Searcharxiv⌕ Search

arXiv subjects

Yuhan Sun

Publications and source records attributed to Yuhan Sun.

30 records · Page 2Linked to original sources
arXiv2025

X-ray imaging of antiferromagnetic octupole domains in Mn$_3$Sn

Novel antiferromagnets with broken time reversal symmetry (TRS) have launched a new direction in spintronics research, combining the advantageous dynamical properties of conventional antiferromagnets with the controllability typically associated with ferromagnets. However, antiferromagnetic domains are notoriously challenging to image in real-space. X-ray magnetic circular dichroism (XMCD) offers a route to overcome this difficulty: XMCD contrast may be finite in TRS-breaking antiferromagnets with an appropriate magnetic space group. Here, we exploit this to image the octupole domains in a focused ion beam-fabricated device of the non-collinear antiferromagnet Mn$_3$Sn. Using scanning transmission x-ray microscopy, we spatially resolve the weak pre-edge XMCD contrast (of 0.2%) that is sensitive to $T_z$, achieving a contrast resolution better than 0.02%. We observe hysteretic switching of the octupole order through both the XMCD contrast and the corresponding anomalous Hall effect within the same device. These results confirm the bulk nature of this contrast, and establish XMCD-based microscopy as a powerful real space imaging method for TRS-breaking antiferromagnets, including altermagnets, enabling future studies of their dynamics, switching, and symmetry-tunable phenomena.

cond-mat.mtrl-sci↗
arXiv2025

LiveThinking: Enabling Real-Time Efficient Reasoning for AI-Powered Livestreaming via Reinforcement Learning

In AI-powered e-commerce livestreaming, digital avatars require real-time responses to drive engagement, a task for which high-latency Large Reasoning Models (LRMs) are ill-suited. We introduce LiveThinking, a practical two-stage optimization framework to bridge this gap. First, we address computational cost by distilling a 670B teacher LRM into a lightweight 30B Mixture-of-Experts (MoE) model (3B active) using Rejection Sampling Fine-Tuning (RFT). This reduces deployment overhead but preserves the teacher's verbose reasoning, causing latency. To solve this, our second stage employs reinforcement learning with Group Relative Policy Optimization (GRPO) to compress the model's reasoning path, guided by a multi-objective reward function balancing correctness, helpfulness, and brevity. LiveThinking achieves a 30-fold reduction in computational cost, enabling sub-second latency. In real-world application on Taobao Live, it improved response correctness by 3.3% and helpfulness by 21.8%. Tested by hundreds of thousands of viewers, our system led to a statistically significant increase in Gross Merchandise Volume (GMV), demonstrating its effectiveness in enhancing user experience and commercial performance in live, interactive settings.

cs.LG↗
arXiv2025

Spectral diameter of negatively monotone manifolds

For a closed negatively monotone symplectic manifold, we construct quasi-isometric embeddings from the Euclidean spaces to its Hamiltonian diffeomorphism group, assuming it contains an incompressible heavy Lagrangian. We also show the super-heaviness of its skeleton with respect to a Donaldson hypersurface.

math.SG↗
arXiv2025

Technical Report of TeleChat2, TeleChat2.5 and T1

We introduce the latest series of TeleChat models: \textbf{TeleChat2}, \textbf{TeleChat2.5}, and \textbf{T1}, offering a significant upgrade over their predecessor, TeleChat. Despite minimal changes to the model architecture, the new series achieves substantial performance gains through enhanced training strategies in both pre-training and post-training stages. The series begins with \textbf{TeleChat2}, which undergoes pretraining on 10 trillion high-quality and diverse tokens. This is followed by Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to further enhance its capabilities. \textbf{TeleChat2.5} and \textbf{T1} expand the pipeline by incorporating a continual pretraining phase with domain-specific datasets, combined with reinforcement learning (RL) to improve performance in code generation and mathematical reasoning tasks. The \textbf{T1} variant is designed for complex reasoning, supporting long Chain-of-Thought (CoT) reasoning and demonstrating substantial improvements in mathematics and coding. In contrast, \textbf{TeleChat2.5} prioritizes speed, delivering rapid inference. Both flagship models of \textbf{T1} and \textbf{TeleChat2.5} are dense Transformer-based architectures with 115B parameters, showcasing significant advancements in reasoning and general task performance compared to the original TeleChat. Notably, \textbf{T1-115B} outperform proprietary models such as OpenAI's o1-mini and GPT-4o. We publicly release \textbf{TeleChat2}, \textbf{TeleChat2.5} and \textbf{T1}, including post-trained versions with 35B and 115B parameters, to empower developers and researchers with state-of-the-art language models tailored for diverse applications.

