Searcharxiv⌕ Search

arXiv subjects

Kai Yang

Publications and source records attributed to Kai Yang.

At least 37 records · Page 2Linked to original sources
arXiv2026

TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for enhancing reasoning and agentic behavior in large language models. However, rollout-intensive policy optimization is often limited by insufficient reward contrast, arising when overly simple or complex prompts generate low-variance feedback and when outcome-only rewards assign the same terminal assessment to every decision in a multi-turn rollout. Past efforts have focused on allocating available rollout resources to promising prompts, yet they only leverage sample informativeness at the prompt level and neglect variation in prefix-level informativeness across turns within the same rollout. This work targets multi-turn agentic RL by modeling each ReAct-style thought-action-observation turn as a semantically distinct node, allowing budget allocation to extend from prompt roots to turn-level prefixes with further continuations, which naturally forms tree-structured rollouts. We introduce Tree Rollout Allocation for Contrastive Exploration (TRACE), a unified rollout allocation framework that enhances reward contrast within a fixed sampling budget. Technically, TRACE allocates rollout budget to both prompt roots and intermediate prefixes that are most likely to yield mixed terminal rewards. A shared generalizable predictor estimates conditional success probability at these anchors from prefix histories to guide this allocation. The resulting adaptive tree structure enriches outcome-only feedback and amplifies the policy-update signal. Empirically, TRACE achieves competitive performance and efficiency gains on typical agentic benchmarks, e.g., improving Qwen3-14B Multi-Hop QA average accuracy by 2.8 points over competitive baselines at equal sampling cost.

cs.LG↗
arXiv2026

A new matrix-variate integer-valued autoregressive process with matrical negative binomial thinning

To address the overdispersion problem in matrix-variate integer-valued time series data arising in sociology, medicine, and related fields, this paper proposes a matrix integer-valued autoregressive model based on the negative binomial thinning operator. By introducing left and right matricial negative binomial thinning operators, the proposed model not only preserves the integer-valued nature of the data but also effectively handles overdispersion. The probabilistic and statistical properties of the proposed model are systematically investigated. Two estimation methods, namely projection estimation and iterative conditional least squares estimation are developed, and the corresponding asymptotic theories are established. Simulation studies provide concrete numerical results to evaluate the finite-sample performance of the estimators. Real data analysis demonstrates that the proposed model outperforms both the continuous matrix autoregressive model and the multivariate integer-valued autoregressive model in fitting matrix-variate integer-valued time series, while also proving effective in accommodating overdispersed count data.

math.ST↗
arXiv2026

Maximum Tsallis Entropy Distributions for Robust and Efficient Sparse Learning from Correlated Data

This paper addresses the limitations of Gaussian distribution assumptions in statistical sparse learning, particularly in modeling correlated and heterogeneous data. Conventional Gaussian models often lack robustness towards outliers and underlying distribution assumptions. To overcome these limitations, we propose the use of the $q$Gaussian distribution, derived from Tsallis entropy maximization, as a robust alternative. This is notably relevant in biostatistics, where the presence of correlated observations and heterogeneity, such as in genetic and longitudinal studies, is prevalent. Our contributions include modeling of correlated data through the re-derived multivariate probability density function from Tsallis entropy maximization, thereby addressing the limitations inherent in conventional Gaussian models. Furthermore, we introduce a novel framework that adapts numerical methods designed to find equilibria in flows to tackle composite optimization problems prevalent in statistical sparse learning. Applying this framework to the Hager-Zhang conjugate gradient algorithm \cite{Hager2005}, we develop a numerically stable and efficient algorithm for sparse statistical learning. The $q$Gaussian distribution, informed by the principle of maximizing Tsallis entropy, presents a viable and flexible alternative to Gaussian-based methods. This paper not only contributes to the theoretical understanding of statistical distributions and optimization techniques, but also paves the way for practical data analysis.

math.OC↗
arXiv2026

Can Retrievers Find the Same Paper from Different Aspects? A Multi-Aspect Full-Paper Scientific Retrieval Benchmark

