Zhenpeng Su
YOU?
Author Swipe
View article: Entropy Ratio Clipping as a Soft Global Constraint for Stable Reinforcement Learning
Entropy Ratio Clipping as a Soft Global Constraint for Stable Reinforcement Learning Open
Large language model post-training relies on reinforcement learning to improve model capability and alignment quality. However, the off-policy training paradigm introduces distribution shift, which often pushes the policy beyond the trust …
View article: Wave Heating of Magnetotail Current Sheet Electrons at Mars
Wave Heating of Magnetotail Current Sheet Electrons at Mars Open
The Martian magnetotail current sheet serves as a critical pathway for ionospheric ion escape. Contrary to the conventional view that external magnetic pressure is balanced mainly by internal ion thermal pressure, we present novel observat…
View article: Three-step Acceleration of the Radiation Belt Relativistic Electrons by Interplanetary Shocks
Three-step Acceleration of the Radiation Belt Relativistic Electrons by Interplanetary Shocks Open
Inward radial diffusion driven by ultra-low-frequency (ULF) waves is one of the dominant acceleration mechanisms of relativistic (>1.0 MeV) electrons in the Earth’s outer radiation belt. However, the role of interplanetary (IP) shocks in t…
View article: Electromagnetic Ion Cyclotron Waves in a Magnetic Reconnection Exhaust at Earth's Magnetopause
Electromagnetic Ion Cyclotron Waves in a Magnetic Reconnection Exhaust at Earth's Magnetopause Open
Plasma waves can initiate, regulate, or reflect magnetic reconnection efficiently converting magnetic energy into plasma energy. While waves ranging from below the ion cyclotron frequency to above the electron plasma frequency are commonly…
View article: Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization
Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization Open
We present Klear-Reasoner, a model with long reasoning capabilities that demonstrates careful deliberation during problem solving, achieving outstanding performance across multiple benchmarks. Although there are already many excellent work…
View article: Statistical Study on the Solar Wind Turbulence Spectra Upstream of Mars
Statistical Study on the Solar Wind Turbulence Spectra Upstream of Mars Open
We statistically analyze the power spectral density (PSD) of magnetic field turbulence in the upstream solar wind of the Martian bow shock by investigating the data from Tianwen-1 and Mars Atmosphere and Volatile Evolution (MAVEN) during 2…
View article: Statistical Study on the Solar Wind Turbulence Spectra upstream of Mars
Statistical Study on the Solar Wind Turbulence Spectra upstream of Mars Open
We statistically analyze the power spectral density (PSD) of magnetic field turbulence in the upstream solar wind of the Martian bow shock by investigating the data from Tianwen-1 and MAVEN during November 13 and December 31, 2021. The spe…
View article: LightRetriever: A LLM-based Text Retrieval Architecture with Extremely Faster Query Inference
LightRetriever: A LLM-based Text Retrieval Architecture with Extremely Faster Query Inference Open
Large Language Models (LLMs)-based text retrieval retrieves documents relevant to search queries based on vector similarities. Documents are pre-encoded offline, while queries arrive in real-time, necessitating an efficient online query en…
View article: Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval
Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval Open
Large Language Model-based Dense Retrieval (LLM-DR) optimizes over numerous heterogeneous fine-tuning collections from different domains. However, the discussion about its training data distribution is still minimal. Previous studies rely …
View article: Scaffold-BPE: Enhancing Byte Pair Encoding for Large Language Models with Simple and Effective Scaffold Token Removal
Scaffold-BPE: Enhancing Byte Pair Encoding for Large Language Models with Simple and Effective Scaffold Token Removal Open
Byte Pair Encoding (BPE) serves as a foundation method for text tokenization in the Natural Language Processing (NLP) field. Despite its wide adoption, the original BPE algorithm harbors an inherent flaw: it inadvertently introduces a freq…
View article: An Efficient Positivity-Preserving Finite Difference Scheme for Solving the Fokker-Planck Diffusion Equation
An Efficient Positivity-Preserving Finite Difference Scheme for Solving the Fokker-Planck Diffusion Equation Open
The Fokker-Planck diffusion equation is widely used for simulating the evolution of Earth's radiation belt electrons, which pose significant hazards to space-borne systems. To preserve the positivity of the numerical solution of the electr…
View article: Inner Magnetospheric Oxygen Torus Induced by Electromagnetic Ion Cyclotron Waves
Inner Magnetospheric Oxygen Torus Induced by Electromagnetic Ion Cyclotron Waves Open
Cold oxygen ions escaping from the ionosphere and temporarily trapped near the plasmapause form the oxygen torus. Mass‐loading by these oxygen ions significantly affects magnetospheric plasma processes. However, due to the technical challe…
View article: DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs
DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs Open
As large language models continue to scale, computational costs and resource consumption have emerged as significant challenges. While existing sparsification methods like pruning reduce computational overhead, they risk losing model knowl…
View article: UniAttn: Reducing Inference Costs via Softmax Unification for Post-Training LLMs
UniAttn: Reducing Inference Costs via Softmax Unification for Post-Training LLMs Open
Post-training is essential for adapting Large Language Models (LLMs) to real-world applications. Deploying post-trained models faces significant challenges due to substantial memory overhead and noticeable inference latency. Existing work …
