arXiv subjects

Jian Zhang

Publications and source records attributed to Jian Zhang.

At least 163 records · Page 9Linked to original sources

arXiv2026

LLMdoctor: Token-Level Flow-Guided Preference Optimization for Efficient Test-Time Alignment of Large Language Models

Aligning Large Language Models (LLMs) with human preferences is critical, yet traditional fine-tuning methods are computationally expensive and inflexible. While test-time alignment offers a promising alternative, existing approaches often rely on distorted trajectory-level signals or inefficient sampling, fundamentally capping performance and failing to preserve the generative diversity of the base model. This paper introduces LLMdoctor, a novel framework for efficient test-time alignment that operates via a patient-doctor paradigm. It integrates token-level reward acquisition with token-level flow-guided preference optimization (TFPO) to steer a large, frozen patient LLM with a smaller, specialized doctor model. Unlike conventional methods that rely on trajectory-level rewards, LLMdoctor first extracts fine-grained, token-level preference signals from the patient model's behavioral variations. These signals then guide the training of the doctor model via TFPO, which establishes flow consistency across all subtrajectories, enabling precise token-by-token alignment while inherently preserving generation diversity. Extensive experiments demonstrate that LLMdoctor significantly outperforms existing test-time alignment methods and even surpasses the performance of full fine-tuning approaches like DPO.

cs.AI↗

arXiv2026

MAXS: Meta-Adaptive Exploration with LLM Agents

Large Language Model (LLM) Agents exhibit inherent reasoning abilities through the collaboration of multiple tools. However, during agent inference, existing methods often suffer from (i) locally myopic generation, due to the absence of lookahead, and (ii) trajectory instability, where minor early errors can escalate into divergent reasoning paths. These issues make it difficult to balance global effectiveness and computational efficiency. To address these two issues, we propose meta-adaptive exploration with LLM agents https://github.com/exoskeletonzj/MAXS, a meta-adaptive reasoning framework based on LLM Agents that flexibly integrates tool execution and reasoning planning. MAXS employs a lookahead strategy to extend reasoning paths a few steps ahead, estimating the advantage value of tool usage, and combines step consistency variance and inter-step trend slopes to jointly select stable, consistent, and high-value reasoning steps. Additionally, we introduce a trajectory convergence mechanism that controls computational cost by halting further rollouts once path consistency is achieved, enabling a balance between resource efficiency and global effectiveness in multi-tool reasoning. We conduct extensive empirical studies across three base models (MiMo-VL-7B, Qwen2.5-VL-7B, Qwen2.5-VL-32B) and five datasets, demonstrating that MAXS consistently outperforms existing methods in both performance and inference efficiency. Further analysis confirms the effectiveness of our lookahead strategy and tool usage.

cs.AI↗

arXiv2026

$A^3$-Bench: Benchmarking Memory-Driven Scientific Reasoning via Anchor and Attractor Activation

Scientific reasoning relies not only on logical inference but also on activating prior knowledge and experiential structures. Memory can efficiently reuse knowledge and enhance reasoning consistency and stability. However, existing benchmarks mainly evaluate final answers or step-by-step coherence, overlooking the \textit{memory-driven} mechanisms that underlie human reasoning, which involves activating anchors and attractors, then integrating them into multi-step inference. To address this gap, we propose $A^3$-Bench~ https://a3-bench.github.io, a benchmark designed to evaluate scientific reasoning through dual-scale memory-driven activation, grounded in Anchor and Attractor Activation. First, we annotate 2,198 science reasoning problems across domains using the SAPM process(subject, anchor & attractor, problem, and memory developing). Second, we introduce a dual-scale memory evaluation framework utilizing anchors and attractors, along with the AAUI(Anchor--Attractor Utilization Index) metric to measure memory activation rates. Finally, through experiments with various base models and paradigms, we validate $A^3$-Bench and analyze how memory activation impacts reasoning performance, providing insights into memory-driven scientific reasoning.

cs.AI↗

arXiv2026

MemFine: Memory-Aware Fine-Grained Scheduling for MoE Training

The training of large-scale Mixture of Experts (MoE) models faces a critical memory bottleneck due to severe load imbalance caused by dynamic token routing. This imbalance leads to memory overflow on GPUs with limited capacity, constraining model scalability. Existing load balancing methods, which cap expert capacity, compromise model accuracy and fail on memory-constrained hardware. To address this, we propose MemFine, a memory-aware fine-grained scheduling framework for MoE training. MemFine decomposes the token distribution and expert computation into manageable chunks and employs a chunked recomputation strategy, dynamically optimized through a theoretical memory model to balance memory efficiency and throughput. Experiments demonstrate that MemFine reduces activation memory by 48.03% and improves throughput by 4.42% compared to full recomputation-based baselines, enabling stable large-scale MoE training on memory-limited GPUs.