Scientific papers contain multiple searchable facets such as background, methods. However, many paper retrieval benchmarks merely evaluate individual query-paper relevance, while overlooking other facets of the same paper. To bridge this gap, we introduce MAPLE, an expert-validated benchmark for multi-aspect, full-paper retrieval that evaluates whether retrievers can consistently recover the same paper from queries targeting its motivation, method, and experimental findings. MAPLE contains 2,095 queries about recent ML and NLP papers, grounded in both textual and multimodal content. We further propose MAPLE-Synth, a retrieval-based in-context learning pipeline that leverages OpenReview discussions and human-written query exemplars to generate realistic queries reflecting researchers' interests in different aspects of a paper. Our expert validation shows that these queries are comparable in realism to human-written queries and highly relevant to the target papers. Experiments across lexical, scientific-domain, general-purpose text, and multimodal retrievers reveal a substantial gap between retrieving a paper from any one aspect and retrieving it from all aspects: the strongest model achieves 98.1% AnyAspect@20 but only 15.7% AllAspect@20. Experiment/result queries and table-referenced queries are particularly difficult across retrievers. Although multi-chunk aggregation improves multi-aspect paper retrieval, considerable failures persist. MAPLE provides a testbed for evaluating and developing retrievers that represent scientific papers more comprehensively.

cs.IR↗
arXiv2026

SGEMM-cube: Precision-Recovery FP32 GEMM Approximation on Ascend NPUs with FP16 Matrix Engines

Modern AI accelerators provide high-throughput low-precision matrix engines, but often lack efficient support for FP32 GEMM. This paper presents SGEMM-cube, an FP32-accuracy GEMM approximation for Ascend NPUs built on FP16 Cube units. Following the fixed-length two-word splitting of Ootomo and Yokota, each FP32 operand is represented by an FP16 high component and a scaled FP16 residual. The product is reconstructed from three FP16 GEMMs while omitting the residual-residual term; the method therefore targets FP32-level accuracy rather than bit-exact emulation. Under an RN FP32-accumulation model, we provide a componentwise error analysis showing that the omitted term is no larger than the rounding error of a short FP32 inner product and that, for practically relevant inner-product lengths, the overall error is dominated by ordinary FP32 accumulation. We further analyze residual underflow and scaling under round-to-nearest conversion, compare two accumulation orders, and adapt L1-aware blocking and double buffering to Ascend's software-managed memory hierarchy. On Ascend 910A, SGEMM-cube is substantially more accurate than native FP16 GEMM, is comparable to the tested OpenBLAS FP32 SGEMM baseline for the evaluated input distributions and exponent range, and reaches 65.3\,TFLOP/s, or 77\% of the three-GEMM FP32-equivalent peak.

cs.DC↗
arXiv2026

GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs

On-policy rollout methods such as GRPO are central to post-training of large language models, yet they frequently suffer from training instabilities, cross-task capability degradation, and response-length inflation. Although prior work has characterized the subspace geometry of aggregate updates, the stepwise variation of this geometry and its relationship to model performance remain unclear. We introduce Principal-Subspace Overlap, a dimension-corrected measure of individual rollout updates relative to the dominant singular subspaces of pretrained weights. Despite low average overlap, transient spikes often precede performance degradation. To address this, we propose GCPO (Geometrically Constrained Policy Optimization), which applies hard bilateral orthogonal projections to constrain updates to the complementary subspaces, preventing such excursions by construction. Across mathematical reasoning, code generation, and tool-use tasks on Qwen3-8B and GLM4-9B, GCPO consistently outperforms GRPO and recent variants, including DAPO and GSPO, improving over the base models and the strongest baseline by up to 27.69 and 2.37 points, respectively. Furthermore, GCPO preserves general capabilities, eliminates response-length inflation, and stabilizes policy entropy. Our findings provide a new diagnostic lens and a principled design perspective for stable reinforcement learning post-training.

