arXiv subjects

Yuan Wei

Publications and source records attributed to Yuan Wei.

At least 19 recordsLinked to original sources

arXiv2026

Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models

On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning behavior rather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student: same-origin pairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereas cross-origin pairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher's influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosing multi-teacher OPD.

cs.CL↗

arXiv2026

FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents

Training deep search agents requires verifiable questions whose answers remain unavailable until sufficient evidence has been acquired through search. Existing synthesis methods often increase apparent difficulty by enriching graph structures, but structural complexity alone does not guarantee realized search difficulty: the intended search process can collapse through a cheaper identifying route. We formalize this gap with a shortcut-aware difficulty framework and identify four actionable shortcut risks: evidence co-coverage, single-clue selectivity, exposed constants, and prior-knowledge binding. To diagnose their realized effects, we use trajectory signatures including solving cost, answer hit time, and prior-shortcut rate. Guided by this framework, we introduce FORT, a Framework of Shortcut-Resistant Training-Data Synthesis. FORT constructs shortcut-resistant training data by controlling shortcut risks across entity selection, evidence graph construction, question formulation, and adversarial refinement. Experiments show that FORT induces longer pre-answer search and fewer shortcut patterns than existing open-source deep search datasets. Using the resulting trajectories, we train FORT-Searcher with supervised fine-tuning (SFT) only, and it achieves the best overall performance among comparable-size open-source search agents on challenging deep search benchmarks. Relevant resources will be made available at https://github.com/RUCAIBox/FORT-Searcher.

cs.CL↗

arXiv2026

TMAS: Scaling Test-Time Compute via Multi-Agent Synergy

Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks show that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, with hybrid reward training further improving scaling effectiveness and stability across iterations. Code and data are available at https://github.com/IQuestLab/tmas.

cs.AI↗

arXiv2026

ClawGym: A Scalable Framework for Building Effective Claw Agents

Claw-style environments support multi-step workflows over local files, tools, and persistent workspace states. However, scalable development around these environments remains constrained by the absence of a systematic framework, especially one for synthesizing verifiable training data and integrating it with agent training and diagnostic evaluation. To address this challenge, we present ClawGym, a scalable framework that supports the full lifecycle of Claw-style personal agent development. Concretely, we construct ClawGym-SynData, a diverse dataset of 13.5K filtered tasks synthesized from persona-driven intents and skill-grounded operations, paired with realistic mock workspaces and hybrid verification mechanisms. We then train a family of capable Claw-style models, termed ClawGym-Agents, through supervised fine-tuning on black-box rollout trajectories, and further explore reinforcement learning via a lightweight pipeline that parallelizes rollouts across per-task sandboxes. To support reliable evaluation, we further construct ClawGym-Bench, a benchmark of 200 instances calibrated through automated filtering and human-LLM review. Relevant resources have been released at https://github.com/ClawGym.

cs.CL↗

arXiv2026

OpAgent: Operator Agent for Web Navigation

To fulfill user instructions, autonomous web agents must contend with the inherent complexity and volatile nature of real-world websites. Conventional paradigms predominantly rely on Supervised Fine-Tuning (SFT) or Offline Reinforcement Learning (RL) using static datasets. However, these methods suffer from severe distributional shifts, as offline trajectories fail to capture the stochastic state transitions and real-time feedback of unconstrained wide web environments. In this paper, we propose a robust Online Reinforcement Learning WebAgent, designed to optimize its policy through direct, iterative interactions with unconstrained wide websites. Our approach comprises three core innovations: 1) Hierarchical Multi-Task Fine-tuning: We curate a comprehensive mixture of datasets categorized by functional primitives -- Planning, Acting, and Grounding -- establishing a Vision-Language Model (VLM) with strong instruction-following capabilities for Web GUI tasks. 2) Online Agentic RL in the Wild: We develop an online interaction environment and fine-tune the VLM using a specialized RL pipeline. We introduce a Hybrid Reward Mechanism that combines a ground-truth-agnostic WebJudge for holistic outcome assessment with a Rule-based Decision Tree (RDT) for progress reward. This system effectively mitigates the credit assignment challenge in long-horizon navigation. Notably, our RL-enhanced model achieves a 38.1\% success rate (pass@5) on WebArena, outperforming all existing monolithic baselines. 3) Operator Agent: We introduce a modular agentic framework, namely \textbf{OpAgent}, orchestrating a Planner, Grounder, Reflector, and Summarizer. This synergy enables robust error recovery and self-correction, elevating the agent's performance to a new State-of-the-Art (SOTA) success rate of \textbf{71.6\%}.

