Searcharxiv⌕ Search

arXiv subjects

Hao Liang

Publications and source records attributed to Hao Liang.

At least 73 records · Page 4Linked to original sources
arXiv2026

Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions

Recent advances in large language model (LLM) reasoning have shown that sophisticated behaviors such as planning and self-reflection can emerge through reinforcement learning (RL). However, despite these successes, RL in its current form remains insufficient to induce capabilities that exceed the limitations of the base model, as it is primarily optimized based on existing knowledge of the model rather than facilitating the acquisition of new information. To address this limitation, we employ supervised fine-tuning (SFT) to learn what RL cannot, which enables the incorporation of new knowledge and reasoning patterns by leveraging high-quality demonstration data. We analyze the training dynamics of RL and SFT for LLM reasoning and find that RL excels at maintaining and improving performance on questions within the model's original capabilities, while SFT is more effective at enabling progress on questions beyond the current scope of the model. Motivated by the complementary strengths of RL and SFT, we introduce a novel training approach, \textbf{ReLIFT} (\textbf{Re}inforcement \textbf{L}earning \textbf{I}nterleaved with Online \textbf{F}ine-\textbf{T}uning). In ReLIFT, the model is primarily trained using RL, but when it encounters challenging questions, high-quality solutions are collected for fine-tuning, and the training process alternates between RL and fine-tuning to enhance the model's reasoning abilities. ReLIFT achieves an average improvement of over +5.2 points across five competition-level benchmarks and one out-of-distribution benchmark compared to other zero-RL models. Furthermore, we demonstrate that ReLIFT outperforms both RL and SFT while using only 13\% of the detailed demonstration data, highlighting its scalability. These results provide compelling evidence that ReLIFT overcomes the fundamental limitations of RL and underscores the significant potential.

cs.AI↗
arXiv2026

One-Eval: An Agentic System for Automated and Traceable LLM Evaluation

Reliable evaluation is essential for developing and deploying large language models, yet in practice it often requires substantial manual effort: practitioners must identify appropriate benchmarks, reproduce heterogeneous evaluation codebases, configure dataset schema mappings, and interpret aggregated metrics. To address these challenges, we present One-Eval, an agentic evaluation system that converts natural-language evaluation requests into executable, traceable, and customizable evaluation workflows. One-Eval integrates (i) NL2Bench for intent structuring and personalized benchmark planning, (ii) BenchResolve for benchmark resolution, automatic dataset acquisition, and schema normalization to ensure executability, and (iii) Metrics \& Reporting for task-aware metric selection and decision-oriented reporting beyond scalar scores. The system further incorporates human-in-the-loop checkpoints for review, editing, and rollback, while preserving sample evidence trails for debugging and auditability. Experiments show that One-Eval can execute end-to-end evaluations from diverse natural-language requests with minimal user effort, supporting more efficient and reproducible evaluation in industrial settings. Our framework is publicly available at https://github.com/OpenDCAI/One-Eval.

cs.CL↗
arXiv2026

ROMA: a Read-Only-Memory-based Accelerator for QLoRA-based On-Device LLM

As large language models (LLMs) demonstrate powerful capabilities, deploying them on edge devices has become increasingly crucial, offering advantages in privacy and real-time interaction. QLoRA has emerged as the standard approach for on-device LLMs, leveraging quantized models to reduce memory and computational costs while utilizing LoRA for task-specific adaptability. In this work, we propose ROMA, a QLoRA accelerator with a hybrid storage architecture that uses ROM for quantized base models and SRAM for LoRA weights and KV cache. Our insight is that the quantized base model is stable and converged, making it well-suited for ROM storage. Meanwhile, LoRA modules offer the flexibility to adapt to new data without requiring updates to the base model. To further reduce the area cost of ROM, we introduce a novel B-ROM design and integrate it with the compute unit to form a fused cell for efficient use of chip resources. ROMA can effectively store both a 4-bit 3B and a 2-bit 8B LLaMA model entirely on-chip, achieving a notable generation speed exceeding 20,000 tokens/s without requiring external memory.

cs.AR↗
arXiv2026

BrowseComp-$V^3$: A Visual, Vertical, and Verifiable Benchmark for Multimodal Browsing Agents