cs.LG↗
arXiv2026

First-order Constrained Trilevel Optimization Over Distributed Networks for Robust Coreset Selection

With the rapid advancement of the Internet of Things (IoT), massive amounts of data are generated across distributed edge networks. Training models on full data incurs significant computational overhead and storage bottlenecks, rendering coreset selection a critical paradigm. Furthermore, given the privacy-sensitive nature of local data and the escalating demand for model robustness in real-world deployments, developing an effective distributed optimization framework for robust coreset selection is vital, yet remains largely unexplored. To this end, this work first characterizes the hierarchical dependencies among coreset selection, robust optimization, and distributed learning, and formulates the distributed robust coreset selection as a trilevel optimization problem with level-wise constraints. Furthermore, to effectively solve the trilevel problem in a distributed manner, the \underline{F}ederated \underline{F}irst-order \underline{C}onstrained \underline{T}rilevel \underline{O}ptimization (F$^2$CTO) is proposed, which synergistically integrates a hierarchical composite value-function reformulation and a distributed alternating projected gradient algorithm. To the best of our knowledge, F$^2$CTO is the first method developed for distributed robust coreset selection, as well as the first distributed optimization approach for trilevel optimization problems with level-wise constraints. Additionally, we prove that the proposed method achieves a non-asymptotic convergence rate of $\mathcal{O}(ε^{-3/2})$ for finding an $ε$-stationary point. Extensive empirical evaluations on reliable continual learning demonstrate the effectiveness and efficiency of the proposed F$^2$CTO.

cs.LG↗
arXiv2026

Silent Failures in Multimodal Agentic Search:A Diagnostic Taxonomy and Cross-Judge Evaluation

Multimodal agentic search systems increasingly rely on external tools to answer knowledge-intensive visual questions. However, existing evaluations mainly focus on final-answer accuracy and may miss failures in the search trajectory. In this work, we study such hidden reliability issues as silent failures. We introduce a six-category taxonomy covering modality shortcuts, phantom grounding, wrong-evidence-right-answer cases, over-retrieval laundering, cross-modal contradiction, and provenance hallucination. Based on this taxonomy, we build a trajectory-level diagnostic pipeline that evaluates both answer correctness and evidence-grounding quality under a unified ReAct-style scaffold. Experiments on MMSearch-Plus trajectories across four frontier multimodal models show that surface accuracy consistently overestimates true trajectory-level correctness. We further use cross-judge validation, blank-image stress tests, and tool ablations to show that silent failures are capability-dependent and often shift rather than disappear. Home-page: https://github.com/DingWu1021/silent-failures-multimodal-agentic-search

cs.AI↗
arXiv2026

Ascend to Science: Exploration of AI Chips for Scientific Computing

The rapid rise of AI-oriented accelerators has reshaped compute systems around low-precision tensor engines, raising a practical question for the HPC community: under what conditions can such hardware support scientific workloads that demand numerical robustness, irregular memory access, and scalability? Using the Ascend 910 NPU series as a representative tensor-centric platform, we characterize precision, execution, and memory-hierarchy bottlenecks that hinder the direct deployment of scientific codes. We then develop and evaluate workload-specific mappings across five application studies -- HPL-MxP, LRSVD, SGEMM-cube, PQSim, and SMC-X -- combining heterogeneous execution, mixed-precision numerical formulations, precision emulation, hierarchical memory orchestration, and communication--computation overlap. These studies show that AI-native NPUs can achieve numerical robustness, competitive performance, and satisfactory scalability when numerical formulation, execution placement, and data movement are addressed in a coordinated manner. Our results provide a state-of-the-practice case study of how scientific workloads can be adapted to tensor-centric architectures, while distinguishing transferable optimization principles from Ascend-specific implementation details.

cs.DC↗
arXiv2026

Joint Chirp Parameter Selection and Low-Complexity MMSE Receiver Design for AFDM Systems

