arXiv subjects

Sihan Zhao

Publications and source records attributed to Sihan Zhao.

At least 19 recordsLinked to original sources

arXiv2026

From Concentration to Differentiation and Back: Routing Effective Rank in MoE Reasoning Cohorts

Test-time scaling produces cohorts of reasoning rollouts, yet there is no standard label-free account of how their internal computation reorganizes as inference unfolds. We introduce routing effective rank deff, the entropy-effective dimensionality of a cross-rollout graph built from MoE expert-routing similarity. Across ten MoE configurations and five math/science benchmarks, deff exhibits a reproducible low-high-low trajectory, with a prominent interior maximum in 98.5% of 3,105 model-question cohorts: routing similarity is concentrated early, maximally differentiated at intermediate budgets, and reconcentrated later, and the timing of this maximum varies systematically with architecture and reasoning effort. An exact decomposition separates cohort-wide common-mode mass from residual spectral dimensionality: common-mode reallocation accounts for about two thirds of the trajectory, while the residual spectrum contributes about one quarter and retains substantial variation beyond the common mode. The decomposition further localizes behavior: among non-unanimous cohorts, increases in common-mode concentration strongly predict same-answer recoverability, and higher reasoning effort delays the maximum by 2.59 octaves (doublings of the token budget) and consistently expands the high-rank period across all four tested architectures, locating the effort effect in timing and duration rather than peak amplitude. Correctness comparisons separate structural monitoring from answer selection, positioning routing effective rank as a decomposable, label-free diagnostic of cohort organization - a principled spectral lens on how MoE reasoning cohorts differentiate and reconcentrate over inference time.

cs.AI↗

arXiv2026

Beyond the Trace: Coupling an Interpretable Reasoning-State Readout to Native MoE Routing

What a reasoning model writes is only a partial record of the process that produces it. We introduce a two-level internal readout for mixture-of-experts reasoning. We first distill vocabulary-scale J-space into J64, a 64-axis semantic frame learned from the model's own reasoning states. J64 reveals readable process state that the emitted trace does not show: it separates inference effort from problem-induced strain. It also adds 0.096 to 0.135 held-out AUC over a baseline that reads the same rollout as token occupancy and aggregates it in exactly the same way. We then reconstruct J64 from native expert-routing statistics. The result is R64, a low-overhead proxy: its median per-axis correlation with J64 is 0.69 to 0.86 across three models and two families, and on gpt-oss-20b it preserves 95 to 100% of J64's predictive gain. The readout supports test-time decisions at two temporal resolutions. Over completed candidate sets, J64 and R64 improve single-branch selection, and R64-weighted voting improves plain majority voting in seven of eight settings. During generation, rolling readout windows drive a cumulative stop-and-resample policy whose operating point is fixed on training questions alone. J64 improves accuracy by 1.1 to 5.9 points over a sibling-permuted control, and the routing-only R64 proxy retains 0.9 to 3.2 of those points. Finally, router edits aimed at the mechanism J64 names induce the predicted reasoning behaviors and shift a diagnosed stall from numerical guessing toward exact symbolic execution. Together, J64 makes latent process state readable, while routing makes it deployable and actionable.

cs.AI↗

arXiv2026

PP-Motion: Physical-Perceptual Fidelity Evaluation for Human Motion Generation

Human motion generation has found widespread applications in AR/VR, film, sports, and medical rehabilitation, offering a cost-effective alternative to traditional motion capture systems. However, evaluating the fidelity of such generated motions is a crucial, multifaceted task. Although previous approaches have attempted at motion fidelity evaluation using human perception or physical constraints, there remains an inherent gap between human-perceived fidelity and physical feasibility. Moreover, the subjective and coarse binary labeling of human perception further undermines the development of a robust data-driven metric. We address these issues by introducing a physical labeling method. This method evaluates motion fidelity by calculating the minimum modifications needed for a motion to align with physical laws. With this approach, we are able to produce fine-grained, continuous physical alignment annotations that serve as objective ground truth. With these annotations, we propose PP-Motion, a novel data-driven metric to evaluate both physical and perceptual fidelity of human motion. To effectively capture underlying physical priors, we employ Pearson's correlation loss for the training of our metric. Additionally, by incorporating a human-based perceptual fidelity loss, our metric can capture fidelity that simultaneously considers both human perception and physical alignment. Experimental results demonstrate that our metric, PP-Motion, not only aligns with physical laws but also aligns better with human perception of motion fidelity than previous work.

