arXiv subjects

Yuxuan Zheng

Publications and source records attributed to Yuxuan Zheng.

14 recordsLinked to original sources

arXiv2026

RewardWeaver: Long-Horizon Interactive Learning for Language Agents via Self-Evolving Reward Adaptation

Reinforcement learning with verifiable rewards (RLVR) has driven substantial progress in domains where task outcomes can be reliably evaluated, but long-horizon interaction remains challenging due to sparse terminal feedback and difficult credit assignment. Process rewards provide denser supervision, yet the capabilities most relevant for training can change as the policy evolves: a behavior that is easy to evaluate or frequently deficient need not be the bottleneck currently limiting task success. We introduce RewardWeaver, a self-evolving reward adaptation framework for language agents in long-horizon interaction. RewardWeaver maintains a validated capability space in which the semantics of admitted Rubrics remain fixed, and closes the loop between policy optimization, task evaluation, failure attribution, and reward adaptation. After each training stage, it performs outcome-grounded backward attribution on low-outcome trajectories, aggregates recurrent and policy-controlled capability bottlenecks, and dynamically selects the corresponding process rewards for the next stage. Recurrent failures not covered by the existing capability space trigger a separate, controlled expansion procedure. We evaluate REWARDWEAVER on SOTOPIA, Amazon?HistoryPrice, and a newly constructed Sales Benchmark. Across social interaction, bilateral bargaining, and domain-specific sales, REWARDWEAVER establishes new state-of-the-art (SOTA) results. Ablations further demonstrate the importance of dynamic reward allocation, failure-grounded attribution, and stable semantics for admitted capabilities.

cs.AI↗

arXiv2026

PiERN: Token-Level Routing for Integrating High-Precision Computation and Reasoning

Tasks on complex systems require high-precision numerical computation to support decisions. However, current large language models (LLMs), even with enhanced reasoning capabilities, cannot integrate such computations as an intrinsic and interpretable capability with existing architectures. To this end, we propose Physically-isolated Experts Routing Network (PiERN), an architecture that directs computation and reasoning at token level, thereby enabling iterative alternation within a single chain of thought. We systematically evaluate PiERN on representative computation-reasoning tasks, including PDEBench and battery management tasks. Results show that PiERN achieves not only higher accuracy than directly finetuning LLMs but also significant improvements in response latency, token usage, GPU energy consumption, and experts routing accuracy compared with mainstream multi-agent approaches, while exhibiting no significant degradation in performance on MMLU and GLUE benchmarks. PiERN offers an efficient, interpretable, and scalable paradigm for interfacing language models with scientific systems.

cs.LG↗

arXiv2026

Full Gabor frames, its existence problem, and a non-uniform Balian-Low type theorem

For a broad class of Delone sets in $\mathbb{R}^n$ that are of significance in both mathematics and physics, we prove a non-uniform Balian-Low type theorem and settle the converse problem on the existence of Gabor frames, for arbitrary dimension $n$. To this end, we introduce a class of Gabor frames, termed full Gabor frames, and prove that the existence of such a frame on the Delone set with Schwartz window functions is equivalent to the condition that the lower Beurling density be strictly greater than one. In fact, the usual Balian-Low direction using window functions from the Feichtinger's algebra can be proven for arbitrary point sets, thereby improving an earlier density theorem by Christensen, Deng, and Heil. The corresponding dual result for Riesz sequences is also obtained. The main technical tools employed in this paper are tiling groupoid constructions and $C^*$-algebraic methods. As a byproduct, we resolve an open question from Ito's thesis concerning the bounded dynamical asymptotic dimension of tiling groupoids. Furthermore, this result allows us to extend the classification theorem of Ito, Whittaker, and Zacharias to the twisted case.