Multimodal large language models (MLLMs), equipped with increasingly advanced planning and tool-use capabilities, are evolving into autonomous agents capable of performing multimodal web browsing and deep search in open-world environments. However, existing benchmarks for multimodal browsing remain limited in task complexity, evidence accessibility, and evaluation granularity, hindering comprehensive and reproducible assessments of deep search capabilities. To address these limitations, we introduce BrowseComp-$V^3$, a novel benchmark consisting of 300 carefully curated and challenging questions spanning diverse domains. The benchmark emphasizes deep, multi-level, and cross-modal multi-hop reasoning, where critical evidence is interleaved across textual and visual modalities within and across web pages. All supporting evidence is strictly required to be publicly searchable, ensuring fairness and reproducibility. Beyond final-answer accuracy, we incorporate an expert-validated, subgoal-driven process evaluation mechanism that enables fine-grained analysis of intermediate reasoning behaviors and systematic characterization of capability boundaries. In addition, we propose OmniSeeker, a unified multimodal browsing agent framework integrating diverse web search and visual perception tools. Comprehensive experiments demonstrate that even state-of-the-art models achieve only 36% accuracy on our benchmark, revealing critical bottlenecks in multimodal information integration and fine-grained perception. Our results highlight a fundamental gap between current model capabilities and robust multimodal deep search in real-world settings.

cs.AI↗
arXiv2026

MathScape: Benchmarking Multimodal Large Language Models in Real-World Mathematical Contexts

With the rapid progress of Multimodal LLMs, evaluating their mathematical reasoning capabilities has become an increasingly important research direction. In particular, visual-textual mathematical reasoning serves as a key indicator of an MLLM's ability to comprehend and solve complex, multi-step quantitative problems. While existing benchmarks such as MathVista and MathVerse have advanced the evaluation of multimodal math proficiency, they primarily rely on digitally rendered content and fall short in capturing the complexity of real-world scenarios. To bridge this gap, we introduce MathScape, a novel benchmark focused on assessing MLLMs' reasoning ability in realistic mathematical contexts. MathScape comprises 1,369 high-quality math problems paired with human-captured real-world images, closely reflecting the challenges encountered in practical educational settings. We conduct a thorough multi-dimensional evaluation across nine leading closed-source MLLMs, three open-source MLLMs with over 20 billion parameters, and seven smaller-scale MLLMs. Our results show that even state-of-the-art models struggle with real-world math tasks, lagging behind human performance, highlighting critical limitations in current model capabilities. Moreover, we find that strong performance on synthetic or digitally rendered images does not guarantee similar effectiveness on real-world tasks. This underscores the necessity of MathScape in the next stage of multimodal mathematical reasoning.

cs.CV↗
arXiv2026

MC-LLaVA: Multi-Concept Personalized Vision-Language Model

Current vision-language models (VLMs) show exceptional abilities across diverse tasks, such as visual question answering. To enhance user experience, recent studies have investigated VLM personalization to understand user-provided concepts. However, they mainly focus on single concepts, neglecting the existence and interplay of multiple concepts, which limits real-world applicability. This paper proposes MC-LLaVA, a multi-concept personalization paradigm. Specifically, MC-LLaVA employs a multi-concept instruction tuning strategy, effectively integrating multiple concepts in a single training step. To reduce the training costs, we propose a personalized textual prompt that uses visual token information to initialize concept tokens. Additionally, we introduce a personalized visual prompt during inference, aggregating location maps for enhanced recognition and grounding capabilities. To further push the performance upper bound, we incorporate an optional auxiliary loss, better enhancing the proposed personalized prompts. To decorate the VLM personalization research, we contribute a high-quality dataset. We carefully collect images with multiple characters and objects from movies and manually create question-answer samples for multi-concept scenarios, featuring superior diversity. Comprehensive experiments demonstrate that MC-LLaVA achieves impressive multi-concept personalized responses, paving the way for VLMs to become better user assistants. The code and dataset will be released at \href{https://github.com/arctanxarc/MC-LLaVA}{https://github.com/arctanxarc/MC-LLaVA}.

