arXiv subjects

Yu Fu

Publications and source records attributed to Yu Fu.

At least 55 records · Page 3Linked to original sources

arXiv2026

LingLanMiDian: Systematic Evaluation of LLMs on TCM Knowledge and Clinical Reasoning

Large language models (LLMs) are advancing rapidly in medical NLP, yet Traditional Chinese Medicine (TCM) with its distinctive ontology, terminology, and reasoning patterns requires domain-faithful evaluation. Existing TCM benchmarks are fragmented in coverage and scale and rely on non-unified or generation-heavy scoring that hinders fair comparison. We present the LingLanMiDian (LingLan) benchmark, a large-scale, expert-curated, multi-task suite that unifies evaluation across knowledge recall, multi-hop reasoning, information extraction, and real-world clinical decision-making. LingLan introduces a consistent metric design, a synonym-tolerant protocol for clinical labels, a per-dataset 400-item Hard subset, and a reframing of diagnosis and treatment recommendation into single-choice decision recognition. We conduct comprehensive, zero-shot evaluations on 14 leading open-source and proprietary LLMs, providing a unified perspective on their strengths and limitations in TCM commonsense knowledge understanding, reasoning, and clinical decision support; critically, the evaluation on Hard subset reveals a substantial gap between current models and human experts in TCM-specialized reasoning. By bridging fundamental knowledge and applied reasoning through standardized evaluation, LingLan establishes a unified, quantitative, and extensible foundation for advancing TCM LLMs and domain-specific medical AI research. All evaluation data and code are available at https://github.com/TCMAI-BJTU/LingLan and http://tcmnlp.com.

cs.AI↗

arXiv2026

Thinking Out of Order: When Output Order Stops Reflecting Reasoning Order in Diffusion Language Models

Autoregressive (AR) language models enforce a fixed left-to-right generation order, creating a fundamental limitation when the required output structure conflicts with natural reasoning (e.g., producing answers before explanations due to presentation or schema constraints). In such cases, AR models must commit to answers before generating intermediate reasoning, and this rigid constraint forces premature commitment. Masked diffusion language models (MDLMs), which iteratively refine all tokens in parallel, offer a way to decouple computation order from output structure. We validate this capability on GSM8K, Math500, and ReasonOrderQA, a benchmark we introduce with controlled difficulty and order-level evaluation. When prompts request answers before reasoning, AR models exhibit large accuracy gaps compared to standard chain-of-thought ordering (up to 67% relative drop), while MDLMs remain stable ($\leq$14% relative drop), a property we term "order robustness". Using ReasonOrderQA, we present evidence that MDLMs achieve order robustness by stabilizing simpler tokens (e.g., reasoning steps) earlier in the diffusion process than complex ones (e.g., final answers), enabling reasoning tokens to stabilize before answer commitment. Finally, we identify failure conditions where this advantage weakens, outlining the limits required for order robustness.

cs.CL↗

arXiv2026

SurfAge-Net: A Hierarchical Surface-Based Network for Interpretable Fine-Grained Brain Age Prediction

Brain age prediction serves as a powerful framework for assessing brain status and detecting deviations associated with neurodevelopmental and neurodegenerative disorders. However, most existing approaches emphasize whole-brain age prediction and therefore overlook the pronounced regional heterogeneity of brain maturation that is crucial for detecting localized atypical trajectories. To address this limitation, we propose a novel spherical surface-based brain age prediction network (SurfAge-Net) that leverages multiple morphological metrics to capture region-specific developmental patterns with enhanced robustness and clinical interpretability. SurfAge-Net establishes a new modeling paradigm by incorporating the connectomic principles of cortical organization: it explicitly models both intra- and inter-hemispheric dependencies through a spatial-channel mixing and a lateralization-aware attention mechanism, enabling the network to characterize the coordinate maturation pattern uniquely associated with each target region. Validated on three fetal and neonatal datasets, SurfAge-Net outperforms existing approaches (global MAE = 0.54, regional MAE = 0.45 in gestational/postmenstrual weeks) and demonstrates strong generalizability across external cohorts. Importantly, it provides spatially precise and biologically interpretable maps of cortical maturation, effectively identifying heterogeneous delays and regional-specific abnormalities in atypical developmental populations. These results established fine-grained brain age prediction as a promising paradigm for advancing neurodevelopmental research and supporting early clinical assessment.