cs.DC↗

arXiv2026

DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models

Recent advancements in slow thinking reasoning models have shown exceptional performance in complex reasoning tasks. However, these models often exhibit overthinking (generating redundant reasoning steps for simple problems), leading to excessive computational resource usage. While current mitigation strategies uniformly reduce reasoning tokens, they risk degrading performance on challenging tasks that require extended reasoning. This paper introduces Difficulty-Adaptive Slow Thinking (DAST), a novel framework that enables models to autonomously adjust the length of Chain-of-Thought (CoT) based on problem difficulty. We first propose a Token Length Budget (TLB) metric to quantify difficulty, then leverage budget-aware reward shaping and budget preference optimization to implement DAST. DAST penalizes overlong responses for simple tasks while incentivizing sufficient reasoning for complex problems. Experiments on diverse datasets and model scales demonstrate that DAST effectively mitigates overthinking (reducing token usage by over 30\% on average) while preserving reasoning accuracy on complex problems. Our codes and models are available at https://github.com/AnonymousUser0520/AnonymousRepo01.

cs.LG↗

arXiv2026

Central elements and evaluation map for the quantum queer superalgebras

We consider the $R$-matrix presentations of the quantum queer superalgebra $U_q(q_n)$ and its affine counterpart $U_q(\widehat q_n)$. We derive crossing symmetry relations for the $R$-matrices and use them to construct central elements in both superalgebras. We also produce an epimorphism $ev:U_q(\widehat q_n)\to U_q(q_n)$ identical on the subalgebra isomorphic to $U_q(q_n)$.

math.QA↗

arXiv2026

LLM-Guided Quantified SMT Solving over Uninterpreted Functions

Quantified formulas with Uninterpreted Functions (UFs) over non-linear real arithmetic pose fundamental challenges for Satisfiability Modulo Theories (SMT) solving. Traditional quantifier instantiation methods struggle because they lack semantic understanding of UF constraints, forcing them to search through unbounded solution spaces with limited guidance. We present AquaForte, a framework that leverages Large Language Models to provide semantic guidance for UF instantiation by generating instantiated candidates for function definitions that satisfy the constraints, thereby significantly reducing the search space and complexity for solvers. Our approach preprocesses formulas through constraint separation, uses structured prompts to extract mathematical reasoning from LLMs, and integrates the results with traditional SMT algorithms through adaptive instantiation. AquaForte maintains soundness through systematic validation: LLM-guided instantiations yielding SAT solve the original problem, while UNSAT results generate exclusion clauses for iterative refinement. Completeness is preserved by fallback to traditional solvers augmented with learned constraints. Experimental evaluation on SMT-COMP benchmarks demonstrates that AquaForte solves numerous instances where state-of-the-art solvers like Z3 and CVC5 timeout, with particular effectiveness on satisfiable formulas. Our work shows that LLMs can provide valuable mathematical intuition for symbolic reasoning, establishing a new paradigm for SMT constraint solving.

cs.AI↗

arXiv2026

Validating Generalist Robots with Situation Calculus and STL Falsification

Generalist robots are becoming a reality, capable of interpreting natural language instructions and executing diverse operations. However, their validation remains challenging because each task induces its own operational context and correctness specification, exceeding the assumptions of traditional validation methods. We propose a two-layer validation framework that combines abstract reasoning with concrete system falsification. At the abstract layer, situation calculus models the world and derives weakest preconditions, enabling constraint-aware combinatorial testing to systematically generate diverse, semantically valid world-task configurations with controllable coverage strength. At the concrete layer, these configurations are instantiated for simulation-based falsification with STL monitoring. Experiments on tabletop manipulation tasks show that our framework effectively uncovers failure cases in the NVIDIA GR00T controller, demonstrating its promise for validating general-purpose robot autonomy.

cs.RO↗

arXiv2026

TalkPhoto: A Versatile Training-Free Conversational Assistant for Intelligent Image Editing

