arXiv subjects

Lu Ma

Publications and source records attributed to Lu Ma.

At least 19 recordsLinked to original sources

arXiv2026

From Passive Retrieval to Active Memory Navigation: Learning to Use Memory as a Structured Action Space

Long-term user memory is essential for personalized conversational agents, yet many memory systems still expose memory through passive retrieval interfaces, making the model a consumer of pre-selected evidence. We introduce NapMem, a framework for learning to use long-term user memory as a structured action space rather than passively retrieved context. NapMem organizes user history into a linked multi-granularity memory pyramid, where raw conversations, typed memory records, topic tracks, and user profiles are connected through provenance relations, and exposes these levels through memory tools. The agent is trained to select memory according to the query and intermediate evidence, allowing it to inspect different memory granularities before answering. Experiments on PersonaMem-v2, LongMemEval, and LoCoMo show that a NapMem agent trained with memory-tool reinforcement learning is competitive across diverse memory-intensive tasks, while evaluations on non-memory tasks suggest that the learned policy largely preserves general reasoning and tool-use abilities. Additional analyses examine storage, inference cost, tool-use behavior, and ablations over navigation, memory granularity, and RL training. Our results suggest that long-term user memory benefits from coupling structured storage with a learned policy for using memory at the appropriate granularity.

cs.AI↗

arXiv2026

ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment

AI agents are increasingly being tasked with automating AI research itself, particularly the critical post-training phase that transforms base LLMs into aligned assistants. However, recent evaluations reveal that even frontier agents struggle to perform this task. While the success of post-training fundamentally relies on acquiring high-quality data, relying on agents to autonomously curate targeted training datasets from the open web introduces severe challenges. Executing the long-horizon tasks of searching, filtering, and balancing data within noisy web environments frequently overwhelms an agent's limited context, ultimately leading to degraded dataset quality and suboptimal downstream training performance. To bridge this gap, we introduce Andes (Agent Native Data Evolving Synthesis), a framework that reimagines data generation as a plug-and-play \emph{agent skill}. Rather than forcing agents to devise complex data-gathering strategies from scratch, \textsc{Andes} provides an intelligent abstraction layer. By leveraging a self-evolving World Tree routing mechanism and actionable diagnostic reports, it allows trainer agents to dynamically steer data synthesis through an interactive, closed-loop interface. We demonstrate that under strict compute constraints, equipping foundationally weaker agents with Andes improves automated alignment, securing state-of-the-art performance on PostTrainBench and robust cross-task generalization. Our project is available at https://github.com/zzy1127/ANDES.

cs.AI↗

arXiv2026

Unified Flux Control Architecture for Fluxonium Qubits

Control architectures that reduce hardware overhead while maintaining high-fidelity operations are essential for the continued scaling of superconducting quantum processors. Here we experimentally realize a unified control architecture for fluxonium qubits, in which both transverse ($XY$) and longitudinal ($Z$) operations are implemented through a single flux-control channel driven by a single arbitrary waveform generator channel. This architecture imposes competing requirements on the shared control channel, which must simultaneously support low-frequency flux transmission for reset operations while strongly attenuating broadband noise near the qubit transition frequency. We address this challenge through frequency-selective cryogenic filtering together with compensated waveform synthesis that corrects the pulse distortion introduced by the filtered control line. Experimentally, this approach preserves coherence times above 100 $\mu$s while enabling active reset with approximately 98% fidelity and 20-ns single-qubit gates with fidelities exceeding 99.99%. We further demonstrate FPGA-native instruction-level waveform synthesis based on reusable pulse primitives for unified flux control. These results establish unified flux control as a scalable architecture for fluxonium qubits that reduces control hardware overhead while preserving high-fidelity operation.

quant-ph↗

arXiv2026

Training with Harnesses: On-Policy Harness Self-Distillation for Complex Reasoning