cs.CV↗

arXiv2026

NEX: Neuron Explore-Exploit Scoring for Label-Free Chain-of-Thought Selection and Model Ranking

Large language models increasingly spend inference compute sampling multiple chain-of-thought traces or searching over merged checkpoints. This shifts the bottleneck from generation to selection, often without supervision on the target distribution. We show entropy-based exploration proxies follow an inverted-U with accuracy, suggesting extra exploration can become redundant and induce overthinking. We propose NEX, a white-box label-free unsupervised scoring framework that views reasoning as alternating E-phase (exploration) and X-phase (exploitation). NEX detects E-phase as spikes in newly activated MLP neurons per token from sparse activation caches, then uses a sticky two-state HMM to infer E-X phases and credits E-introduced neurons by whether they are reused in the following X span. These signals yield interpretable neuron weights and a single Good-Mass Fraction score to rank candidate responses and merged variants without task answers. Across reasoning benchmarks and Qwen3 merge families, NEX computed on a small unlabeled activation set predicts downstream accuracy and identifies better variants; we further validate the E-X signal with human annotations and provide causal evidence via "Effective-vs-Redundant" neuron transfer.

cs.AI↗

arXiv2026

ARM: Role-Conditioned Neuron Transplantation for Training-Free Generalist LLM Agent Merging

Interactive large language model agents have advanced rapidly, but most remain specialized to a single environment and fail to adapt robustly to other environments. Model merging offers a training-free alternative by integrating multiple experts into a single model. In this paper, we propose Agent-Role Merging (ARM), an activation-guided, role-conditioned neuron transplantation method for model merging in LLM agents. ARM improves existing merging methods from static natural language tasks to multi-turn agent scenarios, and over the generalization ability across various interactive environments. This is achieved with a well designed 3-step framework: 1) constructing merged backbones, 2) selection based on its role-conditioned activation analysis, and 3) neuron transplantation for fine-grained refinements. Without gradient-based optimization, ARM improves cross-benchmark generalization while enjoying efficiency. Across diverse domains, the model obtained via ARM merging outperforms prior model merging methods and domain-specific expert models, while demonstrating strong out-of-domain generalization.

cs.AI↗

arXiv2025

Emergent electronic insulating states in a one-dimensional moir\'e superlattice

Two-dimensional (2D) van der Waals (vdW) moir\'e superlattices have provided a powerful knob to engineer a plethora of new quantum states. However, extending such moir\'e engineering to one-dimensional (1D) vdW systems has remained challenging. Here we report the moir\'e-engineered electronic insulating states in a new 1D moir\'e superlattice, by crystallographically aligning an armchair single-walled carbon nanotube (SWNT) to 2D hexagonal boron nitride (hBN) substrate. Remarkably, we observe the emergence of pronounced insulating states at charge neutrality point (CNP), full and half moir\'e fillings in lattice-aligned armchair SWNT/hBN heterostructures by low-temperature electrical transport measurements. In strong contrast, armchair SWNT devices without hBN alignment do not show any of these insulating behaviors, providing compelling evidence for the significant 1D moir\'e effect. Our density functional theory (DFT) and tight-binding calculations reveal that synergetic nanotube partial flattening and in-plane lattice reconstruction at 1D moir\'e interface expand the most stable AB' stacking regions (carbon on top of boron) and open sizable band gaps at both CNP and full moir\'e fillings at the single-particle level. Our one-body theory predicts no band gaps at half moir\'e fillings, suggesting that electron correlation and/or electron-phonon interaction may give rise to these emergent insulating behaviors in our 1D moir\'e systems. Our work establishes a new and definite moir\'e engineering route for 1D vdW materials and opens an exciting avenue for exploring interaction-induced quantum phases in 1D.

