model, separating cluster-level offsets from the differential information carried by the member sequences. We first evaluate the L21 baseline at a fixed reference color, which assigns the remaining color-dependent correction to the residual model. Then we model the residual zero point, $δZ(G,ν_{\rm eff})$, with B-splines functions in color and magnitude--color interaction terms. The uncorrected cluster residuals exhibit a negative offset of approximately $25$--$40~μ{\rm as}$. The L21 baseline removes most of this offset but leaves coherent color- and magnitude-dependent structure. Applying the residual model brings the mean residuals close to zero, especially across most of the sampled color range. In five-fold cross-validation that excludes entire clusters from training, the correction modestly reduces the magnitude- and color-trend RMS relative to the L21 correction. The resulting model therefore provides a differential refinement of the \textit{Gaia} DR3 parallax zero-point within the calibrated domain for GCs, rather than an independent determination of its absolute value.

astro-ph.GA↗
arXiv2026

When Does Pre-Trained Capacity Help Causal Learning? From Scalar Prediction to Counterfactual Post-Training

How does pre-trained model capacity relate to performance after adaptation for causal tasks? We study two systems built on the Qwen 2.5 family: CausalLLM, which estimates scalar potential outcomes with treatment-specific heads, and CausalGen, which generates treatment-conditioned text or code using soft prefixes and factual supervised fine-tuning (SFT), counterfactual SFT, and preference alignment. Increasing backbone size from 0.5B to 1.5B yields mixed scalar-estimation results on IHDP, but improves CodeGen pass@1 from 0.267 to 0.575 and treatment-effect score from 0.231 to 0.515 after the complete generative pipeline. Within 0.5B CausalLLM, narrative inputs reduce mean sqrt(PEHE) from 1.842 to 1.780, a 3.4% improvement over tabular inputs; the 5.8% improvement over a no-LLM baseline combines changes in representation and input format. These descriptive comparisons motivate a distinction between available and convertible capacity. They do not isolate the contribution of post-training: the two systems differ in dataset, architecture, output modality, and supervision, and intermediate-stage evaluations are unavailable. With only two small backbones and three replications, the results support a hypothesis about capacity--adaptation alignment rather than a general scaling law or a demonstrated causal mechanism.

cs.LG↗
arXiv2026

Any 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale

Scene-level 3D generation has long been dominated by 2D multi-view or video diffusion models. Typically, these 2D-based approaches perform scene generation in a 2D latent space, which introduces two fundamental issues: (i) representing 3D scenes via 2D views leads to significant representation redundancy, and (ii) latent space rooted in 2D inherently limits the spatial consistency of the generated scenes. In this paper, we propose to perform 3D scene generation directly within a high-dimensional implicit 3D latent space derived from powerful 2D foundation models (e.g., DINOv2, SigLIP2, DA3). Then we tame diffusion transformers to perform diffusion modeling directly in the 3D latent space, enabling 3D-grounded scene generation. In our experiments, we not only demonstrate the advantages of our method over 2D-based approaches in terms of efficiency and spatial consistency, but also comprehensively analyze the impact of 3D latent representation on 3D scene generation performance. Furthermore, we validate the favorable scalability of our method with respect to representation and model capacity.

cs.CV↗
arXiv2026

Regularity of asymptotically axisymmetric Navier-Stokes flows with rotationally invariant curl of the force

We answer a question of Constantin, Ignatova, and Vicol concerning regularity of asymptotically axisymmetric solutions to the three-dimensional Navier-Stokes equations. For suitable weak solutions that are smooth before a terminal time and have a force uniformly bounded in $C_x^2$, we replace spatial analyticity of the force by rotational invariance of its curl. Under the same anisotropic Type II bounds on the angular mean and exact axial symmetry on a possibly collapsing core, the solution is axisymmetric on the entire cylinder and regular at the putative singular point. No time regularity or quantitative lower bound on the core radius is required. The proof combines unique continuation for linear Stokes systems with a Baire category argument and the axisymmetric regularity theorem due to Constantin, Ignatova, and Vicol.

math.AP↗
arXiv2026

Agentic neural architecture search for clinical prediction from electronic health records: Cross-hospital knowledge transfer and computational efficiency