Inference-time harnesses substantially improve large language models on complex reasoning tasks. However, the intrinsic capabilities of the underlying model remain unchanged by the addition of these external workflows. To bridge this gap, we introduce \emph{On-Policy Harness Self-Distillation} (OPHSD), which employs the harness-augmented current model as a teacher for self-distillation, thereby introducing extra supervisory signals from the harness beyond training data. OPHSD internalizes task-specific harness capabilities into the student model, yielding robust generalizability and strong standalone performance across diverse reasoning tasks. Evaluated across draft--verify harness for text classification and plan--solve for mathematical reasoning tasks, OPHSD consistently outperforms strong baselines (e.g., +10.83\% over OPSD on HMMT25). Our analysis further indicates that reattaching the harness during inference yields no additional benefits and can even degrade performance, suggesting that complex harnesses need not always be permanent fixtures; instead, they can serve as temporary training scaffolds whose benefits are permanently fed back into the base model. Our code and training data are available at https://github.com/zzy1127/OPHSD-On-Policy-Harness-Self-Distillation.

cs.CL↗

arXiv2026

Millikelvin digital-to-analog converter for superconducting quantum processors

Scaling superconducting quantum processors is increasingly constrained by the wiring, heat load, and calibration overhead associated with delivering high-resolution analog signals from room temperature to qubits at millikelvin temperature. Here we demonstrate a superconducting digital-to-analog converter (DAC) integrated with high-coherence fluxonium qubits in a multi-chip module architecture. The DACs generate persistent analog flux signals for tuning qubit parameters and are programmed deterministically using single-flux-quantum (SFQ) pulses, providing a digital interface compatible with established SFQ routing and demultiplexing technologies. Operating at millikelvin temperature, the DACs enable in-situ tuning of fluxonium qubits without measurable degradation of qubit coherence. The presented device provides a static control primitive for flux-tunable qubits, enabling parameter homogenization and eliminating the need for individual room-temperature DC bias lines. These results establish SFQ-programmable millikelvin DACs as a building block for digitally controlled superconducting quantum processors.

quant-ph↗

arXiv2026

Plasmonic Photocatalysis Enables Selective Oxidative Coupling of Methane with Nitrous Oxide under Ambient Conditions

Methane (CH4) and nitrous oxide (N2O) are potent greenhouse gases that represent substantial chemical energy. Conversion of these abundant waste gases to high-value chemicals typically requires high temperatures up to 1000 C, producing substantial CO2 emissions and limited selectivity toward desirable multi-carbon products. Here we demonstrate a plasmonic photocatalyst that enables CH4 and N2O conversion under ambient conditions to form C2 and C3 hydrocarbons. By systematically tuning AuPd alloys on TiO2, we identify an optimal composition (AuPd0.05) where Au enhances light harvesting and Pd enables selective C-H activation and C-C coupling. Under visible-light illumination, this catalyst produces C2H4, C2H6, C3H6, and C3H8 with ~80% selectivity while suppressing CO2 formation. In-situ spectroscopy and hot-carrier calculations show that plasmon-generated carriers redistribute interfacial hydroxyl intermediates, shifting the hydrophilic center to suppress overoxidation. Ab-initio calculations further reveal the reduction in C-C coupling barriers from 2.7 eV to 0.7 eV under illumination. Our work illustrates how engineering interfacial electronic and adsorbate dynamics enables selective multicarbon formation.

cond-mat.mtrl-sci↗

arXiv2026

Fractal Characterization of Low-Correlation Signals in AI-Generated Image Detection

AI-generated imagery has reached near-photorealistic fidelity, yet this technology poses significant threats to information security and societal trust. Existing deepfake detection methods often exhibit limited robustness in open-world scenarios. To address this limitation, this paper investigates intrinsic discrepancies between synthetic and authentic images from a signal-level perspective. Our analysis reveals that low-correlation signals serve as distinctive markers for differentiating AI-generated imagery from real photographs. Building on this insight, we introduce a novel method for quantifying these signals based on fractal theory. By analyzing the fractal characteristics of low-correlation signals, our method effectively captures the subtle statistical anomalies inherent to the synthesis process. Extensive experimental results demonstrate the method's robustness and superior detection performance. This work emphasizes the need to shift research focus to a new signal-level direction for deepfake detection. Theoretically, this proposed approach is not limited to face image identification but can be applied to all AI-generated image detection tasks. This study provides a new research direction for deepfake detection.