cs.AI↗

arXiv2026

Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics

Clinician skepticism toward opaque AI hinders adoption in high-stakes healthcare. We present AICare, an interactive and interpretable AI copilot for collaborative clinical decision-making. By analyzing longitudinal electronic health records, AICare grounds dynamic risk predictions in scrutable visualizations and LLM-driven diagnostic recommendations. Through a within-subjects counterbalanced study with 16 clinicians across nephrology and obstetrics, we comprehensively evaluated AICare using objective measures (task completion time and error rate), subjective assessments (NASA-TLX, SUS, and confidence ratings), and semi-structured interviews. Our findings indicate AICare's reduced cognitive workload. Beyond performance metrics, qualitative analysis reveals that trust is actively constructed through verification, with interaction strategies diverging by expertise: junior clinicians used the system as cognitive scaffolding to structure their analysis, while experts engaged in adversarial verification to challenge the AI's logic. This work offers design implications for creating AI systems that function as transparent partners, accommodating diverse reasoning styles to augment rather than replace clinical judgment.

cs.HC↗

arXiv2025

$Agent^2$: An Agent-Generates-Agent Framework for Reinforcement Learning Automation

Reinforcement learning (RL) agent development traditionally requires substantial expertise and iterative effort, often leading to high failure rates and limited accessibility. This paper introduces Agent$^2$, an LLM-driven agent-generates-agent framework for fully automated RL agent design. Agent$^2$ autonomously translates natural language task descriptions and environment code into executable RL solutions without human intervention. The framework adopts a dual-agent architecture: a Generator Agent that analyzes tasks and designs agents, and a Target Agent that is automatically generated and executed. To better support automation, RL development is decomposed into two stages, MDP modeling and algorithmic optimization, facilitating targeted and effective agent generation. Built on the Model Context Protocol, Agent$^2$ provides a unified framework for standardized agent creation across diverse environments and algorithms, incorporating adaptive training management and intelligent feedback analysis for continuous refinement. Extensive experiments on benchmarks including MuJoCo, MetaDrive, MPE, and SMAC show that Agent$^2$ outperforms manually designed baselines across all tasks, achieving up to 55\% performance improvement with consistent average gains. By enabling a closed-loop, end-to-end automation pipeline, this work advances a new paradigm in which agents can design and optimize other agents, underscoring the potential of agent-generates-agent systems for automated AI development.

cs.AI↗

arXiv2025

Persistent quantum vibronic dynamics in a $5d^1$ double perovskite oxide

Quantum entanglement between the spin, orbital, and lattice degrees of freedom in condensed matter systems can emerge due to an interplay between spin-orbit and vibronic interactions. Heavy transition metal ions decorated on a face-centered cubic lattice, for example, in $5d^1$ double perovskites, are particularly suited to support these quantum entangled states, but direct evidence has not yet been presented. In this work, we report additional peaks in the low-energy spectra of a $5d^1$ double perovskite, Ba$_2$CaReO$_6$, which cannot be explained by adopting a purely classical description of lattice vibrations. Instead, our theoretical analysis demonstrates that these spectroscopic signatures are characteristic of orbital-lattice entangled states in Ba$_2$CaReO$_6$. Crucially, both theory and experiment demonstrate that these quantum-entangled states persist to low temperatures, despite the onset of multipolar order.

cond-mat.str-el↗

arXiv2025

LERO: LLM-driven Evolutionary framework with Hybrid Rewards and Enhanced Observation for Multi-Agent Reinforcement Learning

Multi-agent reinforcement learning (MARL) faces two critical bottlenecks distinct from single-agent RL: credit assignment in cooperative tasks and partial observability of environmental states. We propose LERO, a framework integrating Large language models (LLMs) with evolutionary optimization to address these MARL-specific challenges. The solution centers on two LLM-generated components: a hybrid reward function that dynamically allocates individual credit through reward decomposition, and an observation enhancement function that augments partial observations with inferred environmental context. An evolutionary algorithm optimizes these components through iterative MARL training cycles, where top-performing candidates guide subsequent LLM generations. Evaluations in Multi-Agent Particle Environments (MPE) demonstrate LERO's superiority over baseline methods, with improved task performance and training efficiency.