q-bio.NC↗

arXiv2025

Nonperturbative Heavy-Flavor Transport Approach for Hot QCD Matter

The heavy charm and bottom quarks are unique probes of the transport properties of the quark-gluon plasma (QGP) and its hadronization in high-energy nuclear collisions. A key challenge in this context is to embed the interactions of the heavy quarks in the expanding medium compatible with the strong-coupling nature of the QGP, and thus to unravel the underlying microscopic mechanisms. In the present work we progress toward this goal by combining recent $T$-matrix interactions for elastic scattering with an effective transport implementation of gluon radiation, and apply these in a Langevin framework in a viscous hydrodynamic evolution. Hadronization of heavy quarks is evaluated using a modern recombination model with 4-momentum conservation, supplemented with fragmentation constrained by data in proton-proton collisions. Deploying this approach to charm-hadron observables in Pb-Pb collisions at the LHC yields fair agreement with experiment while also identifying areas of further systematic improvement of the simulations and its current input.

nucl-th↗

arXiv2025

Understanding Cellular Noise with Optical Perturbation and Deep Learning

Noise plays a crucial role in the regulation of cellular and organismal function and behavior. Exploring noise's impact is key to understanding fundamental biological processes, such as gene expression, signal transduction, and the mechanisms of development and evolution. Currently, a comprehensive method to quantify dynamical behavior of cellular noise within these biochemical systems is lacking. In this study, we introduce an optically-controlled perturbation system utilizing the light-sensitive Phytochrome B (PhyB) from \textit{Arabidopsis thaliana}, which enables precise noise modulation with high spatial-temporal resolution. Our system exhibits exceptional sensitivity to light, reacting consistently to pulsed light signals, distinguishing it from other photoreceptor-based promoter systems that respond to a single light wavelength. To characterize our system, we developed a stochastic model for phytochromes that accounts for photoactivation/deactivation, thermal reversion, and the dynamics of the light-activated gene promoter system. To precisely control our system, we determined the rate constants for this model using an omniscient deep neural network that can directly map rate constant combinations to time-dependent state joint distributions. By adjusting the activation rates through light intensity and degradation rates via N-terminal mutagenesis, we illustrate that out optical-controlled perturbation can effectively modulate molecular expression level as well as noise. Our results highlight the potential of employing an optically-controlled gene perturbation system as a noise-controlled stimulus source. This approach, when combined with the analytical capabilities of a sophisticated deep neural network, enables the accurate estimation of rate constants from observational data in a broad range of biochemical reaction networks.

q-bio.MN↗

arXiv2025

Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning

Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark. Codes are available at https://github.com/FYYFU/HeadKV

cs.CL↗

arXiv2025

Generalized Eckardt points on del Pezzo surfaces of degree 1

We study intersections of exceptional curves on del Pezzo surfaces of degree 1, motivated by questions in arithmetic geometry. Outside characteristics 2 and 3, at most 10 exceptional curves can intersect in a point. We classify the different ways in which 10 exceptional curves can intersect, construct a new family of surfaces with 10 exceptional curves intersecting in a point, and discuss strategies for finding more such examples.

math.AG↗

arXiv2025

Open heavy-flavor transport and hadronization in heavy-ion collisions

We develop a comprehensive model for heavy-quark evolution in a realistic QGP, from their production in the initial collision to hadronic freeze-out. Heavy-quark transport is described by a Langevin approach including medium-induced radiation, coupled to a 2+1D viscous hydrodynamic bulk evolution. Transport coefficients are obtained from non-perturbative $T$-matrix calculations with resonant correlations near the transition temperature. Hadronization is implemented via two fragmentation+recombination schemes: an improved sudden coalescence model and a resonance recombination model. We present results for key open heavy-flavor observables, i.e., the nuclear modification factor and elliptic flow, and compare to LHC Pb-Pb data at $\sqrt{s_{\NN}}$=5.02\,TeV.}

hep-ph↗

arXiv2025

Jet Energy-Energy Correlator in Cold QCD Matter

We present a study of medium-induced modifications to the energy-energy correlator (EEC) for jets in cold nuclear matter. For electron-nucleus collisions, at leading order in the QCD coupling and in the jet-medium interaction, we derive an analytic expression for the EEC modification as a function of the opening angle and show that the modification is strongest at large angles within the jet cone. The dependence on jet energy, the transport properties of cold nuclear matter, and the in-medium path length is made explicit. We further extend the analysis to gluon jets in proton-nucleus collisions and compare with preliminary proton-lead data from the LHC. Possible effects of comovers on the EEC in proton-nucleus collisions are also discussed.

