arXiv subjects

Kaiyu Wu

Publications and source records attributed to Kaiyu Wu.

7 recordsLinked to original sources

arXiv2026

Weather-R1: Logically Consistent Reinforcement Fine-Tuning for Multimodal Reasoning in Meteorology

While Vision Language Models (VLMs) show advancing reasoning capabilities, their application in meteorology is constrained by a domain gap and a reasoning faithfulness gap. Specifically, mainstream Reinforcement Fine-Tuning (RFT) can induce Self-Contradictory Reasoning (Self-Contra), where the model's reasoning contradicts its final answer, which is unacceptable in such a high-stakes domain. To address these challenges, we construct WeatherQA, a novel multimodal reasoning benchmark in meteorology. We also propose Logically Consistent Reinforcement Fine-Tuning (LoCo-RFT), which resolves Self-Contra by introducing a logical consistency reward. Furthermore, we introduce Weather-R1, the first reasoning VLM with logical faithfulness in meteorology, to the best of our knowledge. Experiments demonstrate that Weather-R1 improves performance on WeatherQA by 9.8 percentage points over the baseline, outperforming Supervised Fine-Tuning and RFT, and even surpassing the original Qwen2.5-VL-32B. These results highlight the effectiveness of our LoCo-RFT and the superiority of Weather-R1. Our benchmark and code are available at https://github.com/Marcowky/Weather-R1.

cs.CV↗

arXiv2025

Enhancing Visual Programming for Visual Reasoning via Probabilistic Graphs

Recently, Visual Programming (VP) based on large language models (LLMs) has rapidly developed and demonstrated significant potential in complex Visual Reasoning (VR) tasks. Previous works to enhance VP have primarily focused on improving the quality of LLM-generated visual programs. However, they have neglected to optimize the VP-invoked pre-trained models, which serve as modules for the visual sub-tasks decomposed from the targeted tasks by VP. The difficulty is that there are only final labels of targeted VR tasks rather than labels of sub-tasks. Besides, the non-differentiable nature of VP impedes the direct use of efficient gradient-based optimization methods to leverage final labels for end-to-end learning of the entire VP framework. To overcome these issues, we propose EVPG, a method to Enhance Visual Programming for visual reasoning via Probabilistic Graphs. Specifically, we creatively build a directed probabilistic graph according to the variable dependency relationships during the VP executing process, which reconstructs the non-differentiable VP executing process into a differentiable exact probability inference process on this directed probabilistic graph. As a result, this enables the VP framework to utilize the final labels for efficient, gradient-based optimization in end-to-end supervised learning on targeted VR tasks. Extensive and comprehensive experiments demonstrate the effectiveness and advantages of our EVPG, showing significant performance improvements for VP on three classical complex VR tasks: GQA, NLVRv2, and Open Images.

cs.CV↗

arXiv2022

Shortest Beer Path Queries in Interval Graphs

Our interest is in paths between pairs of vertices that go through at least one of a subset of the vertices known as beer vertices. Such a path is called a beer path, and the beer distance between two vertices is the length of the shortest beer path. We show that we can represent unweighted interval graphs using $2n \log n + O(n) + O(|B|\log n)$ bits where $|B|$ is the number of beer vertices. This data structure answers beer distance queries in $O(\log^\varepsilon n)$ time for any constant $\varepsilon > 0$ and shortest beer path queries in $O(\log^\varepsilon n + d)$ time, where $d$ is the beer distance between the two nodes. We also show that proper interval graphs may be represented using $3n + o(n)$ bits to support beer distance queries in $O(f(n)\log n)$ time for any $f(n) \in ω(1)$ and shortest beer path queries in $O(d)$ time. All of these results also have time-space trade-offs. Lastly we show that the information theoretic lower bound for beer proper interval graphs is very close to the space of our structure, namely $\log(4+2\sqrt{3})n - o(n)$ (or about $ 2.9 n$) bits.