Thanks to the powerful language comprehension capabilities of Large Language Models (LLMs), existing instruction-based image editing methods have introduced Multimodal Large Language Models (MLLMs) to promote information exchange between instructions and images, ensuring the controllability and flexibility of image editing. However, these frameworks often build a multi-instruction dataset to train the model to handle multiple editing tasks, which is not only time-consuming and labor-intensive but also fails to achieve satisfactory results. In this paper, we present TalkPhoto, a versatile training-free image editing framework that facilitates precise image manipulation through conversational interaction. We instruct the open-source LLM with a specially designed prompt template to analyze user needs after receiving instructions and hierarchically invoke existing advanced editing methods, all without additional training. Moreover, we implement a plug-and-play and efficient invocation of image editing methods, allowing complex and unseen editing tasks to be integrated into the current framework, achieving stable and high-quality editing results. Extensive experiments demonstrate that our method not only provides more accurate invocation with fewer token consumption but also achieves higher editing quality across various image editing tasks.

cs.CV↗

arXiv2026

Leveraging Flatness to Improve Information-Theoretic Generalization Bounds for SGD

Information-theoretic (IT) generalization bounds have been used to study the generalization of learning algorithms. These bounds are intrinsically data- and algorithm-dependent so that one can exploit the properties of data and algorithm to derive tighter bounds. However, we observe that although the flatness bias is crucial for SGD's generalization, these bounds fail to capture the improved generalization under better flatness and are also numerically loose. This is caused by the inadequate leverage of SGD's flatness bias in existing IT bounds. This paper derives a more flatness-leveraging IT bound for the flatness-favoring SGD. The bound indicates the learned models generalize better if the large-variance directions of the final weight covariance have small local curvatures in the loss landscape. Experiments on deep neural networks show our bound not only correctly reflects the better generalization when flatness is improved, but is also numerically much tighter. This is achieved by a flexible technique called "omniscient trajectory". When applied to Gradient Descent's minimax excess risk on convex-Lipschitz-Bounded problems, it improves representative IT bounds' $Ω(1)$ rates to $O(1/\sqrt{n})$. It also implies a by-pass of memorization-generalization trade-offs.

cs.LG↗

arXiv2025

Evaluating Large Language Models for Line-Level Vulnerability Localization

Recently, Automated Vulnerability Localization (AVL) has attracted growing attention, aiming to facilitate diagnosis by pinpointing the specific lines of code responsible for vulnerabilities. Large Language Models (LLMs) have shown potential in various domains, yet their effectiveness in line-level vulnerability localization remains underexplored. In this work, we present the first comprehensive empirical evaluation of LLMs for AVL. Our study examines 19 leading LLMs suitable for code analysis, including ChatGPT and multiple open-source models, spanning encoder-only, encoder-decoder, and decoder-only architectures, with model sizes from 60M to 70B parameters. We evaluate three paradigms including few-shot prompting, discriminative fine-tuning, and generative fine-tuning with and without Low-Rank Adaptation (LoRA), on both a BigVul-derived dataset for C/C++ and a smart contract vulnerability dataset.} Our results show that discriminative fine-tuning achieves substantial performance gains over existing learning-based AVL methods when sufficient training data is available. In low-data settings, prompting advanced LLMs such as ChatGPT proves more effective. We also identify challenges related to input length and unidirectional context during fine-tuning, and propose two remedial strategies: a sliding window approach and right-forward embedding, both of which yield significant improvements. Moreover, we provide the first assessment of LLM generalizability in AVL, showing that certain models can transfer effectively across Common Weakness Enumerations (CWEs) and projects. However, performance degrades notably for newly discovered vulnerabilities containing unfamiliar lexical or structural patterns, underscoring the need for continual adaptation.

cs.SE↗

arXiv2025

Co-Evolution of Types and Dependencies: Towards Repository-Level Type Inference for Python Code