cs.CV↗
arXiv2026

Canvas-of-Thought: Grounding Reasoning via Mutable Structured States

While Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), relying solely on linear text sequences remains a bottleneck for complex tasks. We observe that even when auxiliary visual elements are interleaved, they are often treated as static snapshots within a one-dimensional, unstructured reasoning chain. We argue that such approaches treat reasoning history as an immutable stream: correcting a local error necessitates either generating verbose downstream corrections or regenerating the entire context. This forces the model to implicitly maintain and track state updates, significantly increasing token consumption and cognitive load. This limitation is particularly acute in high-dimensional domains, such as geometry and SVG design, where the textual expression of CoT lacks explicit visual guidance, further constraining the model's reasoning precision. To bridge this gap, we introduce \textbf{Canvas-of-Thought (Canvas-CoT)}. By leveraging a HTML Canvas as an external reasoning substrate, Canvas-CoT empowers the model to perform atomic, DOM-based CRUD operations. This architecture enables in-place state revisions without disrupting the surrounding context, allowing the model to explicitly maintain the "ground truth". Furthermore, we integrate a rendering-based critique loop that serves as a hard constraint validator, providing explicit visual feedback to resolve complex tasks that are difficult to articulate through text alone. Extensive experiments on VCode, RBench-V, and MathVista demonstrate that Canvas-CoT significantly outperforms existing baselines, establishing a new paradigm for context-efficient multimodal reasoning.

cs.CL↗
arXiv2026

Text2SQL-Flow: A Robust SQL-Aware Data Augmentation Framework for Text-to-SQL

The data-centric paradigm has emerged as a pivotal direction in artificial intelligence (AI), emphasizing the role of high-quality training data. This shift is especially critical in the Text-to-SQL task, where the scarcity, limited diversity, and structural simplicity of existing datasets constrain model performance. To address these challenges, we propose Text2SQL-Flow, a SQL-aware data augmentation framework that systematically generates large-scale, semantically valid, and structurally diverse Text-to-SQL pairs from limited seed data. Our framework spans six augmentation dimensions and integrates an end-to-end pipeline with auxiliary database selection, SQL executability verification, natural language (NL) question generation, NL-SQL correspondence verification, and chain-of-thought (CoT) reasoning trace generation. Leveraging this framework, we construct SQLFlow, a high-quality dataset comprising 75,386 annotated examples. We demonstrate the utility of SQLFlow in both fine-tuning and prompt-based settings. (1) For open-source large language models (LLMs), fine-tuning with SQLFlow improves problem-solving ability, delivering competitive gains across multiple benchmarks under the same data budget. (2) For closed-source LLMs, we propose a masked alignment retrieval method that uses SQLFlow as both a knowledge base and training data for the retrieval model, enabling structure-aware example matching via fine-grained NL-SQL alignments. Experiments show that our retrieval strategy outperforms existing example retrieval methods, highlighting the combined value of SQLFlow's data quality and our retrieval technique. Overall, our work provides a scalable, data-centric foundation for advancing Text-to-SQL systems and underscores the importance of structured, high-fidelity data in modern AI development. Our code is available at https://github.com/TechNomad-ds/Text2SQL-Flow.

cs.CL↗
arXiv2026

An Algorithm for Diagonalizing Matrices of Formal Power Series

This paper studies the unitary diagonalization of matrices over formal power series rings. Our main result shows that a normal matrix is unitarily diagonalizable if and only if its minimal polynomial completely splits over the ring and the associated spectral projections have entries in the ring. Building on this characterization, we develop an algorithm for deciding the unitary diagonalizability of matrices over regular local rings of algebraic varieties. A central ingredient of the algorithm is a decision procedure for determining whether a polynomial splits over a formal power series ring; we establish this using techniques from prime decomposition and the relative smoothness of integral closures in ramification theory.

math.AC↗
arXiv2026

M2A: Multimodal Memory Agent with Dual-Layer Hybrid Memory for Long-Term Personalized Interactions