hep-ph↗

arXiv2025

A Survey of Reinforcement Learning for Large Reasoning Models

In this paper, we survey recent advances in Reinforcement Learning (RL) for reasoning with Large Language Models (LLMs). RL has achieved remarkable success in advancing the frontier of LLM capabilities, particularly in addressing complex logical tasks such as mathematics and coding. As a result, RL has emerged as a foundational methodology for transforming LLMs into LRMs. With the rapid progress of the field, further scaling of RL for LRMs now faces foundational challenges not only in computational resources but also in algorithm design, training data, and infrastructure. To this end, it is timely to revisit the development of this domain, reassess its trajectory, and explore strategies to enhance the scalability of RL toward Artificial SuperIntelligence (ASI). In particular, we examine research applying RL to LLMs and LRMs for reasoning abilities, especially since the release of DeepSeek-R1, including foundational components, core problems, training resources, and downstream applications, to identify future opportunities and directions for this rapidly evolving area. We hope this review will promote future research on RL for broader reasoning models. Github: https://github.com/TsinghuaC3I/Awesome-RL-for-LRMs

cs.CL↗

arXiv2025

Vector Copula Variational Inference and Dependent Block Posterior Approximations

The key to VI is the selection of a tractable density to approximate the Bayesian posterior. For large and complex models a common choice is to assume independence between multivariate blocks in a partition of the parameter space. While this simplifies the problem it can reduce accuracy. This paper proposes using vector copulas to capture dependence between the blocks parsimoniously. Tailored multivariate marginals are constructed using learnable transport maps. We call the resulting joint distribution a ``dependent block posterior'' approximation. Vector copula models are suggested that make tractable and flexible variational approximations. They allow for differing marginals, numbers of blocks, block sizes and forms of between block dependence. They also allow for solution of the variational optimization using efficient stochastic gradient methods. The approach is demonstrated using four different statistical models and 16 datasets which have posteriors that are challenging to approximate. This includes models that use global-local shrinkage priors for regularization, and hierarchical models for smoothing and heteroscedastic time series. In all cases, our method produces more accurate posterior approximations than benchmark VI methods that either assume block independence or factor-based dependence, at limited additional computational cost. A python package implementing the method is available on GitHub at https://github.com/YuFuOliver/VCVI_Rep_PyPackage.

stat.ML↗

arXiv2025

MACE: A Hybrid LLM Serving System with Colocated SLO-aware Continuous Retraining Alignment

Large language models (LLMs) deployed on edge servers are increasingly used in latency-sensitive applications such as personalized assistants, recommendation, and content moderation. However, the non-stationary nature of user data necessitates frequent retraining, which introduces a fundamental tension between inference latency and model accuracy under constrained GPU resources. Existing retraining strategies either delay model updates, over-commit resources to retraining, or overlook iteration-level retraining granularity. In this paper, we identify that iteration-level scheduling is crucial for adapting retraining frequency to model drift without violating service-level objectives (SLOs). We propose MACE, a hybrid LLM system that colocates concurrent inference (prefill, decode) and fine-tuning, with intelligent memory management to maximize task performance while promising inference throughput. MACE leverages the insight that not all model updates equally affect output alignment and allocates GPU cycles accordingly to balance throughput, latency, and update freshness. Our trace-driven evaluation shows that MACE matches or exceeds continuous retraining while reducing inference latency by up to 63% and maintaining throughput under resource constraints. Compared to periodic retraining, MACE improves latency breakdown across prefill, decode, and finetune stages, and sustains GPU utilization above 85% in NVIDIA AGX Orin. These results demonstrate that iteration-level hybrid scheduling is a promising direction for deploying LLMs with continual learning capabilities on edge platforms.

cs.LG↗

arXiv2025

From "What to Eat?" to Perfect Recipe: ChefMind's Chain-of-Exploration for Ambiguous User Intent in Recipe Recommendation

Personalized recipe recommendation faces challenges in handling fuzzy user intent, ensuring semantic accuracy, and providing sufficient detail coverage. We propose ChefMind, a hybrid architecture combining Chain of Exploration (CoE), Knowledge Graph (KG), Retrieval-Augmented Generation (RAG), and a Large Language Model (LLM). CoE refines ambiguous queries into structured conditions, KG offers semantic reasoning and interpretability, RAG supplements contextual culinary details, and LLM integrates outputs into coherent recommendations. We evaluate ChefMind on the Xiachufang dataset and manually annotated queries, comparing it with LLM-only, KG-only, and RAG-only baselines. Results show that ChefMind achieves superior performance in accuracy, relevance, completeness, and clarity, with an average score of 8.7 versus 6.4-6.7 for ablation models. Moreover, it reduces unprocessed queries to 1.6%, demonstrating robustness in handling fuzzy demands.

cs.AI↗

arXiv2025

Inverse IFEval: Can LLMs Unlearn Stubborn Training Conventions to Follow Real Instructions?