math.FA↗

arXiv2026

The Time-Frequency Covariance Principle on Unimodular Kac Algebras

This paper extends the short-time Fourier transform (STFT), a fundamental tool in time-frequency analysis, to the quantum group setting of unimodular Kac algebras. For a unimodular Kac algebra \mathbb{G}, we introduce a time-frequency shift operator that combines left translation and modulation operators. Using a window vector in the Hilbert space L^2(\mathbb{G}), we define the corresponding STFT and establish its essential analytic properties, including a Plancherel theorem, the Moyal identity, an inversion formula, and a fundamental identity. Furthermore, we explore the projective corepresentation structure of the time-frequency shift operator, and prove that its reflected version induces a continuous projective left representation of the dual quantum group of the quantum double. Finally, we derive the covariance principle and several uncertainty principles.

math.FA↗

arXiv2026

Reversing Heat Flow by Coherence in a Multipartite Quantum System

The second law of thermodynamics dictates that heat flows spontaneously from a high-temperature entity to a lower-temperature one. Yet, recent advances have demonstrated that quantum correlations between a system and its thermal environment can induce a reversal of heat flow, challenging classical thermodynamic expectations. Here, we experimentally demonstrate that internal quantum coherence in a multipartite spin system can also reverse heat flow, without relying on initial correlations with the environment. Under the collision model with cascade interaction, we verify that both the strength and the phase of the coherence term determine the direction and magnitude of energy transfer. These results enable precise control of heat flow using only local quantum properties.

quant-ph↗

arXiv2025

Noise-Resilient Quantum Metrology with Quantum Computing

Quantum computing has made remarkable strides in recent years, as demonstrated by quantum supremacy experiments and the realization of high-fidelity, fault-tolerant gates. However, a major obstacle persists: practical real-world applications remain scarce, largely due to the inefficiency of loading classical data into quantum processors. Here, we propose an alternative strategy that shifts the focus from classical data encoding to directly processing quantum data. We target quantum metrology, a practical quantum technology whose precision is often constrained by realistic noise. We develop an experimentally feasible scheme in which a quantum computer optimizes information acquired from quantum metrology, thereby enhancing performance in noisy quantum metrology tasks and overcoming the classical-data-loading bottleneck. We demonstrate this approach through experimental implementation with nitrogen-vacancy centers in diamond and numerical simulations using models of distributed superconducting quantum processors. Our results show that this method improves the accuracy of sensing estimates and significantly boosts sensitivity, as quantified by the quantum Fisher information, thus offering a new pathway to harness near-term quantum computers for realistic quantum metrology.

quant-ph↗

arXiv2025

Experimental Virtual Quantum Broadcasting

The quantum no-broadcasting theorem states that it is fundamentally impossible to perfectly replicate an arbitrary quantum state, even if correlations between the copies are allowed. While quantum broadcasting cannot occur through any physical process, it can be achieved via postprocessing of experimental data using a process called virtual quantum broadcasting (VQB). In this work, we report the experimental implementation of a quantum circuit based on the linear combination of unitaries, integrated with a post-processing protocol, to realize VQB in a nuclear magnetic resonance system. VQB can be expressed as a linear combination of two channels: the universal cloner, which broadcasts the target quantum state, and the universal antisymmetrizer, which reduces broadcasting error. We implement both channels within the same circuit and demonstrate that the universal cloner is the closest physical map to VQB. In addition, we show how the universal antisymmetrizer can be utilized to mitigate imperfections in the cloner, enabling near-ideal fidelity. Our method is applicable to broadcasting quantum systems of any dimension.