This work addresses the challenge of personalized question answering in long-term human-machine interactions: when conversational history spans weeks or months and exceeds the context window, existing personalization mechanisms struggle to continuously absorb and leverage users' incremental concepts, aliases, and preferences. Current personalized multimodal models are predominantly static-concepts are fixed at initialization and cannot evolve during interactions. We propose M2A, an agentic dual-layer hybrid memory system that maintains personalized multimodal information through online updates. The system employs two collaborative agents: ChatAgent manages user interactions and autonomously decides when to query or update memory, while MemoryManager breaks down memory requests from ChatAgent into detailed operations on the dual-layer memory bank, which couples a RawMessageStore (immutable conversation log) with a SemanticMemoryStore (high-level observations), providing memories at different granularities. In addition, we develop a reusable data synthesis pipeline that injects concept-grounded sessions from Yo'LLaVA and MC-LLaVA into LoCoMo long conversations while preserving temporal coherence. Experiments show that M2A significantly outperforms baselines, demonstrating that transforming personalization from one-shot configuration to a co-evolving memory mechanism provides a viable path for high-quality individualized responses in long-term multimodal interactions. The code is available at https://github.com/Little-Fridge/M2A.

cs.AI↗
arXiv2026

LoVR: A Benchmark for Long Video Retrieval in Multimodal Contexts

Long videos contain a vast amount of information, making video-text retrieval an essential and challenging task in multimodal learning. However, existing benchmarks suffer from limited video duration, low-quality captions, and coarse annotation granularity, which hinder the evaluation of advanced video-text retrieval methods. To address these limitations, we introduce LoVR, a benchmark specifically designed for long video-text retrieval. LoVR contains 467 long videos and over 40,804 fine-grained clips with high-quality captions. To overcome the issue of poor machine-generated annotations, we propose an efficient caption generation framework that integrates VLM automatic generation, caption quality scoring, and dynamic refinement. This pipeline improves annotation accuracy while maintaining scalability. Furthermore, we introduce a semantic fusion method to generate coherent full-video captions without losing important contextual information. Our benchmark introduces longer videos, more detailed captions, and a larger-scale dataset, presenting new challenges for video understanding and retrieval. Extensive experiments on various advanced embedding models demonstrate that LoVR is a challenging benchmark, revealing the limitations of current approaches and providing valuable insights for future research. We release the code and dataset link at https://lovrbench.github.io/

cs.CV↗
arXiv2026

Research on World Models Is Not Merely Injecting World Knowledge into Specific Tasks

World models have emerged as a critical frontier in AI research, aiming to enhance large models by infusing them with physical dynamics and world knowledge. The core objective is to enable agents to understand, predict, and interact with complex environments. However, current research landscape remains fragmented, with approaches predominantly focused on injecting world knowledge into isolated tasks, such as visual prediction, 3D estimation, or symbol grounding, rather than establishing a unified definition or framework. While these task-specific integrations yield performance gains, they often lack the systematic coherence required for holistic world understanding. In this paper, we analyze the limitations of such fragmented approaches and propose a unified design specification for world models. We suggest that a robust world model should not be a loose collection of capabilities but a normative framework that integrally incorporates interaction, perception, symbolic reasoning, and spatial representation. This work aims to provide a structured perspective to guide future research toward more general, robust, and principled models of the world.

cs.CV↗
arXiv2026

Is Pure Exploitation Sufficient in Exogenous MDPs with Linear Function Approximation?

Exogenous MDPs (Exo-MDPs) capture sequential decision-making where uncertainty comes solely from exogenous inputs that evolve independently of the learner's actions. This structure is especially common in operations research applications such as inventory control, energy storage, and resource allocation, where exogenous randomness (e.g., demand, arrivals, or prices) drives system behavior. Despite decades of empirical evidence that greedy, exploitation-only methods work remarkably well in these settings, theory has lagged behind: all existing regret guarantees for Exo-MDPs rely on explicit exploration or tabular assumptions. We show that exploration is unnecessary. We propose Pure Exploitation Learning (PEL) and prove the first general finite-sample regret bounds for exploitation-only algorithms in Exo-MDPs. In the tabular case, PEL achieves $\widetilde{O}(H^2|Ξ|\sqrt{K})$. For large, continuous endogenous state spaces, we introduce LSVI-PE, a simple linear-approximation method whose regret is polynomial in the feature dimension, exogenous state space, and horizon, independent of the endogenous state and action spaces. Our analysis introduces two new tools: counterfactual trajectories and Bellman-closed feature transport, which together allow greedy policies to have accurate value estimates without optimism. Experiments on synthetic and resource-management tasks show that PEL consistently outperforming baselines. Overall, our results overturn the conventional wisdom that exploration is required, demonstrating that in Exo-MDPs, pure exploitation is enough.