Transformer architectures for clinical prediction from electronic health records (EHRs) often require costly manual tuning and may vary across hospitals and tasks. We introduce ATHENA (Agentic Transfer across Hospitals for EHR Neural Architecture Search), a knowledge-guided agentic neural architecture search framework for Transformer-based EHR modeling. A weight-sharing supernet is pretrained once per hospital, allowing candidate architectures to inherit weights and be evaluated through fine-tuning rather than independent pretraining. A two-layer cross-hospital prior supplies high-performing source architecture examples and task-specific component preferences summarized by SHapley Additive exPlanations (SHAP)-based meta-regression. These priors guide a multi-agent large language model search using target-hospital validation feedback. Across six clinical prediction tasks at one held-out OneFlorida+ site and one external MIMIC-IV site, ATHENA significantly outperforms all four baselines in 9 of 12 site-task evaluations under a strict equal-compute comparison. In a separate comparison using a common pretrained AutoFormer supernet to isolate search strategy performance, ATHENA ranks first in 9 of 12 evaluations at a maximum search budget of 30 candidate evaluations. It also shows more consistent architecture selection across repeated searches. ATHENA supports more efficient architecture selection for Transformer-based clinical prediction models.

cs.AI↗
arXiv2026

Quantitative Parabolic Decay With Bounded Drift

A bounded time-dependent drift together with an arbitrarily small real potential can produce double exponential decay of the full spatial $L^2$ norm. We prove matching quantitative two-time lower bounds for energy solutions of $\partial_tu-Δu=vu+w\cdot\nabla u$ on $\mathbb{R}^n$, for $n\geq3$, when $w\in L^\infty_{t,x}$ and either $v$ is sufficiently small in $L^\infty_tL^{n/2}_x$ or $v\in L^\infty_tL^p_x$ for some $n/2<p\leq\infty$, without a smallness assumption. In particular, a nonzero global solution satisfies a double exponential lower bound at large times. This decay scale is optimal for these coefficient classes: In every dimension $n\geq3$, we construct a real bounded drift, a real potential arbitrarily small in the critical space and in any prescribed subcritical $L^p$ space, and a nonzero real energy solution with double exponential decay. The construction localizes Rowan's torus example by absorbing the localization error.

math.AP↗
arXiv2026

Iterative Policy Refinement through Semantic Rollout Analysis

Structured policies improve efficiency, robustness, and interpretability in imitation learning by introducing task-specific inductive bias, but existing structure generation methods rely either on extensive human input or on static domain knowledge encoded in LLMs, which may be inconsistent with the expert demonstrations. We propose a closed-loop framework that iteratively refines structured policies using LLM-guided analysis of policy rollouts. By logging rollouts as semantically meaningful tabular data and prompting the LLM to generate diagnostic analysis code, our method identifies suboptimalities in the policy structure and iteratively corrects them without requiring human instruction. Experiments on car racing and door opening tasks show that our approach improves imitation learning performance by up to 15% over zero-shot LLM-generated structures and requires 75% less compute to achieve the same reinforcement learning performance. These results demonstrate that tabular rollout analysis provides an effective feedback signal to align LLM-generated policy structures with expert demonstrations, and we can utilize it to generate good policy structures automatically.

cs.LG↗
arXiv2026

Quantitative Decay for Linear Parabolic Equations

We establish sharp temporal lower bounds for the full spatial $L^2$-norm of energy solutions to $\partial_tu-Δu=vu$ on $\mathbb{R}^n$, with $n\geq3$. For global solutions, nonvanishing at a single time implies the lower bound $ce^{-Ct}$ when $v$ is small in the scale-invariant space $L^\infty_tL^{n/2}_x$, and $ce^{-Ct^2}$ throughout the subcritical range $v\in L^\infty_tL^p_x$, with $p>n/2$, without smallness. The linear time exponent is optimal already for the free heat equation, while we show that the quadratic exponent is optimal for bounded complex potentials on $\mathbb{R}^3$. No additional regularity of the potential is required. The proof uses a convexified mixed Carleman estimate whose norms are designed together with the final time truncation. Integration over spatial centers removes the heat-kernel weight and yields quantitative comparisons of the full solution norms. To prove quadratic sharpness in the Euclidean energy class, we construct a time-dependent function $g$ that localizes Meshkov's periodic parabolic example while keeping the induced potential bounded across the zeros of the periodic profile.

math.AP↗
arXiv2026

GOLF: Global Observation with Local Focus for Calibration-Aware Stereo Interaction Field Estimation