cs.CL↗
arXiv2025

Index bounded relative symplectic cohomology

We study the relative symplectic cohomology with the help of an index bounded contact form. For a Liouville domain with an index bounded boundary, we construct a spectral sequence which starts from its classical symplectic cohomology and converges to the relative symplectic cohomology of it inside a Calabi-Yau manifold.

math.SG↗
arXiv2024

Heavy sets and index bounded relative symplectic cohomology

We use relative symplectic cohomology to detect heavy sets, with the help of index bounded contact forms. This establishes a relation between two notions SH-heaviness and heaviness, which partly answers a conjecture of Dickstein-Ganor-Polterovich-Zapolsky in the symplectically aspherical setting.

math.SG↗
arXiv2024

Displacement energy of Lagrangian 3-spheres

We estimate the displacement energy of Lagrangian 3-spheres in a symplectic 6-manifold $X$, by estimating the displacement energy of a one-parameter family $L_λ$ of Lagrangian tori near the sphere. The proof establishes a new version of Lagrangian Floer theory with cylinder corrections, which is motivated by the change of open Gromov-Witten invariants under the conifold transition. We also make observations and computations on the classical Floer theory by using symplectic sum formula and Welschinger invariants.

math.SG↗
arXiv2024

A characterization of heaviness in terms of relative symplectic cohomology

For a compact subset $K$ of a closed symplectic manifold $(M, ω)$, we prove that $K$ is heavy if and only if its relative symplectic cohomology over the Novikov field is non-zero. As an application we show that if two compact sets are not heavy and Poisson commuting, then their union is also not heavy. A discussion on superheaviness together with some partial results are also included.

math.SG↗
arXiv2023

To be or not to be? an exploration of continuously controllable prompt engineering

As the use of large language models becomes more widespread, techniques like parameter-efficient fine-tuning and other methods for controlled generation are gaining traction for customizing models and managing their outputs. However, the challenge of precisely controlling how prompts influence these models is an area ripe for further investigation. In response, we introduce ControlPE (Continuously Controllable Prompt Engineering). ControlPE enables finer adjustments to prompt effects, complementing existing prompt engineering, and effectively controls continuous targets. This approach harnesses the power of LoRA (Low-Rank Adaptation) to create an effect akin to prompt weighting, enabling fine-tuned adjustments to the impact of prompts. Our methodology involves generating specialized datasets for prompt distillation, incorporating these prompts into the LoRA model, and carefully adjusting LoRA merging weight to regulate the influence of prompts. This provides a dynamic and adaptable tool for prompt control. Through our experiments, we have validated the practicality and efficacy of ControlPE. It proves to be a promising solution for control a variety of prompts, ranging from generating short responses prompts, refusal prompts to chain-of-thought prompts.

cs.CL↗
arXiv2018

$A_{n}$-type surface singularity and nondisplaceable Lagrangian tori

We prove the existence of a one-parameter family of nondisplaceable Lagrangian tori near a linear chain of Lagrangian 2-spheres in a symplectic 4-manifold. When the symplectic structure is rational we prove that the deformed Floer cohomology groups of these tori are nontrivial. The proof uses the idea of toric degeneration to analyze the full potential functions with bulk deformations of these tori.

math.SG↗
arXiv2016

GeoReach: An Efficient Approach for Evaluating Graph Reachability Queries with Spatial Range Predicates

Graphs are widely used to model data in many application domains. Thanks to the wide spread use of GPS-enabled devices, many applications assign a spatial attribute to graph vertices (e.g., geo-tagged social media). Users may issue a Reachability Query with Spatial Range Predicate (abbr. RangeReach). RangeReach finds whether an input vertex can reach any spatial vertex that lies within an input spatial range. An example of a RangeReach query is: Given a social graph, find whether Alice can reach any of the venues located within the geographical area of Arizona State University. The paper proposes GeoReach an approach that adds spatial data awareness to a graph database management system (GDBMS). GeoReach allows efficient execution of RangeReach queries, yet without compromising a lot on the overall system scalability (measured in terms of storage size and initialization/maintenance time). To achieve that, GeoReach is equipped with a light-weight data structure, namely SPA-Graph, that augments the underlying graph data with spatial indexing directories. When a RangeReach query is issued, the system employs a pruned-graph traversal approach. Experiments based on real system implementation inside Neo4j proves that GEOREACH exhibits up to two orders of magnitude better query response time and up to four times less storage than the state-of-the-art spatial and reachability indexing approaches.

cs.DB↗