cs.LG↗
arXiv2026

An Improved Privacy and Utility Analysis of Differentially Private SGD with Bounded Domain and Smooth Losses

Differentially Private Stochastic Gradient Descent (DPSGD) is widely used to protect sensitive data during the training of machine learning models, but its privacy guarantee often comes at a large cost of model performance due to the lack of tight theoretical bounds quantifying privacy loss. While recent efforts have achieved more accurate privacy guarantees, they still impose some assumptions prohibited from practical applications, such as convexity and complex parameter requirements, and rarely investigate in-depth the impact of privacy mechanisms on the model's utility. In this paper, we provide a rigorous privacy characterization for DPSGD with general L-smooth and non-convex loss functions, revealing converged privacy loss with iteration in bounded-domain cases. Specifically, we track the privacy loss over multiple iterations, leveraging the noisy smooth-reduction property, and further establish comprehensive convergence analysis in different scenarios. In particular, we show that for DPSGD with a bounded domain, (i) the privacy loss can still converge without the convexity assumption, (ii) a smaller bounded diameter can improve both privacy and utility simultaneously under certain conditions, and (iii) the attainable big-O order of the privacy utility trade-off for DPSGD with gradient clipping (DPSGD-GC) and for DPSGD-GC with bounded domain (DPSGD-DC) and mu-strongly convex population risk function, respectively. Experiments via membership inference attack (MIA) in a practical setting validate insights gained from the theoretical results.

cs.LG↗
arXiv2026

MathMixup: Boosting LLM Mathematical Reasoning with Difficulty-Controllable Data Synthesis and Curriculum Learning

In mathematical reasoning tasks, the advancement of Large Language Models (LLMs) relies heavily on high-quality training data with clearly defined and well-graded difficulty levels. However, existing data synthesis methods often suffer from limited diversity and lack precise control over problem difficulty, making them insufficient for supporting efficient training paradigms such as curriculum learning. To address these challenges, we propose MathMixup, a novel data synthesis paradigm that systematically generates high-quality, difficulty-controllable mathematical reasoning problems through hybrid and decomposed strategies. Automated self-checking and manual screening are incorporated to ensure semantic clarity and a well-structured difficulty gradient in the synthesized data. Building on this, we construct the MathMixupQA dataset and design a curriculum learning strategy that leverages these graded problems, supporting flexible integration with other datasets. Experimental results show that MathMixup and its curriculum learning strategy significantly enhance the mathematical reasoning performance of LLMs. Fine-tuned Qwen2.5-7B achieves an average score of 52.6\% across seven mathematical benchmarks, surpassing previous state-of-the-art methods. These results fully validate the effectiveness and broad applicability of MathMixup in improving the mathematical reasoning abilities of LLMs and advancing data-centric curriculum learning.

cs.LG↗
arXiv2025

Prospect for measurement of CP-violating parameters of $B_s^0 \to ϕγ$ at the Tera Z factory

$b \to sγ$ transition is a critical flavor-changing neutral current (FCNC) process that could be used to probe CP violation (CPV) and new physics (NP). We quantify the anticipated precision for measuring $B_s^0 \to ϕγ$ at the CEPC Z pole operation, showing that the relative statistical uncertainty could be as low as 0.16\%, improved by approximately two orders of magnitude compared to existing measurements. Additionally, we perform a time-dependent analysis of the $B_s^0 \to ϕγ$ decay, accounting for $B_s^0/\bar{B}_s^0$ mixing extract the mixing-induced and CP-violating parameters $\boldsymbol{\mathcal{A}_{ϕγ}^Δ}$, $\boldsymbol{C_{ϕγ}}$ and $\boldsymbol{S_{ϕγ}}$. Using central value from LHCb measurement as input, we evaluate the anticipated accuracy of measurements of these parameters. The projected statistical uncertainties are $σ_{A_{ϕγ}^Δ{}^{\text{stat}}} = 0.021$, $σ_C^{\text{stat}} = 0.0092$ and $σ_S^{\text{stat}} = 0.0096$, and the systematic uncertainties are $σ_{A_{ϕγ}^Δ{}^{\text{syst}}} = 0.035$, $σ_C^{\text{syst}} = 0.0027$ and $σ_S^{\text{syst}} = 0.0064$. Furthermore, the 1$σ$ sensitivity boundaries for NP in this study are found to be $\mathcal{A}_{ϕγ}^Δ< -0.05$ or $\mathcal{A}_{ϕγ}^Δ> 0.15$, $\mathcal{C}_{ϕγ} < -0.02$ or $\mathcal{C}_{ϕγ} > 0.04$, and $\mathcal{S}_{ϕγ} < -0.04$ or $\mathcal{S}_{ϕγ} > 0.04$. We also conduct a relevant detector optimization study by establishing the correlation between the anticipated precision and the intrinsic resolution of the ECAL, as well as the performance of the PID system.