Python's dynamic typing mechanism, while promoting flexibility, is a significant source of runtime type errors that plague large-scale software, which inspires the automatic type inference techniques. Existing type inference tools have achieved advances in type inference within isolated code snippets. However, repository-level type inference remains a significant challenge, primarily due to the complex inter-procedural dependencies that are difficult to model and resolve. To fill this gap, we present \methodName, a novel approach based on LLMs that achieves repository-level type inference through the co-evolution of types and dependencies. \methodName~constructs an Entity Dependency Graph (EDG) to model the objects and type dependencies across the repository. During the inference process, it iteratively refines types and dependencies in EDG for accurate type inference. Our key innovations are: (1) an EDG model designed to capture repository-level type dependencies; (2) an iterative type inference approach where types and dependencies co-evolve in each iteration; and (3) a type-checker-in-the-loop strategy that validates and corrects inferences on-the-fly, thereby reducing error propagation. When evaluated on 12 complex Python repositories, \methodName~significantly outperformed prior works, achieving a \textit{TypeSim} score of 0.89 and a \textit{TypeExact} score of 0.84, representing a 27\% and 40\% relative improvement over the strongest baseline. More importantly, \methodName~removed new type errors introduced by the tool by 92.7\%. This demonstrates a significant leap towards automated, reliable type annotation for real-world Python development.

cs.SE↗

arXiv2025

MAGIC: Achieving Superior Model Merging via Magnitude Calibration

The proliferation of pre-trained models has given rise to a wide array of specialised, fine-tuned models. Model merging aims to merge the distinct capabilities of these specialised models into a unified model, requiring minimal or even no additional training. A core objective of model merging is to ensure the merged model retains the behavioural characteristics of the specialised models, typically achieved through feature alignment. We identify that features consist of two critical components: direction and magnitude. Prior research has predominantly focused on directional alignment, while the influence of magnitude remains largely neglected, despite its pronounced vulnerability to perturbations introduced by common merging operations (e.g., parameter fusion and sparsification). Such perturbations to magnitude inevitably lead to feature deviations in the merged model from the specialised models, resulting in subsequent performance degradation. To address this, we propose MAGnItude Calibration (MAGIC), a plug-and-play framework that rectifies layer-wise magnitudes in feature and weight spaces, with three variants. Specifically, our Feature Space Calibration (FSC) realigns the merged model's features using a small set of unlabelled data, while Weight Space Calibration (WSC) extends this calibration to the weight space without requiring additional data. Combining these yields Dual Space Calibration (DSC). Comprehensive experiments demonstrate that MAGIC consistently boosts performance across diverse Computer Vision tasks (+4.3% on eight datasets) and NLP tasks (+8.0% on Llama) without additional training. Our code is available at: https://github.com/lyymuwu/MAGIC

cs.LG↗

arXiv2025

WeMusic-Agent: Efficient Conversational Music Recommendation via Knowledge Internalization and Agentic Boundary Learning

Personalized music recommendation in conversational scenarios usually requires a deep understanding of user preferences and nuanced musical context, yet existing methods often struggle with balancing specialized domain knowledge and flexible tool integration. This paper proposes WeMusic-Agent, a training framework for efficient LLM-based conversational music recommendation. By integrating the knowledge internalization and agentic boundary learning, the framework aims to teach the model to intelligently decide when to leverage internalized knowledge and when to call specialized tools (e.g., music retrieval APIs, music recommendation systems). Under this framework, we present WeMusic-Agent-M1, an agentic model that internalizes extensive musical knowledge via continued pretraining on 50B music-related corpus while acquiring the ability to invoke external tools when necessary. Additionally, considering the lack of open-source benchmarks for conversational music recommendation, we also construct a benchmark for personalized music recommendations derived from real-world data in WeChat Listen. This benchmark enables comprehensive evaluation across multiple dimensions, including relevance, personalization, and diversity of the recommendations. Experiments on real-world data demonstrate that WeMusic-Agent achieves significant improvements over existing models.

cs.AI↗

arXiv2025

Artificial Intelligence-Enabled Holistic Design of Catalysts Tailored for Semiconducting Carbon Nanotube Growth

Catalyst design is crucial for materials synthesis, especially for complex reaction networks. Strategies like collaborative catalytic systems and multifunctional catalysts are effective but face challenges at the nanoscale. Carbon nanotube synthesis contains complicated nanoscale catalytic reactions, thus achieving high-density, high-quality semiconducting CNTs demands innovative catalyst design. In this work, we present a holistic framework integrating machine learning into traditional catalyst design for semiconducting CNT synthesis. It combines knowledge-based insights with data-driven techniques. Three key components, including open-access electronic structure databases for precise physicochemical descriptors, pre-trained natural language processing-based embedding model for higher-level abstractions, and physical - driven predictive models based on experiment data, are utilized. Through this framework, a new method for selective semiconducting CNT synthesis via catalyst - mediated electron injection, tuned by light during growth, is proposed. 54 candidate catalysts are screened, and three with high potential are identified. High-throughput experiments validate the predictions, with semiconducting selectivity exceeding 91% and the FeTiO3 catalyst reaching 98.6%. This approach not only addresses semiconducting CNT synthesis but also offers a generalizable methodology for global catalyst design and nanomaterials synthesis, advancing materials science in precise control.