cs.CV↗

arXiv2026

Scalable Fluxonium Quantum Processors via Tunable-Coupler Architecture

Superconducting quantum processors have largely converged on transmon-based architectures, while alternative qubit modalities with intrinsic error protection have lacked a demonstrated path to scalable system integration. In particular, although tunable-coupler-mediated interactions have been validated for small fluxonium systems, it remains unclear whether such designs can be scaled to a multi-qubit lattice. Here, we establish a scalable fluxonium processor architecture based on a modular qubit-coupler unit cell engineered to suppress residual interactions and spectator errors in a many-qubit lattice. The system enables parallel single-qubit gate fidelities approaching 99.99% and two-qubit CZ gate fidelities around 99%. With an optimized gate duration of 32 ns, the best CZ gate fidelity reaches 99.9%. We further validate this architecture in a 22-qubit processor based on the same configuration, where parallel operations enable the deterministic generation of Greenberger-Horne-Zeilinger states involving up to 10 qubits. Together, these results demonstrate that the fluxonium-tunable-coupler unit cell composes without emergent interaction pathologies and establish fluxonium as a scalable superconducting qubit platform.

quant-ph↗

arXiv2026

DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models

Data-centric training has emerged as a promising direction for improving large language models (LLMs) by optimizing not only model parameters but also the selection, composition, and weighting of training data during optimization. However, existing approaches to data selection, data mixture optimization, and data reweighting are often developed in isolated codebases with inconsistent interfaces, hindering reproducibility, fair comparison, and practical integration. In this paper, we present DataFlex, a unified data-centric dynamic training framework built upon LLaMA-Factory. DataFlex supports three major paradigms of dynamic data optimization: sample selection, domain mixture adjustment, and sample reweighting, while remaining fully compatible with the original training workflow. It provides extensible trainer abstractions and modular components, enabling a drop-in replacement for standard LLM training, and unifies key model-dependent operations such as embedding extraction, inference, and gradient computation, with support for large-scale settings including DeepSpeed ZeRO-3. We conduct comprehensive experiments across multiple data-centric methods. Dynamic data selection consistently outperforms static full-data training on MMLU across both Mistral-7B and Llama-3.2-3B. For data mixture, DoReMi and ODM improve both MMLU accuracy and corpus-level perplexity over default proportions when pretraining Qwen2.5-1.5B on SlimPajama at 6B and 30B token scales. DataFlex also achieves consistent runtime improvements over original implementations. These results demonstrate that DataFlex provides an effective, efficient, and reproducible infrastructure for data-centric dynamic training of LLMs.

cs.LG↗

arXiv2026

Thinking by Subtraction: Confidence-Driven Contrastive Decoding for LLM Reasoning

Recent work on test-time scaling for large language model (LLM) reasoning typically assumes that allocating more inference-time computation uniformly improves correctness. However, prior studies show that reasoning uncertainty is highly localized: a small subset of low-confidence tokens disproportionately contributes to reasoning errors and unnecessary output expansion. Motivated by this observation, we propose Thinking by Subtraction, a confidence-driven contrastive decoding approach that improves reasoning reliability through targeted token-level intervention. Our method, Confidence-Driven Contrastive Decoding, detects low-confidence tokens during decoding and intervenes selectively at these positions. It constructs a contrastive reference by replacing high-confidence tokens with minimal placeholders, and refines predictions by subtracting this reference distribution at low-confidence locations. Experiments show that CCD significantly improves accuracy across mathematical reasoning benchmarks while substantially reducing output length, with minimal KV-cache overhead. As a training-free method, CCD enhances reasoning reliability through targeted low-confidence intervention without computational redundancy. Our code will be made available at: https://github.com/bolo-web/CCD.