hep-ex↗
arXiv2025

Time-domain measurement of Auger electron dynamics in xenon atoms after giant resonant photoionization

Time-resolved measurement of Auger-Meitner (AM) decay [Nature 419, 803 (2002)] marked a milestone in the development of attosecond science. To date, the time constants for the AM decay processes obtained from the time-domain experiments were found to be consistent with the values deduced from conventional energy-domain measurements. One of the main factors limiting the temporal resolution of these studies is the unlocked carrier-envelope-phase (CEP) of the laser pulses used to probe the electronic dynamics triggered by inner-shell photoabsorption. In this work, we report time-resolved inner-shell electron spectroscopy of xenon and krypton using attosecond soft X-ray (atto-SXR) pulses centered at 130 eV in combination with CEP-stabilized few-cycle Yb laser pulses. We observed that the N$_{4,5}$OO Auger electrons from xenon exhibit a clear streaking pattern, but with an unexpected time shift of $\sim$ 1.32 fs relative to the 4$d$ photoelectrons. Furthermore, the energy-integrated yield of streaked Auger electrons from xenon exhibits a pronounced minimum at a pump-probe time delay of 4 fs. Neither of these observations can be explained by current streaking theories and both are inconsistent with lifetimes inferred from energy-domain measurements. The M$_{4,5}$NN Auger electrons from krypton partly overlap in energy with the 3$d$ inner-shell photoelectrons and do not show these anomalous features. This study offers new insights into the inner-shell electron dynamics of heavy atoms in the giant dipole resonance region, laying the groundwork for attosecond soft X-ray spectroscopy of molecular systems containing iodine or bromine atoms.

physics.atom-ph↗
arXiv2025

DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI

The rapidly growing demand for high-quality data in Large Language Models (LLMs) has intensified the need for scalable, reliable, and semantically rich data preparation pipelines. However, current practices remain dominated by ad-hoc scripts and loosely specified workflows, which lack principled abstractions, hinder reproducibility, and offer limited support for model-in-the-loop data generation. To address these challenges, we present DataFlow, a unified and extensible LLM-driven data preparation framework. DataFlow is designed with system-level abstractions that enable modular, reusable, and composable data transformations, and provides a PyTorch-style pipeline construction API for building debuggable and optimizable dataflows. The framework consists of nearly 200 reusable operators and six domain-general pipelines spanning text, mathematical reasoning, code, Text-to-SQL, agentic RAG, and large-scale knowledge extraction. To further improve usability, we introduce DataFlow-Agent, which automatically translates natural-language specifications into executable pipelines via operator synthesis, pipeline planning, and iterative verification. Across six representative use cases, DataFlow consistently improves downstream LLM performance. Our math, code, and text pipelines outperform curated human datasets and specialized synthetic baselines, achieving up to +3\% execution accuracy in Text-to-SQL over SynSQL, +7\% average improvements on code benchmarks, and 1--3 point gains on MATH, GSM8K, and AIME. Moreover, a unified 10K-sample dataset produced by DataFlow enables base models to surpass counterparts trained on 1M Infinity-Instruct data. These results demonstrate that DataFlow provides a practical and high-performance substrate for reliable, reproducible, and scalable LLM data preparation, and establishes a system-level foundation for future data-centric AI development.

cs.LG↗