cs.DS↗

arXiv2020

Distance Oracles for Interval Graphs via Breadth-First Rank/Select in Succinct Trees

We present the first succinct distance oracles for (unweighted) interval graphs and related classes of graphs, using a novel succinct data structure for ordinal trees that supports the mapping between preorder (i.e., depth-first) ranks and level-order (breadth-first) ranks of nodes in constant time. Our distance oracles for interval graphs also support navigation queries -- testing adjacency, computing node degrees, neighborhoods, and shortest paths -- all in optimal time. Our technique also yields optimal distance oracles for proper interval graphs (unit-interval graphs) and circular-arc graphs. Our tree data structure supports all operations provided by different approaches in previous work, as well as mapping to and from level-order ranks and retrieving the last (first) internal node before (after) a given node in a level-order traversal, all in constant time.

cs.DS↗

arXiv2016

The Dixmier-Moeglin equivalence for extensions of scalars and Ore extensions

An algebra $A$ satisfies the Dixmier-Moeglin equivalence if we have the equivalences: $$P~{\rm primitive}\iff P~{\rm rational}\iff P ~{\rm locally~closed~}\qquad~{\rm for}~P\in {\rm Spec}(A).$$ We study the robustness of the Dixmier-Moeglin equivalence under extension of scalars and under the formation of Ore extensions. In particular, we show that the Dixmier-Moeglin equivalence is preserved under base change for finitely generated complex noetherian algebras. We also study Ore extensions of finitely generated complex noetherian algebras $A$. If $T:A\to A$ is either a $\mathbb{C}$-algebra automorphism or a $\mathbb{C}$-linear derivation of $A$, we say that $T$ is \emph{frame-preserving} if there exists a finite-dimensional subspace $V\subseteq A$ that generates $A$ as an algebra such that $T(V)\subseteq V$. We show that if $A$ is of finite Gelfand-Kirillov dimension and has the property that all prime ideals of $A$ are completely prime and $A$ satisfies the Dixmier-Moeglin equivalence then the Ore extension $A[x;T]$ satisfies the Dixmier-Moeglin equivalence whenever $T$ is a frame-preserving derivation or automorphism.

math.RA↗

arXiv2014

Moments of zeta functions associated to hyperelliptic curves over finite fields

Let $q$ be an odd prime power, and $H_{d,q}$ denote the set of square-free monic polynomials $D(x) \in F_q[x]$ of degree $d$. Katz and Sarnak showed that the moments, over $H_{d,q}$, of the zeta functions associated to the curves $y^2=D(x)$, evaluated at the central point, tend, as $q \to \infty$, to the moments of characteristic polynomials, evaluated at the central point, of matrices in $USp(2\lfloor (d-1)/2 \rfloor)$. Using techniques that were originally developed for studying moments of $L$-functions over number fields, Andrade and Keating conjectured an asymptotic formula for the moments for $q$ fixed and $d \to \infty$. We provide theoretical and numerical evidence in favour of their conjecture. In some cases we are able to work out exact formulas for the moments and use these to precisely determine the size of the remainder term in the predicted moments.

math.NT↗

arXiv2013

Low-power photothermal probing of single plasmonic nanostructures with nanomechanical string resonators

We demonstrate the direct photothermal probing and mapping of single plasmonic nanostructures via the temperature induced detuning of nanomechanical string resonators. Single Au nanoslits are illuminated with a low-power polarized focused laser beam (λ = 633 nm). Polarization dependent heat generation in gold nanoslits is then imaged with high sensitivity. A sensitivity of -4.1 ppm/nW with respect to the illuminated light (beam diameter 5.0pm0.8 μm) is determined for a single nanoslit (1 μm long and 53 nm wide), which equals to a total light absorption of 16%. This results in a heating of 0.5 K for an illuminance of 8 nW/μm2. Our results show that nanomechanical resonators are a unique and robust analysis tool for the low-power investigation of thermoplasmonic effects in plasmonic hot spots.

cond-mat.mes-hall↗