cond-mat.mes-hall↗

arXiv2025

Domain Wall Sliding-induced Polarization Switching in Multilayer Graphene

Electric polarization and metallicity are long believed not to coexist until the emergence of exceptionally rare material examples including the bulk polar metals and more recently two-dimensional (2D) van der Waals (vdW) materials such as 1T' WTe2. The electric polarization for the latter represents a new and distinguishable paradigm in materials science and physics because its electric polarization states embedded in the conduction electron sea are able to couple with (and controlled by) the external electric field. However, the microscopic polarization switching process and mechanism in these 2D vdW metallic materials have not been experimentally observed and remain elusive. Here, we report the first direct experimental imaging of the microscopic mechanism behind electric-field-coupled polarization switching in a metallic system. Our gate-tunable nanoscale optical imaging identifies the robust coexistence of electric polarization and appreciable carrier densities in adjacent polar domains hosting opposite electric polarizations in a 2D elemental metallic material, tetralayer graphene. We directly visualize and verify that the sliding domain wall (DW) solitons confined at the middlemost interface are responsible for the polarization switching in tetralayer graphene upon the application of electric fields and mechanical forces. Our work provides the first direct visualization of domain wall sliding-induced polarization switching in 2D elemental carbon at room temperature, significantly expanding and advancing the research of "ferroelectric metal" initially dubbed by P. W. Anderson and coauthors in 1965.

cond-mat.mes-hall↗

arXiv2025

Metallic NbS2 one-dimensional van der Waals heterostructures

This study presents the experimental realization of metallic NbS2-based one-dimensional van der Waals heterostructures applying a modified NaCl-assisted chemical vapor deposition approach. By employing a "remote salt" strategy, precise control over NaCl supply was achieved, enabling the growth of high-quality coaxial NbS2 nanotubes on single-walled carbon nanotube-boron nitride nanotube (SWCNT-BNNT) templates. With the remote salt strategy, the morphologies of as synthesized NbS2 could be controlled from 1D nanotubes to suspended 2D flakes. Structural characterization via high-resolution transmission electron microscopy (HRTEM) and scanning transmission electron microscopy (STEM) confirms the formation of crystalline NbS2 nanotubes, revealing a distinct bi-layer preference compared to monolayer-dominated semiconducting transition metal dichalcogenide analogs. Optical analyses using UV-vis-NIR and FTIR spectroscopy highlight the metallic nature of NbS2. With Raman analysis, oxidation studies demonstrate relative higher degradation rate of 1D NbS2 under ambient conditions. Density functional theory (DFT) calculations further elucidate the stabilization mechanism of bi-layer NbS2 nanotubes, emphasizing interlayer charge transfer and Coulomb interactions. This work establishes a robust framework for synthesizing metallic 1D vdW heterostructures, advancing their potential applications in optoelectronics and nanodevices.

cond-mat.mtrl-sci↗

arXiv2025

Strongly hybridized phonons in one-dimensional van der Waals crystals

The phenomena of pronounced electron-electron and electron-phonon interactions in one-dimensional (1D) systems are ubiquitous, which are well described by frameworks of Luttinger liquid, Peierls instability and concomitant charge density wave. However, the experimental observation of strongly hybridized phonons in 1D was not demonstrated. Herein we report the first observation of strongly hybridized phonons in 1D condensed matters by using double-walled carbon nanotubes (DWNTs), representative 1D van der Waals crystals, with combining the spectroscopic and microscopic tools as well as the ab initio density functional theory (DFT) calculations. We observe uncharted phonon modes in one commensurate and three incommensurate DWNT crystals, three of which concurrently exhibit strongly-reconstructed electronic band structures. Our DFT calculations for the experimentally observed commensurate DWNT (7,7) @ (12,12) reveal that this new phonon mode originates from a (nearly) degenerate coupling between two transverse acoustic modes (ZA modes) of constituent inner and outer nanotubes having approximately trigonal and pentagonal rotational symmetry along the nanotube circumferences. Such coupling strongly hybridizes the two phonon modes in different shells and leads to the formation of a unique lattice motion featuring evenly distributed vibrational amplitudes over inner and outer nanotubes, distinct from any known phonon modes in 1D systems. All four DWNTs that exhibit the pronounced new phonon modes show small chiral angle twists, closely matched diameter ratios of 3/5 and decreased frequencies of new phonon modes with increased diameters, all supporting the uncovered coupling mechanism. Our discovery of strongly hybridized phonons in DWNTs open new opportunities for engineering phonons and exploring novel phonon-related phenomena in 1D condensed matters.