quant-ph↗

arXiv2025

Odyssey: Empowering Minecraft Agents with Open-World Skills

Recent studies have delved into constructing generalist agents for open-world environments like Minecraft. Despite the encouraging results, existing efforts mainly focus on solving basic programmatic tasks, e.g., material collection and tool-crafting following the Minecraft tech-tree, treating the ObtainDiamond task as the ultimate goal. This limitation stems from the narrowly defined set of actions available to agents, requiring them to learn effective long-horizon strategies from scratch. Consequently, discovering diverse gameplay opportunities in the open world becomes challenging. In this work, we introduce Odyssey, a new framework that empowers Large Language Model (LLM)-based agents with open-world skills to explore the vast Minecraft world. Odyssey comprises three key parts: (1) An interactive agent with an open-world skill library that consists of 40 primitive skills and 183 compositional skills. (2) A fine-tuned LLaMA-3 model trained on a large question-answering dataset with 390k+ instruction entries derived from the Minecraft Wiki. (3) A new agent capability benchmark includes the long-term planning task, the dynamic-immediate planning task, and the autonomous exploration task. Extensive experiments demonstrate that the proposed Odyssey framework can effectively evaluate different capabilities of LLM-based agents. All datasets, model weights, and code are publicly available to motivate future research on more advanced autonomous agent solutions.

cs.AI↗

arXiv2025

Bi-level Mean Field: Dynamic Grouping for Large-Scale MARL

Large-scale Multi-Agent Reinforcement Learning (MARL) often suffers from the curse of dimensionality, as the exponential growth in agent interactions significantly increases computational complexity and impedes learning efficiency. To mitigate this, existing efforts that rely on Mean Field (MF) simplify the interaction landscape by approximating neighboring agents as a single mean agent, thus reducing overall complexity to pairwise interactions. However, these MF methods inevitably fail to account for individual differences, leading to aggregation noise caused by inaccurate iterative updates during MF learning. In this paper, we propose a Bi-level Mean Field (BMF) method to capture agent diversity with dynamic grouping in large-scale MARL, which can alleviate aggregation noise via bi-level interaction. Specifically, BMF introduces a dynamic group assignment module, which employs a Variational AutoEncoder (VAE) to learn the representations of agents, facilitating their dynamic grouping over time. Furthermore, we propose a bi-level interaction module to model both inter- and intra-group interactions for effective neighboring aggregation. Experiments across various tasks demonstrate that the proposed BMF yields results superior to the state-of-the-art methods.

cs.AI↗

arXiv2025

Neural quantum embedding via deterministic quantum computation with one qubit

Quantum computing is expected to provide exponential speedup in machine learning. However, optimizing the data loading process, commonly referred to as quantum data embedding, to maximize classification performance remains a critical challenge. In this work, we propose a neural quantum embedding (NQE) technique based on deterministic quantum computation with one qubit (DQC1). Unlike the traditional embedding approach, NQE trains a neural network to maximize the trace distance between quantum states corresponding to different categories of classical data. Furthermore, training is efficiently achieved using DQC1, which is specifically designed for ensemble quantum systems, such as nuclear magnetic resonance (NMR). We validate the NQE-DQC1 protocol by encoding handwritten images into NMR quantum processors, demonstrating a significant improvement in distinguishability compared to traditional methods. Additionally, after training the NQE, we implement a parameterized quantum circuit for classification tasks, achieving 98\% classification accuracy, in contrast to the 54\% accuracy obtained using traditional embedding. Moreover, we show that the NQE-DQC1 protocol is extendable, enabling the use of the NMR system for NQE training due to its high compatibility with DQC1, while subsequent machine learning tasks can be performed on other physical platforms, such as superconducting circuits. Our work opens new avenues for utilizing ensemble quantum systems for efficient classical data embedding into quantum registers.

quant-ph↗

arXiv2024

Experimental Realization of Self-Contained Quantum Refrigeration

A fundamental challenge in quantum thermodynamics is the exploration of inherent dimensional constraints in thermodynamic machines. In the context of two-level systems, the most compact refrigerator necessitates the involvement of three entities, operating under self-contained conditions that preclude the use of external work sources. Here, we build such a smallest refrigerator using a nuclear spin system, where three distinct two-level carbon-13 nuclei in the same molecule are involved to facilitate the refrigeration process. The self-contained feature enables it to operate without relying on net external work, and the unique mechanism sets this refrigerator apart from its classical counterparts. We evaluate its performance under varying conditions and systematically scrutinize the cooling constraints across a spectrum of scenarios, which sheds light on the interplay between quantum information and thermodynamics.