cond-mat.mtrl-sci↗

arXiv2025

A new proof on quasilinear Schrödinger equations with prescribed mass and combined nonlinearities

In this work, we study the quasilinear Schrödinger equation \begin{equation*} \aligned -Δu-Δ(u^2)u=|u|^{p-2}u+|u|^{q-2}u+λu,\,\, x\in\R^N, \endaligned \end{equation*} under the mass constraint \begin{equation*} \int_{\R^N}|u|^2\text{d}x=a, \end{equation*} where $N\geq2$, $2 0$ is a given mass and $λ$ is a Lagrange multiplier. As a continuation of our previous work (Chen et al., 2025, arXiv:2506.07346v1), we establish some results by means of a suitable change of variables as follows: \begin{itemize} \item[{\bf(i) }] {\bf qualitative analysis of the constrained minimization}\\ For $2 0$; \end{itemize} \begin{itemize} \item[{\bf(ii)}]{\bf existence of two radial distinct normalized solutions}\\ For $2<p<2+\frac{4}{N}<4+\frac{4}{N}<q<22^*$, we obtain a local minimizer under the normalized constraint;\\ For $2<p<2+\frac{4}{N}<4+\frac{4}{N}<q\leq2^*$, we obtain a mountain pass type normalized solution distinct from the local minimizer. \end{itemize} Notably, the second result {\bf (ii)} resolves the open problem {\bf(OP1)} posed by (Chen et al., 2025, arXiv:2506.07346v1). Unlike previous approaches that rely on constructing Palais-Smale-Pohozaev sequences by [Jeanjean, 1997, Nonlinear Anal. {\bf 28}, 1633-1659], we obtain the mountain pass solution employing a new method, which lean upon the monotonicity trick developed by (Chang et al., 2024, Ann. Inst. H. Poincaré C Anal. Non Linéaire, {\bf 41}, 933-959). We emphasize that the methods developed in this work can be extended to investigate the existence of mountain pass-type normalized solutions for other classes of quasilinear Schrödinger equations.

math.AP↗

arXiv2025

Audio-Visual Cross-Modal Compression for Generative Face Video Coding

Generative face video coding (GFVC) is vital for modern applications like video conferencing, yet existing methods primarily focus on video motion while neglecting the significant bitrate contribution of audio. Despite the well-established correlation between audio and lip movements, this cross-modal coherence has not been systematically exploited for compression. To address this, we propose an Audio-Visual Cross-Modal Compression (AVCC) framework that jointly compresses audio and video streams. Our framework extracts motion information from video and tokenizes audio features, then aligns them through a unified audio-video diffusion process. This allows synchronized reconstruction of both modalities from a shared representation. In extremely low-rate scenarios, AVCC can even reconstruct one modality from the other. Experiments show that AVCC significantly outperforms the Versatile Video Coding (VVC) standard and state-of-the-art GFVC schemes in rate-distortion performance, paving the way for more efficient multimodal communication systems.

eess.IV↗

arXiv2025

UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement

Image quality assessment (IQA) and image restoration are fundamental problems in low-level vision. Although IQA and restoration are closely connected conceptually, most existing work treats them in isolation. Recent advances in unified multimodal understanding-generation models demonstrate promising results and indicate that stronger understanding can improve generative performance. This motivates a single model that unifies IQA and restoration and explicitly studies how IQA can guide restoration, a setting that remains largely underexplored yet highly valuable. In this paper, we propose UARE, to our knowledge the first Unified vision-language model for image quality Assessment, Restoration, and Enhancement. Built on pretrained unified understanding and generation models, we introduce a two-stage training framework. First, a progressive, easy-to-hard schedule expands from single-type distortions to higher-order mixed degradations, enabling UARE to handle multiple degradations. Second, we perform unified fine-tuning of quality understanding and restoration with interleaved text-image data, aligning IQA signals with restoration objectives. Through multi-task co-training, UARE leverages IQA to boost restoration and enhancement performance. Extensive experiments across IQA, restoration, and enhancement tasks demonstrate the effectiveness of UARE. The code and models will be available at https://github.com/lwq20020127/UARE.

cs.CV↗