cond-mat.mes-hall↗

arXiv2024

OBSeg: Accurate and Fast Instance Segmentation Framework Using Segmentation Foundation Models with Oriented Bounding Box Prompts

Instance segmentation in remote sensing images is a long-standing challenge. Since horizontal bounding boxes introduce many interference objects, oriented bounding boxes (OBBs) are usually used for instance identification. However, based on ``segmentation within bounding box'' paradigm, current instance segmentation methods using OBBs are overly dependent on bounding box detection performance. To tackle this problem, this paper proposes OBSeg, an accurate and fast instance segmentation framework using OBBs. OBSeg is based on box prompt-based segmentation foundation models (BSMs), e.g., Segment Anything Model. Specifically, OBSeg first detects OBBs to distinguish instances and provide coarse localization information. Then, it predicts OBB prompt-related masks for fine segmentation. Since OBBs only serve as prompts, OBSeg alleviates the over-dependence on bounding box detection performance of current instance segmentation methods using OBBs. Thanks to OBB prompts, OBSeg outperforms other current BSM-based methods using HBBs. In addition, to enable BSMs to handle OBB prompts, we propose a novel OBB prompt encoder. To make OBSeg more lightweight and further improve the performance of lightweight distilled BSMs, a Gaussian smoothing-based knowledge distillation method is introduced. Experiments demonstrate that OBSeg outperforms current instance segmentation methods on multiple datasets in terms of instance segmentation accuracy and has competitive inference speed. The code is available at https://github.com/zhen6618/OBBInstanceSegmentation.

cs.CV↗

arXiv2023

ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs

Despite the advancements of open-source large language models (LLMs), e.g., LLaMA, they remain significantly limited in tool-use capabilities, i.e., using external tools (APIs) to fulfill human instructions. The reason is that current instruction tuning largely focuses on basic language tasks but ignores the tool-use domain. This is in contrast to the excellent tool-use capabilities of state-of-the-art (SOTA) closed-source LLMs, e.g., ChatGPT. To bridge this gap, we introduce ToolLLM, a general tool-use framework encompassing data construction, model training, and evaluation. We first present ToolBench, an instruction-tuning dataset for tool use, which is constructed automatically using ChatGPT. Specifically, the construction can be divided into three stages: (i) API collection: we collect 16,464 real-world RESTful APIs spanning 49 categories from RapidAPI Hub; (ii) instruction generation: we prompt ChatGPT to generate diverse instructions involving these APIs, covering both single-tool and multi-tool scenarios; (iii) solution path annotation: we use ChatGPT to search for a valid solution path (chain of API calls) for each instruction. To enhance the reasoning capabilities of LLMs, we develop a novel depth-first search-based decision tree algorithm. It enables LLMs to evaluate multiple reasoning traces and expand the search space. Moreover, to evaluate the tool-use capabilities of LLMs, we develop an automatic evaluator: ToolEval. Based on ToolBench, we fine-tune LLaMA to obtain an LLM ToolLLaMA, and equip it with a neural API retriever to recommend appropriate APIs for each instruction. Experiments show that ToolLLaMA demonstrates a remarkable ability to execute complex instructions and generalize to unseen APIs, and exhibits comparable performance to ChatGPT. Our ToolLLaMA also demonstrates strong zero-shot generalization ability in an out-of-distribution tool-use dataset: APIBench.