View article: CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts
CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts Open
View article: DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs
DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs Open
View article: Temporal Scaling Law for Large Language Models
Temporal Scaling Law for Large Language Models Open
View article: CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts
CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts Open
Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its…
View article: Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval
Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval Open
Large Language Model-based Dense Retrieval (LLM-DR) optimizes over numerous heterogeneous fine-tuning collections from different domains. However, the discussion about its training data distribution is still minimal. Previous studies rely …
View article: MaskMoE: Boosting Token-Level Learning via Routing Mask in Mixture-of-Experts
MaskMoE: Boosting Token-Level Learning via Routing Mask in Mixture-of-Experts Open
Scaling the size of a model enhances its capabilities but significantly increases computation complexity. Mixture-of-Experts models (MoE) address the issue by allowing model size to scale up without substantially increasing training or inf…
View article: Scaffold-BPE: Enhancing Byte Pair Encoding for Large Language Models with Simple and Effective Scaffold Token Removal
Scaffold-BPE: Enhancing Byte Pair Encoding for Large Language Models with Simple and Effective Scaffold Token Removal Open
Byte Pair Encoding (BPE) serves as a foundation method for text tokenization in the Natural Language Processing (NLP) field. Despite its wide adoption, the original BPE algorithm harbors an inherent flaw: it inadvertently introduces a freq…
View article: Long Lifetime Hiss Rays in the Disturbed Plasmasphere
Long Lifetime Hiss Rays in the Disturbed Plasmasphere Open
Plasmaspheric hiss waves are important to shape the Earth’s electron radiation belt. These waves are commonly envisioned to have a long lifetime which allows them to permeate the global plasmasphere from a spatially restricted source. Howe…
View article: Interplanetary shock induced intensification of electron cyclotron harmonic waves in the Earth’s inner magnetosphere
Interplanetary shock induced intensification of electron cyclotron harmonic waves in the Earth’s inner magnetosphere Open
Electron cyclotron harmonic (ECH) waves are electrostatic emissions frequently observed in the Earth’s magnetosphere. By precipitating magnetospheric hot electrons into the ionosphere, ECH waves play a critical role in the formation of dif…
View article: A Substorm‐Dependent Negative Limit of Non‐Eclipse Surface Charging of a Chinese Geosynchronous Satellite
A Substorm‐Dependent Negative Limit of Non‐Eclipse Surface Charging of a Chinese Geosynchronous Satellite Open
Surface charging is one of the most common causes of spacecraft anomalies. When and to what potential the spacecraft is charged are two important questions in space weather. Here, for a Chinese geosynchronous navigation satellite, we infer…
View article: Inferring Whistler‐Mode Chorus Wave Source Regions in the Martian Mini‐Magnetospheres
Inferring Whistler‐Mode Chorus Wave Source Regions in the Martian Mini‐Magnetospheres Open
Martian mini‐magnetospheres contain whistler‐mode chorus waves potentially contributing to atmospheric escape, analogous to the Earth's inner magnetosphere. At Earth, the chorus waves have been found to originate from the near‐equatorial r…
View article: A substorm-dependent negative limit of non-eclipse surface charging of a Chinese geosynchronous satellite
A substorm-dependent negative limit of non-eclipse surface charging of a Chinese geosynchronous satellite Open
Surface charging is one of the most common causes of spacecraft anomalies. When and to what potential the spacecraft is charged are two important questions in space weather. Here, for a Chinese geosynchronous navigation satellite, we infer…
View article: MiLe Loss: a New Loss for Mitigating the Bias of Learning Difficulties in Generative Language Models
MiLe Loss: a New Loss for Mitigating the Bias of Learning Difficulties in Generative Language Models Open
Generative language models are usually pretrained on large text corpus via predicting the next token (i.e., sub-word/word/phrase) given the previous ones. Recent works have demonstrated the impressive performance of large generative langua…
View article: HC3 Plus: A Semantic-Invariant Human ChatGPT Comparison Corpus
HC3 Plus: A Semantic-Invariant Human ChatGPT Comparison Corpus Open
ChatGPT has garnered significant interest due to its impressive performance; however, there is growing concern about its potential risks, particularly in the detection of AI-generated content (AIGC), which is often challenging for untraine…
View article: Martian Bow Shock Oscillations Driven by Solar Wind Variations: Simultaneous Observations From Tianwen‐1 and MAVEN
Martian Bow Shock Oscillations Driven by Solar Wind Variations: Simultaneous Observations From Tianwen‐1 and MAVEN Open
The Martian bow shock stands as the first defense against the solar wind and shapes the Martian magnetosphere. Previous studies showed the correlation between the Martian bow shock location and solar wind parameters. Here we present direct…
View article: Plasmaspheric High‐Frequency Whistlers as a Candidate Cause of Shock Aurora at Earth
Plasmaspheric High‐Frequency Whistlers as a Candidate Cause of Shock Aurora at Earth Open
Auroral brightening driven by interplanetary shocks on Earth's closed magnetic field lines is commonly attributed to the 0.1–10 keV electron precipitations by electron cyclotron harmonic waves and whistler‐mode chorus waves in the low‐dens…