We present GOLF, the first-place solution to the SHOW3D Interaction Field Estimation Challenge at HANDS@ECCV 2026. Given synchronized egocentric stereo views, the task is to predict a 3D vector from each of 21 hand joints to the closest point on the manipulated object. GOLF combines dense global context, locally sampled hand/object evidence, and common-frame Plücker-ray geometry. We adapt DINOv3 ViT-H+/16 with LoRA and trainable LayerNorm parameters, then jointly decode both interaction fields. Our primary model achieves an official score of 27.61 and a mean ADE of 27.96 mm on the hidden test set. An equal-weight ensemble with a complementary directly fine-tuned variant improves these results to an official score of 27.47 and a mean ADE of 27.82 mm, securing first place.

cs.CV↗
arXiv2026

A pilot study on the CSST astrometric capability: Detecting astrometric binaries with Gaia synergy via simulated data

Context. The China Space-station Survey Telescope (CSST) will provide deep, wide-field epoch astrometry during its 10-year mission. Astrometric binary orbits constrain the masses of stellar and compact-object components. Orbital recovery depends on astrometric precision and temporal coverage. Combining CSST and Gaia data extends the baseline and improves binary detection. Aims. We evaluate CSST, Gaia, and joint astrometry for binary-candidate selection and 12-parameter (12p) orbit fitting at faint magnitudes ($g>17.8$). We also test how regular CSST cadences affect the yield of 12p fits satisfying our criteria. Methods. We constructed a mock catalog, simulated CSST and Gaia epoch astrometry, and fitted five-parameter (5p) single-star models to derive astrometric diagnostics, proper-motion anomaly features, and observational-sampling features. A four-stage histogram-based gradient-boosting classifier used these features to select candidates for 12p orbit fitting and assessment. Results. On the independent test set, the classifier reaches a precision of 0.802 and a recall of 0.181 among eligible true binaries. In the scenario-specific fitted samples, joint astrometry raises the fiducial fraction from 6.76% for Gaia alone to 10.46%; for fitted binaries with $P_{\rm true}>15{\rm yr}$, it rises from 2.37% to 6.78%. The current CSST schedule yields few fiducial fits, while idealized regular cadences increase the yield mainly at $g\lesssim21$. Conclusions. In the simulation, joint CSST and Gaia epoch astrometry yields higher fractions of fitted unresolved binaries satisfying the stated criteria than Gaia-only solution. A practical strategy is to select candidates from 5p diagnostics and astrometric anomalies, obtain more regular CSST follow-up observations, and then fit 12p orbital models and apply the selection criteria.

astro-ph.IM↗
arXiv2026

Multi-Task Learning with Covariate-Overlap Regularization

Multi-task learning improves data efficiency by sharing information across related tasks, but indiscriminate sharing can be harmful when their covariate distributions and response relationships differ. We propose COVariate-ovERlap regularized multi-task learning (COVER) to address covariate and posterior heterogeneity. The model combines a common component function with a shared neural representation and low-dimensional task-specific coefficients. Taskwise second-moment matrices summarize covariate heterogeneity and determine the strength of coefficient integration in each representation direction. We derive a covariate-overlap penalty by minimizing the total squared change in two task predictors when their coefficients are replaced by one auxiliary coefficient. An equivalent auxiliary formulation supports end-to-end training without matrix inversion. An exact fixed-representation bias--variance decomposition quantifies how covariate overlap controls variance reduction and how posterior heterogeneity determines shrinkage bias. Global and localized end-to-end oracle inequalities account for jointly learning the neural functions and estimating the overlap matrices from the same observations. We give explicit neural-network rates and sharpen the stochastic prediction term when the regularized oracle risk and overlap-estimation error are small. Simulations across diverse heterogeneity settings show competitive performance against deep-learning and statistical data-integration methods, with the largest gains under joint heterogeneity. In a GTEx central-nervous-system analysis, COVER achieves the lowest response-averaged prediction error among the compared methods and reveals tissue-pair integration patterns.