cs.AI↗

arXiv2022

Nature of novel moiré exciton states in WSe$_2$/WS$_2$ heterobilayers

Moiré patterns of transition metal dichalcogenide (TMD) heterobilayers have proven to be an ideal platform to host unusual correlated electronic phases, emerging magnetism, and correlated exciton physics. While the existence of novel moiré excitonic states is established through optical measurements, the microscopic nature of these states is still poorly understood, often relying on empirically fit models. Here, combining large-scale first-principles GW-BSE calculations and micro-reflection spectroscopy, we identify the nature of the exciton resonances in WSe$_2$/WS$_2$ moiré superlattices, discovering a surprisingly rich set of moiré excitons that cannot be even qualitatively captured by prevailing continuum models. Our calculations reveal moiré excitons with distinct characters, including modulated Wannier excitons and previously unindentified intralayer charge-transfer excitons. Signatures of these distinct excitonic characters are confirmed experimentally via the unique carrier-density and magnetic-field dependences of different moiré exciton resonances. Our study highlights the highly non-trivial exciton states that can emerge in TMD moiré superlattices, and suggests novel ways of tuning many-body physics in moiré systems by engineering excited-states with specific spatial characters.

cond-mat.mes-hall↗

arXiv2021

Interlayer interactions in one-dimensional van der Waals moiré superlattices

Different atomistic registry between the layers forming the inner and outer nanotubes can form one-dimensional (1D) van der Waals (vdW) moiré superlattices. Unlike the two-dimensional (2D) vdW moiré superlattices, effects of 1D vdW moiré superlattices on electronic and optical properties in 1D moiré superlattices are not well understood, and they are often neglected. In this Perspective, we summarize new experimental observations and theoretical perspectives related to interlayer interactions in double-walled carbon nanotubes (DWNTs), a representative 1D vdW moiré system. Our discussion will focus on new optical features emerging from the interlayer electronic interactions in DWNTs. Exciting correlated physics and exotic phases of matter are anticipated to exist in 1D vdW moiré superlattices, analogous with those discovered in the 2D vdW moiré superlattices. We further discuss the future directions in probing and uncovering interesting physical phenomena in 1D moiré superlattices.

cond-mat.mes-hall↗

arXiv2021

Efficient Fizeau Drag from Dirac electrons in monolayer graphene

Fizeau demonstrated in 1850 that the speed of light can be modified when it is propagating in moving media. Can we achieve such control of the light speed efficiently with a fast-moving electron media by passing electrical current? Because the strong electromagnetic coupling between the electron and light leads to the collective excitation of plasmon polaritons, it will manifest as the plasmonic Doppler effect. Experimental observation of the plasmonic Doppler effect in electronic system has been challenge because the plasmon propagation speed is much faster than the electron drift velocity in conventional noble metals. Here, we report direct observation of Fizeau drag of plasmon polaritons in strongly biased graphene by exploiting the high electron mobility and the slow plasmon propagation of massless Dirac electrons. The large bias current in graphene creates a fast drifting Dirac electron medium hosting the plasmon polariton. It results in nonreciprocal plasmon propagation, where plasmons moving with the drifting electron media propagate at an enhanced speed. We measure the Doppler-shifted plasmon wavelength using a cryogenic near-field infrared nanoscopy, which directly images the plasmon polariton mode in the biased graphene at low temperature. We observe a plasmon wavelength difference up to 3.6% between plasmon moving along and against the drifting electron media. Our findings on the plasmonic Doppler effect open new opportunities for electrical control of non-reciprocal surface plasmon polaritons in nonequilibrium systems.

physics.optics↗

arXiv2021

Correlation-Driven Electron-Hole Asymmetry in Graphene Field Effect Devices

Electron-hole asymmetry is a fundamental property in solids that can determine the nature of quantum phase transitions and the regime of operation for devices. The observation of electron-hole asymmetry in graphene and recently in the phase diagram of bilayer graphene has spurred interest into whether it stems from disorder or from fundamental interactions such as correlations. Here, we report an effective new way to access electron-hole asymmetry in 2D materials by directly measuring the quasiparticle self-energy in graphene/Boron Nitride field effect devices. As the chemical potential moves from the hole to the electron doped side, we see an increased strength of electronic correlations manifested by an increase in the band velocity and inverse quasiparticle lifetime. These results suggest that electronic correlations play an intrinsic role in driving electron hole asymmetry in graphene and provide a new insight for asymmetries in more strongly correlated materials.