Affine frequency division multiplexing (AFDM) has emerged as a promising waveform against doubly selective channels under high-mobility communication scenarios. Optimal chirp parameter selection and reduced-complexity receiver design in AFDM are essential for achieving satisfactory bit error rate (BER) performance with low computational complexity. In this paper, we investigate the joint optimization of chirp-parameter selection and low-complexity minimum mean square error (MMSE)-based receiver design by exploiting the structural characteristics of the AFDM effective channel matrix (ECM). First, a simplified BER performance metric is derived by leveraging the diagonal and circulant structure of the discrete affine Fourier transformation (DAFT), based on which a fast circulant-diagonal aggregation (FCDA) algorithm is developed for efficient $c_1$ selection. Then, a low-complexity banded MMSE (LC-BMMSE) receiver is developed by constructing a cyclic-banded ECM through path-wise structured sparsification, where banded Cholesky factorization is employed to avoid direct matrix inversion. Building upon the proposed BER metric and the LC-BMMSE receiver, a hierarchical-search-based joint chirp parameter and structured sparsification (HS-JCPS) algorithm is further proposed to jointly optimize the chirp parameter and sparsification pattern under a given complexity constraint. Simulation results demonstrate that the proposed FCDA reduces the search time for the optimal $c_1$ by an order of magnitude compared with using a BER-based criterion. Moreover, the proposed HS-JCPS algorithm with the LC-BMMSE receiver can identify a near-optimal $c_1$, while attaining a superior performance-complexity tradeoff.

eess.SP↗
arXiv2026

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search attempts fail to yield useful evidence, current single- and multi-agent systems can become trapped in repetitive loops, wasting search budgets and ultimately compromising the quality and completeness of the final output. We introduce SearchOS, a system-level multi-agent framework that turns fragile, implicit search progress into explicit, persistent, and shared state. First, we formulate open-domain information seeking as relational schema completion with grounded citations, where agents discover entities, populate attributes across linked tables, and anchor each value to source evidence. Then we design Search-Oriented Context Management (SOCM), which externalizes the evolving state into Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory. Built on SOCM, SearchOS applies a pipeline-parallel scheduling mechanism that overlaps the execution of sub-agents and continuously refills freed slots with tasks targeting unresolved coverage gaps to improve utilization and throughput. To schedule and control the execution of search agents, SearchOS introduces a Search Tool Middleware Harness that intercepts model and tool interactions to record grounded evidence and react to stalls or budget exhaustion, and provides a reusable hierarchical skill system comprising strategy and access skills to augment the agents' search process and avoid repeating failed search patterns across runs. On WideSearch and GISA, SearchOS leads all metrics among the evaluated single- and multi-agent baselines, paving the way toward robust information-seeking collaboration.

cs.AI↗
arXiv2026

Rydberg Atomic Quantum Radio: A Comprehensive Survey From Wireless Communication Perspective

Next-generation space-air-ground-sea integrated networks (SAGSIN) impose unprecedented demands on advanced radio frequency (RF) receivers for full-spectrum agility, ultra-high sensitivity, and anti-jamming resilience, pushing conventional electronic receivers to their physical limits. To address these challenges, the Rydberg atomic quantum (RAQ) radio has emerged as a promising quantum-enabled receiver paradigm that directly maps electromagnetic fields onto atomic quantum states, offering an alternative to alleviate bottlenecks of conventional RF front ends. To provide a clear research roadmap, this survey presents a comprehensive review of RAQ radios by bridging atomic physics and wireless communications. Specifically, we first introduce the underlying quantum mechanisms, representative architectures, and atomic response models of RAQ radio. On this basis, state-of-the-art techniques for enhancing sensitivity, instantaneous bandwidth, and operating frequency are systematically reviewed, with particular emphasis on the inherent trade-offs among these key metrics. To connect quantum response with communication theory, we further analyze equivalent channel modeling frameworks for characterizing systematic performance limits. From the wireless communication perspective, some RAQ-enabled advanced technologies including cognitive, interference-resilient, low-frequency and multiple-input multiple-output (MIMO) communications are reviewed, alongside emerging deployment scenarios such as satellite networks, integrated sensing and communications, and reconfigurable intelligent surface-assisted systems. Finally, we identify open challenges and provide potential future directions of RAQ radio to inspire the further exploration.