Large Language Models (LLMs) achieve strong performance on diverse tasks but often exhibit cognitive inertia, struggling to follow instructions that conflict with the standardized patterns learned during supervised fine-tuning (SFT). To evaluate this limitation, we propose Inverse IFEval, a benchmark that measures models Counter-intuitive Abilitytheir capacity to override training-induced biases and comply with adversarial instructions. Inverse IFEval introduces eight types of such challenges, including Question Correction, Intentional Textual Flaws, Code without Comments, and Counterfactual Answering. Using a human-in-the-loop pipeline, we construct a dataset of 1012 high-quality Chinese and English questions across 23 domains, evaluated under an optimized LLM-as-a-Judge framework. Experiments on existing leading LLMs demonstrate the necessity of our proposed Inverse IFEval benchmark. Our findings emphasize that future alignment efforts should not only pursue fluency and factual correctness but also account for adaptability under unconventional contexts. We hope that Inverse IFEval serves as both a diagnostic tool and a foundation for developing methods that mitigate cognitive inertia, reduce overfitting to narrow patterns, and ultimately enhance the instruction-following reliability of LLMs in diverse and unpredictable real-world scenarios.

cs.CL↗

arXiv2025

Families of isogenous elliptic curves ordered by height

Given a family of products of elliptic curves over a rational curve defined over a number field $K$, and assuming that there exists no isogeny between the pair of elliptic curves in the generic fiber, we establish an upper bound for the number of special fibers with height at most $B$ where the two factors are isogenous. Our proof provides an upper bound that is dependent on $K$, the family, and the bound of height $B$. Furthermore, by introducing a slight modification to the definition of the height of the parametrizing family, we prove a uniform bound depends solely on the degree of the family, the field $K$, and $B$. Based on the uniformity, and the fact that the idea of using Heath-Brown type bounds on covers and optimizing the cover to count rational points on specific algebraic families has not been exploited much yet, we hope that the paper serves as a good example to illustrate the strengths of the method and will inspire further exploration and application of these techniques in related research.

math.NT↗

arXiv2025

Isogeny classes of non-simple abelian surfaces over finite fields

Let $A=E \times E_{ss}$ be a principally polarized almost ordinary split abelian surface over a finite field $\mathbb{F}_{q}$. We give asymptotic upper and lower bounds on the number of principally polarized abelian surfaces over $\mathbb{F}_{q^{n}}$ that are $\overline{\mathbb{F}}_{q}$-isogenous to $A$ up to isomorphism, which is a refinement of the results in the work of Achter and Howe.

math.NT↗

arXiv2025

Non-simple abelian varieties in a family: arithmetic approaches

Inspired by the work of Ellenberg, Elsholtz, Hall, and Kowalski, we investigate how the property of the generic fiber of a one-parameter family of abelian varieties being geometrically simple extends to other fibers. In \cite{EEHK09}, the authors studied a special case involving specific one-parameter families of Jacobians of curves using analytic methods. We generalize their results, particularly Theorem B, to all families of abelian varieties with big geometric monodromy, employing an arithmetic approach. Our method applies Heath-Brown-type bounds on certain covers with level structures and optimizes the covers to derive the desired results.

math.NT↗

arXiv2025

SSRL: Self-Search Reinforcement Learning

We investigate the potential of large language models (LLMs) to serve as efficient simulators for agentic search tasks in reinforcement learning (RL), thereby reducing dependence on costly interactions with external search engines. To this end, we first quantify the intrinsic search capability of LLMs via structured prompting and repeated sampling, which we term Self-Search. Our results reveal that LLMs exhibit strong scaling behavior with respect to the inference budget, achieving high pass@k on question-answering benchmarks, including the challenging BrowseComp task. Building on these observations, we introduce Self-Search RL (SSRL), which enhances LLMs' Self-Search capability through format-based and rule-based rewards. SSRL enables models to iteratively refine their knowledge utilization internally, without requiring access to external tools. Empirical evaluations demonstrate that SSRL-trained policy models provide a cost-effective and stable environment for search-driven RL training, reducing reliance on external search engines and facilitating robust sim-to-real transfer. We draw the following conclusions: 1) LLMs possess world knowledge that can be effectively elicited to achieve high performance; 2) SSRL demonstrates the potential of leveraging internal knowledge to reduce hallucination; 3) SSRL-trained models integrate seamlessly with external search engines without additional effort. Our findings highlight the potential of LLMs to support more scalable RL agent training.

cs.CL↗