cond-mat.mes-hall↗

arXiv2020

Imaging moiré flat bands in 3D reconstructed WSe2/WS2 superlattices

Moiré superlattices in transition metal dichalcogenide (TMD) heterostructures can host novel correlated quantum phenomena due to the interplay of narrow moiré flat bands and strong, long-range Coulomb interactions1-5. However, microscopic knowledge of the atomically-reconstructed moiré superlattice and resulting flat bands is still lacking, which is critical for fundamental understanding and control of the correlated moiré phenomena. Here we quantitatively study the moiré flat bands in three-dimensional (3D) reconstructed WSe2/WS2 moiré superlattices by comparing scanning tunneling spectroscopy (STS) of high quality exfoliated TMD heterostructure devices with ab initio simulations of TMD moiré superlattices. A strong 3D buckling reconstruction accompanied by large in-plane strain redistribution is identified in our WSe2/WS2 moiré heterostructures. STS imaging demonstrates that this results in a remarkably narrow and highly localized K-point moiré flat band at the valence band edge of the heterostructure. A series of moiré flat bands are observed at different energies that exhibit varying degrees of localization. Our observations contradict previous simplified theoretical models but agree quantitatively with ab initio simulations that fully capture the 3D structural reconstruction. Here the strain redistribution and 3D buckling dominate the effective moiré potential and result in moiré flat bands at the Brillouin zone K points.

cond-mat.mes-hall↗

arXiv2019

Visualization of the flat electronic band in twisted bilayer graphene near the magic angle twist

Bilayer graphene was theorized to host a moire miniband with flat dispersion if the layers are stacked at specific twist angles known as the magic angles. Recently, such twisted bilayer graphene (tBLG) with the first magic angle twist was reported to exhibit correlated insulating state and superconductivity, where the presence of the flat miniband in the system is thought to be essential for the emergence of these ordered phases in the transport measurements. Tunneling spectroscopy and electronic compressibility measurements in tBLG have revealed a van Hove singularity that is consistent with the presence of the flat miniband. However, a direct observation of the flat dispersion in the momentum-space of such moire miniband in tBLG is still elusive. Here, we report the visualization of the flat moire miniband by using angle-resolved photoemission spectroscopy with nanoscale resolution (nanoARPES). The high spatial resolution in nanoARPES enabled the measurement of the local electronic structure of the tBLG. We clearly demonstrate the existence of the flat moire band near the charge neutrality for tBLG close to the magic angle at room temperature.

cond-mat.mes-hall↗

arXiv2019

Optical detection of Mott and generalized Wigner crystal states in WSe2/WS2 moiré superlattices

Moiré superlattices are emerging as a new route for engineering strongly correlated electronic states in two-dimensional van der Waals heterostructures, as recently demonstrated in the correlated insulating and superconducting states in magic-angle twisted bilayer graphene and ABC trilayer graphene/boron nitride moiré superlattices. Transition metal dichalcogenide (TMDC) moiré heterostructures provide another exciting model system to explore correlated quantum phenomena, with the addition of strong light-matter interactions and large spin-orbital coupling. Here we report the optical detection of strongly correlated phases in semiconducting WSe2/WS2 moiré superlattices. Our sensitive optical detection technique reveals a Mott insulator state at one hole per superlattice site (ν = 1), and surprising insulating phases at fractional filling factors ν = 1/3 and 2/3, which we assign to generalized Wigner crystallization on an underlying lattice. Furthermore, the unique spin-valley optical selection rules of TMDC heterostructures allow us to optically create and investigate low-energy spin excited states in the Mott insulator. We reveal an especially slow spin relaxation lifetime of many microseconds in the Mott insulating state, orders-of-magnitude longer than that of charge excitations. Our studies highlight novel correlated physics that can emerge in moiré superlattices beyond graphene.

cond-mat.mes-hall↗