eess.SP↗
arXiv2026

A Turán Theorem for Cayley Graphs

In this note, we give a Turán theorem for Cayley graphs $\Cay(\Z_p,S)$ over prime cyclic groups $\Z_p$. For a graph $F$ and a finite abelian group $G$, define the Cayley--Turán number by \[ \exCay(F,G) = \max\{|S|:S=-S\subseteq G\setminus\{0\},\ \Cay(G,S)\text{ is }F\text{-free}\}. \] Using a polynomial method, we prove that for every odd prime $p$ and every $1\le r\le p-1$, \[ \exCay(K_{r+1},\Z_p) = p-1-2\left\lfloor\frac{p}{r+1}\right\rfloor . \] The extremal construction is the complement of the short-difference interval \[ D_0=\{0,\pm1,\ldots,\pm\lfloor p/(r+1)\rfloor\}. \] We also discuss what changes for general finite abelian groups, showing why the exact prime-cyclic formula does not extend verbatim to composite cyclic groups.

math.CO↗
arXiv2026

Popular Differences and the Croot--Lev Half-Threshold Problem

Let $A$ be a finite non-empty subset of an abelian group $G$, and let $r_A(d)=|\{(a,a')\in A^2:a-a'=d\}|$. Croot and Lev asked whether the pointwise half-threshold condition $r_A(d)\ge |A|/2$ for every $d\in A-A$ forces $A-A$ to be either a subgroup or a union of three cosets. We resolve this open problem in its sharp general form by identifying the essential obstruction: the statement is false in arbitrary abelian groups, but becomes true after excluding non-zero two-torsion. More precisely, if $G$ is two-torsion-free and the half-threshold condition holds, then either $A-A$ is a finite subgroup of $G$, or there are a finite subgroup $H\le G$ and elements $x,g\in G$ such that \[ A=(x+H)\cup(x+g+H). \] The two-torsion-free hypothesis is essential: for every $r\ge1$ we construct $A\subseteq\F_2^{2r+1}$ with $A-A=\F_2^{2r+1}\setminus\{t\}$ such that every non-zero represented difference has exactly $|A|/2$ representations, giving genuine counterexamples to the Croot--Lev conclusion. The proof of the positive result combines a Kneser quotient reduction with Lev's formulation of Kemperman's critical-pair theory.

math.CO↗
arXiv2026

Odd cycles in symmetric Cayley graphs on prime cyclic groups

Let $p$ be an odd prime and let $S\subseteq \Z_p$ be symmetric with $0\notin S$. Let $\Cay(\Z_p,S)$ be the undirected Cayley graph on $\Z_p$ in which $x$ and $y$ are adjacent if and only if $x-y\in S$. For $1\le \ell\le (p-1)/2$, define \[ \ex_{\Cay}(C_{2\ell+1},\Z_p)=\max\{|S|: S=-S,\ 0\notin S,\ \Cay(\Z_p,S)\text{ contains no }C_{2\ell+1}\}. \] Confirming a conjecture of Cashman and Kelley, we prove that if $p=2\ell+1$, then $\ex_{\Cay}(C_{2\ell+1},\Z_p)=0$, while if $p>2\ell+1$, then \[ \ex_{\Cay}(C_{2\ell+1},\Z_p)=2\floor{\frac{p+2\ell+1}{2(2\ell+1)}}. \] The proof combines a sharp additive zero-sum odd-girth argument with weak odd pancyclicity to transfer the result from odd-girth exclusion to fixed odd-cycle exclusion. We also give a canonical extremal family, an exact extremality criterion in terms of odd zero-sum avoidance, and an example showing that extremizers need not be dilates of the canonical construction.

math.CO↗
arXiv2026

HOTDISK. Finding Massive Protostellar Disks with Water and Refractory Molecular Species