cs.LG↗

arXiv2025

Billet Number Recognition Based on Test-Time Adaptation

During the steel billet production process, it is essential to recognize machine-printed or manually written billet numbers on moving billets in real-time. To address the issue of low recognition accuracy for existing scene text recognition methods, caused by factors such as image distortions and distribution differences between training and test data, we propose a billet number recognition method that integrates test-time adaptation with prior knowledge. First, we introduce a test-time adaptation method into a model that uses the DB network for text detection and the SVTR network for text recognition. By minimizing the model's entropy during the testing phase, the model can adapt to the distribution of test data without the need for supervised fine-tuning. Second, we leverage the billet number encoding rules as prior knowledge to assess the validity of each recognition result. Invalid results, which do not comply with the encoding rules, are replaced. Finally, we introduce a validation mechanism into the CTC algorithm using prior knowledge to address its limitations in recognizing damaged characters. Experimental results on real datasets, including both machine-printed billet numbers and handwritten billet numbers, show significant improvements in evaluation metrics, validating the effectiveness of the proposed method.

cs.CV↗

arXiv2025

Spin correlations in La$_3$Ni$_2$O$_7$ thin films

The discovery of ambient-pressure superconductivity with $T_{c,\text{onset}} > 40$ K in La$_3$Ni$_2$O$_7$ (LNO) thin films grown on the SrLaAlO$_4$ (SLAO) substrate with compressive ($\varepsilon\approx-2\%$) epitaxial strain provides a unique platform for investigating the superconducting mechanism in nickelate superconductors. Here, we use resonant inelastic X-ray scattering (RIXS) to unveil the dispersive spin excitations in the LNO/SLAO thin film and establish the strain dependence of the electronic and spin excitations in LNO thin films with strain ranging from $\varepsilon\approx-2\%$ to $+1.9\%$. Compared with bulk LNO, LNO/SLAO exhibits similar $dd$ excitations and spin dynamics, but with a larger spin-excitation bandwidth, whereas tensile-strained LNO/SrTiO$_3$ exhibits a marked suppression of both the spin excitations and the Ni $3d_{z^2}$-derived $dd$ excitations. This evolution reflects a strain-tuned interlayer exchange interaction $J_z$ and Ni $3d_{z^2}$-O 2$p_z$ hybridization. Our results demonstrate how epitaxial strain modulates the interlayer magnetic coupling and are consistent with scenarios in which the interlayer antiferromagnetic superexchange interaction promotes interlayer pairing in bilayer nickelates.

cond-mat.supr-con↗

arXiv2024

Smart Pressure e-Mat for Human Sleeping Posture and Dynamic Activity Recognition

With the emphasis on healthcare, early childhood education, and fitness, non-invasive measurement and recognition methods have received more attention. Pressure sensing has been extensively studied because of its advantages of simple structure, easy access, visualization application, and harmlessness. This paper introduces a Smart Pressure e-Mat (SPeM) system based on piezoresistive material, Velostat, for human monitoring applications, including recognition of sleeping postures, sports, and yoga. After a subsystem scans the e-mat readings and processes the signal, it generates a pressure image stream. Deep neural networks (DNNs) are used to fit and train the pressure image stream and recognize the corresponding human behavior. Four sleeping postures and 13 dynamic activities inspired by Nintendo Switch Ring Fit Adventure (RFA) are used as a preliminary validation of the proposed SPeM system. The SPeM system achieves high accuracies in both applications, demonstrating the high accuracy and generalizability of the models. Compared with other pressure sensor-based systems, SPeM possesses more flexible applications and commercial application prospects, with reliable, robust, and repeatable properties.

cs.CV↗

arXiv2024

Emergence of interfacial magnetism in strongly-correlated nickelate-titanate superlattices