stat.ML↗
arXiv2026

SPD: Single Pass Decoding for Generative Reranking

Large language models (LLMs) achieve state-of-the-art generative ranking quality, but the ranking they produce must be decoded, and autoregressive decoding spends one sequential forward pass per emitted token. We observe that the only tokens a ranker must emit are the $N$ ordinal values naming the items in ranked order, and that this narrow, permutation-structured output format admits decoding strategies which are much more efficient than left-to-right generation. We introduce SPD (Single Forward Pass), a format-specialized decoding strategy that decodes all $N$ ordinals in $O(1)$ forward passes. SPD reads an $N \times K$ item-position score matrix off the LLM's prefill hidden states with a lightweight self-attention head, then decodes the ordinals as the optimal bipartite assignment of that matrix via the Hungarian algorithm, yielding a valid permutation by construction rather than by repair. Through a systematic study of training signals and backbone adaptation, we show that LoRA-based fine-tuning combined with auto-regressive LLM ranking distillation reaches 28 ms end-to-end inference, a speed-up of 64x while maintaining ranking quality on par with the teacher. We provide a complete ablation decomposing the contributions of architecture, training signal, and backbone adaptation. Our framework connects generative ranking to combinatorial optimization, opening a path toward other $O(1)$-decode mechanisms for real-time ranking.

cs.LG↗
arXiv2026

Distributional Treatment Effect Transportability across Heterogeneous Sites

We study distributional transportability of treatment effects in a ``cross-site, one-armed target" design, where both treated and control units are observed in a source site, but only control units are observed in a target site. Our object of interest is not estimating the average treatment effect, but recovering the full treated distribution in the target site using transfer knowledge from the source site, while allowing cross-site heterogeneity in measurement systems, observed features, outcome reporting, population composition, and latent contextual factors. We model cross-site heterogeneity through a transformation between the sites that transports the joint feature--outcome distributions. This transformation is learned from comparing the observed control samples in the source and target sites, using an optimal transport criterion. The learned transformation is then applied to the source treated sample to construct a synthetic sample from the target treated distribution. We establish convergence of the resulting synthetic empirical distribution to the target treated distribution. Simulation studies across multiple data-generating scenarios and a real-world application to patient-derived xenograft (PDX) data demonstrate that our framework recovers the full distributional properties of the target treated population.

stat.ME↗
arXiv2026

FinRiskAtlas: Decision-Aligned Evaluation of Large Language Models for Financial Risk Review

Deploying large language models for professional financial review requires more than measuring general financial competence: models must perform the specific review operation required by a workflow and determine whether available evidence is sufficient for a defensible decision. Existing financial benchmarks cover knowledge, reasoning, compliance, and professional tasks, but their evaluation units are often organized around datasets or task formulations rather than the decisions that deployed systems support. We introduce FinRiskAtlas, a Chinese-language benchmark that evaluates financial LLMs along two complementary dimensions: operation execution under fixed evidence states and evidence-state control under evolving review conditions. The static benchmark contains 9,742 instances across 53 task families, including 42 Domain Knowledge families and eleven downstream review operations defined by explicit evaluation contracts. FinRisk-Ask extends this framework through offline replay of 680 pre-action states from 104 de-identified professional trajectories, withholding future evidence during inference and using it only to construct expert-verified evidence targets. Across 33 model configurations, operation-level evaluation yields non-redundant rankings (mean pairwise Spearman correlation 0.42 across downstream operations), and knowledge-based shortlisting can incur up to 18.01 points of regret on individual operations. FinRisk-Ask further shows that entering the Ask branch more frequently does not necessarily improve request targeting or end-to-end evidence acquisition. These results show that broad financial capability scores do not fully capture where models are reliable in professional workflows, motivating evaluation units aligned with the decisions and evidence states that deployed systems must support.

cs.AI↗
arXiv2026

Representation Finetuning for Continual Learning

The world is inherently dynamic, and continual learning aims to enable models to adapt to ever-evolving data streams. While pre-trained models have shown powerful performance in continual learning, they still require finetuning to adapt effectively to downstream tasks. However, prevailing Parameter-Efficient Fine-Tuning (PEFT) methods operate through empirical, black-box optimization at the weight level. These approaches lack explicit control over representation drift, leading to sensitivity to domain shifts and catastrophic forgetting in continual learning scenarios. In this work, we introduce Continual Representation Learning (CoRe), a novel framework that for the first time shifts the finetuning paradigm from weight space to representation space. Unlike conventional methods, CoRe performs task-specific interventions within a low-rank linear subspace of hidden representations, adopting a learning process with explicit objectives, which ensures stability for past tasks while maintaining plasticity for new ones. By constraining updates to a low-rank subspace, CoRe achieves exceptional parameter efficiency. Extensive experiments across multiple continual learning benchmarks demonstrate that CoRe not only preserves parameter efficiency but also significantly outperforms existing state-of-the-art methods. Our work introduces representation finetuning as a new, more effective and interpretable paradigm for continual learning.

cs.LG↗