We present high-angular-resolution ($\sim0.05^{\prime\prime}$, $\sim 60-250$ au) ALMA Band~6 observations from the HOTDISK project (Hot-Origin Tracer survey of DISKs of massive protostars) aimed at investigating the "hot-disk" chemical pattern traced by vibrationally excited water, NaCl, SiS, and SiO in the innermost regions around massive protostars. Ten targets were selected based on strong CH$_3$CN emission exhibiting clear rotational signatures and centrally concentrated SiO emission from lower-resolution observations. We detect vibrationally excited water emission toward 7 of the 10 sources. In all detections, the blueshifted and redshifted components are compact and located on opposite sides of the 1.3 mm continuum peak, with velocity gradients approximately perpendicular to the outflow axes, consistent with rotation on disk scales. Emission from NaCl and SiS is detected toward 5 of these 7 sources and exhibits similar kinematics, further supporting the presence of compact rotating structures. In contrast, commonly used hot-core tracers (e.g., CH$_3$CN and SO$_2$) primarily probe larger-scale envelope gas. These results demonstrate that vibrationally excited water, NaCl, and SiS are powerful tracers of disk structures on $\sim$100 au scales, when observed at sufficient angular resolution and sensitivity. The high detection rate suggests that hot-disk chemical patterns -- and thus compact rotating disks -- are common in massive star-forming regions, at least among sources with well-developed rotating envelopes.

astro-ph.GA↗
arXiv2026

RLVR Datasets and Where to Find Them: Tracing Data Lineage for Better Training Data

The proliferation of Reinforcement Learning from Verifiable Rewards (RLVR) datasets has exacerbated provenance collapse due to unclear lineage among existing datasets. To bridge this fragmented RLVR data landscape, we propose Atomic-source Tracing via Lineage-Aware Search (ATLAS), a systematic framework for tracing RLVR datasets back to their atomic sources, attributing over 99.7% of 1.45M instances to 20 atomic sources. Our analysis reveals that most RLVR datasets are variants of a small set of shared upstream sources, with few introducing genuinely new data, and many facing data contamination risks. These findings naturally motivate us to curate a new RLVR dataset, DAPO++, and to benchmark existing datasets from a lineage-aware perspective. To this end, we propose Source-level Counterfactual Attribution (SCA) as a guiding principle to curate a decontaminated training dataset with concentrated learning signals. Essentially, SCA measures a sample's marginal utility by comparing per-atomic-source RL checkpoints against a shared base model. Building upon these attribution signals, we further design a composite dataset quality score Q that strongly correlates with downstream RLVR performance. Experiments on Qwen3 series models verify that DAPO++ consistently improves performance on held-out benchmarks, while Q reliably predicts downstream RLVR training effectiveness. Our code and data is available at https://github.com/Celine-hxy/ATLAS.

cs.LG↗
arXiv2026

FD-RAG: Federated Dual-System Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) has emerged as a paradigm for grounding large language models in external knowledge, yet most existing RAG systems assume centralized knowledge access and ample computation. These assumptions break down in edge environments, where knowledge is fragmented across devices, raw data cannot be shared, and repeated LLM calls are prohibitively expensive. We propose FD-RAG, a federated dual-system RAG framework that decouples lightweight memory access from on-demand LLM reasoning for decentralized deployment. Specifically, FD-RAG learns semantic-aware adaptive hypergraphs over local corpora and distills them into compact QA memories. At inference time, it answers well-covered queries via direct memory matching and invokes LLM-based reasoning only when necessary, while tracing retrieved memories to hypergraph-grounded evidence. To mitigate cross-device knowledge fragmentation, FD-RAG aggregates anonymized memories across devices without exposing raw documents. Experiments on QA benchmarks show that FD-RAG improves accuracy by up to 7.8\% while reducing latency by 8.4$\times$ compared with strong local and federated baselines. We also provide theoretical analysis establishing an $\mathcal{O}(1/ε^{2})$ convergence rate for the proposed hypergraph learning, supporting its tractable deployment in edge settings.

cs.IR↗