quant-ph↗

arXiv2024

Self-Consistent Determination of Single-Impurity Anderson Model Using Hybrid Quantum-Classical Approach on a Spin Quantum Simulator

The accurate determination of the electronic structure of strongly correlated materials using first principle methods is of paramount importance in condensed matter physics, computational chemistry, and material science. However, due to the exponential scaling of computational resources, incorporating such materials into classical computation frameworks becomes prohibitively expensive. In 2016, Bauer et al. proposed a hybrid quantum-classical approach to correlated materials Phys. Rev. X 6, 031045 (2016)}] that can efficiently tackle the electronic structure of complex correlated materials. Here, we experimentally demonstrate that approach to tackle the computational challenges associated with strongly correlated materials. By seamlessly integrating quantum computation into classical computers, we address the most computationally demanding aspect of the calculation, namely the computation of the Green's function, using a spin quantum processor. Furthermore, we realize a self-consistent determination of the single impurity Anderson model through a feedback loop between quantum and classical computations. A quantum phase transition in the Hubbard model from the metallic phase to the Mott insulator is observed as the strength of electron correlation increases. As the number of qubits with high control fidelity continues to grow, our experimental findings pave the way for solving even more complex models, such as strongly correlated crystalline materials or intricate molecules.

quant-ph↗

arXiv2021

Multi-modal Self-supervised Pre-training for Regulatory Genome Across Cell Types

In the genome biology research, regulatory genome modeling is an important topic for many regulatory downstream tasks, such as promoter classification, transaction factor binding sites prediction. The core problem is to model how regulatory elements interact with each other and its variability across different cell types. However, current deep learning methods often focus on modeling genome sequences of a fixed set of cell types and do not account for the interaction between multiple regulatory elements, making them only perform well on the cell types in the training set and lack the generalizability required in biological applications. In this work, we propose a simple yet effective approach for pre-training genome data in a multi-modal and self-supervised manner, which we call GeneBERT. Specifically, we simultaneously take the 1d sequence of genome data and a 2d matrix of (transcription factors x regions) as the input, where three pre-training tasks are proposed to improve the robustness and generalizability of our model. We pre-train our model on the ATAC-seq dataset with 17 million genome sequences. We evaluate our GeneBERT on regulatory downstream tasks across different cell types, including promoter classification, transaction factor binding sites prediction, disease risk estimation, and splicing sites prediction. Extensive experiments demonstrate the effectiveness of multi-modal and self-supervised pre-training for large-scale regulatory genomics data.

q-bio.GN↗

arXiv2021

Horizontally Fused Training Array: An Effective Hardware Utilization Squeezer for Training Novel Deep Learning Models

Driven by the tremendous effort in researching novel deep learning (DL) algorithms, the training cost of developing new models increases staggeringly in recent years. We analyze GPU cluster usage statistics from a top research institute for more insights into the hardware efficiency achieved by typical DL training jobs. Our study reveals that single-accelerator training jobs can dominate the cluster-wide resource consumption when launched repetitively (e.g., for hyper-parameter tuning) while severely under-utilizing the hardware. Fortunately, we observe that such workloads have the following unique characteristics: (i) the models among jobs often have the same types of operators with the same shapes, and (ii) the inter-model horizontal fusion of such operators is mathematically equivalent to other already well-optimized operators. Thus, to help DL researchers and practitioners effectively improve the hardware utilization of their novel DL training workloads, we propose Horizontally Fused Training Array (HFTA). HFTA is a new DL framework extension library that horizontally fuses the models from different repetitive jobs deeply down to operators and then trains them simultaneously on a shared accelerator. To show the generality of our solution, we apply HFTA to six DL models training on state-of-the-art accelerators (GPUs and TPUs). Our results indicate that HFTA is highly effective in improving hardware utilization and achieves up to $15.1 \times$ higher training throughput vs. the standard practice of running each job on a separate accelerator.

cs.LG↗