cs.CL↗

arXiv2026

Converting qubit relaxation into erasures with a single fluxonium

Qubits that experience predominantly erasure errors offer distinct advantages for fault-tolerant operation. Indeed, dual-rail encoded erasure qubits in superconducting cavities and transmons have demonstrated high-fidelity operations by converting physical-qubit relaxation into logical-qubit erasures, but this comes at the cost of increased hardware overhead and circuit complexity. Here, we address these limitations by realizing erasure conversion in a single fluxonium operated at zero flux, where the logical state is encoded in its 0-2 subspace. A single, carefully engineered resonator provides both mid-circuit erasure detection and end-of-line (EOL) logical measurement. Post-selection on non-erasure outcomes results in more than four-fold increase of the logical lifetime, from $193~\mu$s to $869~\mu$s. Finally, we characterize measurement-induced logical dephasing as a function of measurement power and frequency, and infer that each erasure check contributes a negligible error of $7.2\times 10^{-5}$. These results establish integer-fluxonium as a promising, resource-efficient platform for erasure-based error mitigation, without requiring additional hardware.

quant-ph↗

arXiv2026

GIFT: Reconciling Post-Training Objectives via Variational Finite-Temperature Gibbs Initialization

The prevailing post-training paradigm for Large Reasoning Models (LRMs)---Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL)-suffers from an intrinsic optimization mismatch: the rigid likelihood maximization in SFT induces distributional collapse, thereby exhausting the exploration space necessary for subsequent RL. Motivated by the Gibbs optimum of KL-regularized RL, we derive a token-level variational surrogate that makes the SFT target structurally compatible with the subsequent RL stage, and propose Gibbs Initialization with Finite Temperature (GIFT). Standard SFT emerges as a degenerate zero-temperature limit of this surrogate, while a finite temperature preserves structural diversity. Our experiments demonstrate that GIFT outperforms standard SFT and surpasses competitive baselines when utilized for RL initialization. Our code is available at https://github.com/zzy1127/GIFT.

cs.LG↗

arXiv2025

Leash: Adaptive Length Penalty and Reward Shaping for Efficient Large Reasoning Model

Existing approaches typically rely on fixed length penalties, but such penalties are hard to tune and fail to adapt to the evolving reasoning abilities of LLMs, leading to suboptimal trade-offs between accuracy and conciseness. To address this challenge, we propose Leash (adaptive LEngth penAlty and reward SHaping), a reinforcement learning framework for efficient reasoning in LLMs. We formulate length control as a constrained optimization problem and employ a Lagrangian primal-dual method to dynamically adjust the penalty coefficient. When generations exceed the target length, the penalty is intensified; when they are shorter, it is relaxed. This adaptive mechanism guides models toward producing concise reasoning without sacrificing task performance. Experiments on Deepseek-R1-Distill-Qwen-1.5B and Qwen3-4B-Thinking-2507 show that Leash reduces the average reasoning length by 60% across diverse tasks - including in-distribution mathematical reasoning and out-of-distribution domains such as coding and instruction following - while maintaining competitive performance. Our work thus presents a practical and effective paradigm for developing controllable and efficient LLMs that balance reasoning capabilities with computational budgets.

cs.AI↗

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↗

arXiv2025

Ming-Flash-Omni: A Sparse, Unified Architecture for Multimodal Perception and Generation

We propose Ming-Flash-Omni, an upgraded version of Ming-Omni, built upon a sparser Mixture-of-Experts (MoE) variant of Ling-Flash-2.0 with 100 billion total parameters, of which only 6.1 billion are active per token. This architecture enables highly efficient scaling (dramatically improving computational efficiency while significantly expanding model capacity) and empowers stronger unified multimodal intelligence across vision, speech, and language, representing a key step toward Artificial General Intelligence (AGI). Compared to its predecessor, the upgraded version exhibits substantial improvements across multimodal understanding and generation. Notably, it achieves strong performance on vision-language understanding benchmarks, with overall scores on par with Gemini 2.5 Pro, and enables seamless switching among multimodal tasks in multi-turn interactions. In speech, it achieves strong performance in contextual and dialect-aware ASR while enabling joint, continuous-generation of speech, sound, and music. In vision, it introduces generative semantic segmentation that achieves competitive standalone performance and enhances spatial control and editing consistency, alongside marked improvements in identity preservation, and high-fidelity in-image text rendering. Together, these capabilities demonstrate that a single unified model can serve as a practical foundation for general-purpose multimodal intelligence.