Strongly-correlated transition-metal oxides are widely known for their various exotic phenomena. This is exemplified by rare-earth nickelates such as LaNiO$_{3}$, which possess intimate interconnections between their electronic, spin, and lattice degrees of freedom. Their properties can be further enhanced by pairing them in hybrid heterostructures, which can lead to hidden phases and emergent phenomena. An important example is the LaNiO$_{3}$/LaTiO$_{3}$ superlattice, where an interlayer electron transfer has been observed from LaTiO$_{3}$ into LaNiO$_{3}$ leading to a high-spin state. However, macroscopic emergence of magnetic order associated with this high-spin state has so far not been observed. Here, by using muon spin rotation, x-ray absorption, and resonant inelastic x-ray scattering, we present direct evidence of an emergent antiferromagnetic order with high magnon energy and exchange interactions at the LaNiO$_{3}$/LaTiO$_{3}$ interface. As the magnetism is purely interfacial, a single LaNiO$_{3}$/LaTiO$_{3}$ interface can essentially behave as an atomically thin strongly-correlated quasi-two-dimensional antiferromagnet, potentially allowing its technological utilisation in advanced spintronic devices. Furthermore, its strong quasi-two-dimensional magnetic correlations, orbitally-polarized planar ligand holes, and layered superlattice design make its electronic, magnetic, and lattice configurations resemble the precursor states of superconducting cuprates and nickelates, but with an $S \rightarrow 1$ spin state instead.

cond-mat.str-el↗

arXiv2024

Spin waves and orbital contribution to ferromagnetism in a topological metal

Special arrangements of atoms with more than one atom per unit cell, including honeycomb or kagome (woven bamboo mat) lattices, can host propagating excitations with non-trivial topology as defined by their evolution along closed paths in momentum space. Excitations on such lattices can also be momentum-independent, meaning that they are localized notwithstanding strong hopping of the underlying disturbances between neighbouring sites. The associated flat bands are interesting because the interactions between the heavy quasiparticles inhabiting them will become much more important than for strong dispersion, resulting in novel quantum solid and liquid states. Different stackings of two-dimensional lattices, for example twisted graphene bilayers, provide routes to further engineer topology and many-body effects. Here, we report the discovery, using circularly polarized x-rays for the unambiguous isolation of magnetic signals, of a nearly flat spin wave band and large (compared to elemental iron) orbital moment for the metallic ferromagnet Fe3Sn2, built from compact AB-stacked kagome bilayers and which has a topologically non-trivial electronic band structure controllable by modest external magnetic fields. As a function of out-of-plane momentum, the nearly flat optical mode and the global rotation symmetry-restoring acoustic mode are out of phase, consistent with a bilayer exchange coupling that is larger than the already large in-plane couplings. The defining units of this topological metal are therefore a triangular lattice of octahedral iron clusters rather than weakly coupled kagome planes. The spin waves are strongly damped when compared to elemental iron, opening the topic of interactions of topological bosons (spin waves) and fermions (electrons) with the very specific target of explaining boson lifetimes.

cond-mat.str-el↗

arXiv2024

Dynamic Jahn-Teller effect in the strong spin-orbit coupling regime

Exotic quantum phases, arising from a complex interplay of charge, spin, lattice and orbital degrees of freedom, are of immense interest to a wide research community. A well-known example of such an entangled behavior is the Jahn-Teller effect, where the lifting of orbital degeneracy proceeds through lattice distortions, often accompanied by ordering of spins and metal-insulator transitions. Static distortions, including cooperative behavior, have been associated with colossal magneto-resistance, multiferroicity, high-$T_\mathrm{C}$ superconductivity and other correlated phenomena. Realizations of the dynamic Jahn-Teller effect, on the other hand, are scarce since the preservation of vibronic symmetries requires subtle tuning of the local environment. Here we demonstrate that a highly-symmetrical 5d$^1$ double perovskite Ba$_2$MgReO$_6$, comprising of a 3D array of isolated ReO$_6$ octahedra, fulfils these requirements, resulting in a unique case of a dynamic Jahn-Teller system with strong spin-orbit coupling. Thermodynamic and resonant inelastic x-ray scattering experiments undoubtedly show that the Jahn-Teller instability leads to a ground-state doublet, invoking a paradigm shift for this family of compounds. The restoration of vibronic degrees of freedom arises from a quantum-mechanical zero-point motion, as revealed by detailed quantum chemistry calculations. The dynamic state of ReO$_6$ octahedra persists down to the lowest temperatures, where a multipolar order sets in, allowing for investigations of the interplay between a dynamic JT effect and strongly correlated electron behavior.

cond-mat.str-el↗

arXiv2024

Spin-orbital excitations encoding the magnetic phase transition in the van der Waals antiferromagnet FePS$_{3}$