cs.CV↗

arXiv2025

DARO: Difficulty-Aware Reweighting Policy Optimization

Recent advances in large language models (LLMs) have shown that reasoning ability can be significantly enhanced through Reinforcement Learning with Verifiable Rewards (RLVR). Group Relative Policy Optimization (GRPO) has emerged as the de facto approach for RLVR, inspiring numerous variants. However, our mathematical analysis reveals that these methods are fundamentally weighted variations of GRPO. We provide a unified view, demonstrating that their reliance on static or overly simplistic weighting schemes tied to sample difficulty prevents adaptation to a model's evolving capabilities. This creates a significant loss scale issue, where training disproportionately focuses on certain difficulty levels at the expense of others, hindering overall performance. To address these limitations, we introduce \textbf{Difficulty-Aware Reweighting Policy Optimization (DARO)}, a method that dynamically adjusts the loss contribution of each difficulty group based on the model's learning state. Extensive experiments on Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, and Llama3.1-8B show that DARO outperforms four leading baselines across six math benchmarks, achieving significantly faster convergence and superior final performance.

cs.CL↗

arXiv2025

Global Mean-Amplitude Enhanced Spiking Neural Network Coherent Ising Machine

The coherent Ising machine (CIM) is a quantum-inspired computing platform that leverages optical parametric oscillation dynamics to solve combinatorial optimization problems by searching for the ground state of an Ising Hamiltonian. Conventional CIM implementations face challenges in handling non-uniform coupling strengths and maintaining amplitude stability during computation. In this paper, a new global mean-amplitude feedback-enhanced spiking neural network CIM (GFSNN-CIM) is introduced with a physics-driven amplitude stabilization mechanism to dynamically balance nonlinear gain saturation and coupling effects. This modification enhances synchronization in the optical pulse network, leading to more robust convergence under varying interaction strengths. Experimental validation on Max-Cut problems demonstrates that the GFSNN-CIM achieves up to a 27% improvement in solution success rates compared to conventional spiking neural network CIM, with scalability improving as problem complexity increases. Further application to the traffic assignment problem (TAP) confirms the method's generality; the GFSNN-CIM achieves near-continuous accuracy (deviations < 0.035%) even at coarse discretization, while large-scale tests on Beijing's road network (481 spins) validate its real-world applicability. These advances establish a physics-consistent optimization framework, where optical pulse dynamics directly encode combinatorial problems, paving the way for scalable, high-performance CIM implementations in complex optimization tasks.

quant-ph↗

arXiv2025

Reliable and Real-Time Highway Trajectory Planning via Hybrid Learning-Optimization Frameworks

Autonomous highway driving involves high-speed safety risks due to limited reaction time, where rare but dangerous events may lead to severe consequences. This places stringent requirements on trajectory planning in terms of both reliability and computational efficiency. This paper proposes a hybrid highway trajectory planning (H-HTP) framework that integrates learning-based adaptability with optimization-based formal safety guarantees. The key design principle is a deliberate division of labor: a learning module generates a traffic-adaptive velocity profile, while all safety-critical decisions including collision avoidance and kinematic feasibility are delegated to a Mixed-Integer Quadratic Program (MIQP). This design ensures that formal safety constraints are always enforced, regardless of the complexity of multi-vehicle interactions. A linearization strategy for the vehicle geometry substantially reduces the number of integer variables, enabling real-time optimization without sacrificing formal safety guarantees. Experiments on the HighD dataset demonstrate that H-HTP achieves a scenario success rate above 97% with an average planning-cycle time of approximately 54 ms, reliably producing smooth, kinematically feasible, and collision-free trajectories in safety-critical highway scenarios.

cs.RO↗