In the rich phases of van der Waals (vdW) materials featuring intertwined electronic order and collective phenomena, characterizing elementary dynamics that entail the low-energy Hamiltonian and electronic degrees of freedom is of paramount importance. Here we performed resonant inelastic X-ray scattering (RIXS) to elaborate the spin-orbital ground and excited states of the vdW antiferromagnetic insulator FePS$_{3}$, as well as their relation to magnetism. We observed the spectral enhancement of spin-orbital multiplet transitions about $\sim$ 100 and $\sim$ 220 meV, as well as quasielastic response, when entering the zig-zag antiferromagnetic phase, where the spectral changes develop an order-parameter-like evolution with temperature. By comparing with ligand field theory calculations, we discovered the essential role of trigonal lattice distortion and negative metal-ligand charge-transfer to account for these emergent excitations. Such spectral profiles are further examined upon confinement by mechanical exfoliation. We reveal their spectral robustness down to the few atomic layer limit, in accordance with the persistent antiferromagnetic state previously reported in optical measurements. Our study demonstrates the versatile RIXS capability that resolves magneto-crystalline anisotropy and charge-transfer energetics. These provide the crucial insight to understand how the spontaneous magnetic symmetry-breaking stabilizes in the quasi-two-dimensional limit for the vdW magnet FePS$_{3}$.

cond-mat.str-el↗

arXiv2024

Spectral evidence for Dirac spinons in a kagome lattice antiferromagnet

Emergent quasiparticles with a Dirac dispersion in condensed matter systems can be described by the Dirac equation for relativistic electrons, in analogy with Dirac particles in high-energy physics. For example, electrons with a Dirac dispersion have been intensively studied in electronic systems such as graphene and topological insulators. However, charge is not a prerequisite for Dirac fermions, and the emergence of Dirac fermions without charge degree of freedom has been theoretically predicted to be realized in Dirac quantum spin liquids. These quasiparticles carry a spin of 1/2 but are charge-neutral, and so are called spinons. Here we show that the spin excitations of a kagome antiferromagnet, YCu$_3$(OD)$_6$Br$_2$[Br$_{0.33}$(OD)$_{0.67}$], are conical with a spin continuum inside, which is consistent with the convolution of two Dirac spinons. The predictions of a Dirac spin liquid model with a spinon velocity obtained from the spectral measurements are in agreement with the low-temperature specific heat of the sample. Our results thus provide spectral evidence for the Dirac quantum spin liquid state emerging in this kagome lattice antiferromagnet. However, the locations of the conical spin excitations differ from those calculated by the nearest neighbor Heisenberg model, suggesting the Dirac spinons have an unexpected origin.

cond-mat.str-el↗

arXiv2023

Nematic charge-density-wave correlations in FeSe$_{1-x}$S$_{x}$

The occurrence of charge-density-wave (CDW) order is a common thread in the phase diagram of cuprate high-transition-temperature ($T_c$) superconductors. In iron-based superconductors (FeSCs), nematic order and fluctuations play a decisive role in driving other emergent orders. CDW order has been observed by scanning tunneling microscopy for various FeSCs such as FeSe thin films, uniaxially strained LiFeAs, and tetragonal FeSe$_{0.81}$S$_{0.19}$. However, it remains elusive if the CDW in these materials is a bulk phenomenon as well as if and how it intertwines with the electronic nematicity. Using energy-resolved resonant X-ray scattering at the Fe-L$_3$ edge, we report the discovery of a local-strain-induced incommensurate isotropic CDW order in FeSe$_{0.82}$S$_{0.18}$. A highly anisotropic CDW response under uniaxial strain unambiguously manifests that the CDW is directly coupled to the nematicity. Transforming part of Fe$^{2+}$ to Fe$^{3+}$ on the surface of FeSe$_{1-x}$S$_{x}$ reveals that the same isotropic CDW can be induced, enhanced, and stabilized in the whole nematic regime measured ($x=0-0.19$). As Fe$^{3+}$ can create local lattice distortions on the surface, the CDW could arise from the interaction between the local strain around Fe$^{3+}$ and the nematic electron correlations. Our experimental observation of a local-strain-induced CDW gives vital information for understanding the interplay between electron correlations and the electronic nematicity in FeSCs.

cond-mat.supr-con↗