arXiv subjects

Xiaoqing Sun

Publications and source records attributed to Xiaoqing Sun.

16 recordsLinked to original sources

arXiv2026

Getting the Parameters Right: A Difficulty-Graded Benchmark and Probe-Guided Training for LLM Tool Calls

Large language model agents derive much of their capability from tool use. Existing research on tool use has largely focused on selecting the right tool and orchestrating the order of calls. However, correctly filling the parameters of a tool call is equally critical for successful execution and has received far less attention. In domains such as cloud networking, even frontier models correctly complete fewer than half of tool calls. Inspired by recent analyses showing that LLM hidden states encode rich information about model predictions, we discover that while the model generates a parameter value, its hidden state contains a strong correctness signal: a simple linear probe can accurately predict whether the value will be correct. Based on this observation, we propose a unified probe-guided framework with two complementary approaches: probe-filtered bootstrapped training (PBT), which uses the probe to filter reliable self-generated calls for fine-tuning, and probe-guided reranking (PGR), which uses the probe to select better candidates during inference. To support systematic evaluation, we release ParamBench, a benchmark built from real cloud-network APIs that categorizes every instance into five difficulty levels according to parameter nesting depth, cross-parameter dependencies, and the reasoning required to derive values from earlier calls. Extensive experiments across 5 open models on ParamBench and 6 external benchmarks demonstrate that our method substantially improves parameter generation, raising the average exact match from 19.7% to 59.6%.

cs.AI↗

arXiv2026

Scalable LLM Agent Tool Access in the Cloud

LLM agents increasingly rely on tool calling to act on external systems, and the Model Context Protocol (MCP) has quickly become its de facto interface. Operating MCP at cloud scale, however, becomes difficult. On the tool provider side, legacy services are not directly callable through MCP; the rapid protocol development also creates ongoing compatibility cost. On the agent side, the number of accessible tool is limited by the LLM context window and inference overhead; mounting a large tool set increases token usage and inference latency and can reduce task success rate. Moreover, for stateful MCP backends with multiple replicas, preserving session affinity increases client-side complexity. We present a cloud-scale gateway system for MCP service. It breaks the direct-connect model on the data plane and offloads legacy service integration, consolidating incompatible MCP variants, access control, tool recommendation, and session-aware routing to the gateway. Hybrid retrieval sustains 98% Top-15 recall; it scales agent tool access to 3,000+ with high tool selection accuracy, and reduces tool selection time by $8.9\times$ and token usage by $23.8\times$, with low per-call overhead, stable under scale-out. Finally, we share the lessons learned from deploying the gateway system in production.

cs.DC↗

arXiv2025

Interpretable Embeddings with Sparse Autoencoders: A Data Analysis Toolkit

Analyzing large-scale text corpora is a core challenge in machine learning, crucial for tasks like identifying undesirable model behaviors or biases in training data. Current methods often rely on costly LLM-based techniques (e.g. annotating dataset differences) or dense embedding models (e.g. for clustering), which lack control over the properties of interest. We propose using sparse autoencoders (SAEs) to create SAE embeddings: representations whose dimensions map to interpretable concepts. Through four data analysis tasks, we show that SAE embeddings are more cost-effective and reliable than LLMs and more controllable than dense embeddings. Using the large hypothesis space of SAEs, we can uncover insights such as (1) semantic differences between datasets and (2) unexpected concept correlations in documents. For instance, by comparing model responses, we find that Grok-4 clarifies ambiguities more often than nine other frontier models. Relative to LLMs, SAE embeddings uncover bigger differences at 2-8x lower cost and identify biases more reliably. Additionally, SAE embeddings are controllable: by filtering concepts, we can (3) cluster documents along axes of interest and (4) outperform dense embeddings on property-based retrieval. Using SAE embeddings, we study model behavior with two case studies: investigating how OpenAI model behavior has changed over time and finding "trigger" phrases learned by Tulu-3 (Lambert et al., 2024) from its training data. These results position SAEs as a versatile tool for unstructured data analysis and highlight the neglected importance of interpreting models through their data.

cs.AI↗

arXiv2025

Investigation of the Physical Mechanism behind Retention Loss in FeFETs with MIFIFIS Gate Structure

A Metal-Gate Blocking Layer (GBL)- Ferroelectric-Tunnel Dielectric Layer (TDL)-Ferroelectric -Channel Insulator (Ch.IL)-Si (MIFIFIS) structure is proposed to achieve a larger MW for applications in Fe-NAND. However, the large retention loss (RL) in the MIFIFIS structure restricts its application. In this work, we vary the physical thickness of the GBL and TDL, and conduct an in-depth analysis of the energy bands of the gate structure to investigate the physical mechanism behind the RL in FeFETs with the MIFIFIS structure. The physical origin of the RL is that the electric field direction across the TDL reduces the potential barrier provided by the ferroelectric near the silicon substrate. Based on the above physical mechanism, the RL can be reduced to 12% and 0.2% by redesigning the gate structure or reducing the pulse amplitude, respectively. Our work contributes to a deeper understanding of the physical mechanism behind the RL in FeFETs with the MIFIFIS gate structure. It provides guidance for enhancing the reliability of FeFETs.

physics.app-ph↗

arXiv2025

Lumina-mGPT 2.0: Stand-Alone AutoRegressive Image Modeling

We present Lumina-mGPT 2.0, a stand-alone, decoder-only autoregressive model that revisits and revitalizes the autoregressive paradigm for high-quality image generation and beyond. Unlike existing approaches that rely on pretrained components or hybrid architectures, Lumina-mGPT 2.0 is trained entirely from scratch, enabling unrestricted architectural design and licensing freedom. It achieves generation quality on par with state-of-the-art diffusion models such as DALL-E 3 and SANA, while preserving the inherent flexibility and compositionality of autoregressive modeling. Our unified tokenization scheme allows the model to seamlessly handle a wide spectrum of tasks-including subject-driven generation, image editing, controllable synthesis, and dense prediction-within a single generative framework. To further boost usability, we incorporate efficient decoding strategies like inference-time scaling and speculative Jacobi sampling to improve quality and speed, respectively. Extensive evaluations on standard text-to-image benchmarks (e.g., GenEval, DPG) demonstrate that Lumina-mGPT 2.0 not only matches but in some cases surpasses diffusion-based models. Moreover, we confirm its multi-task capabilities on the Graph200K benchmark, with the native Lumina-mGPT 2.0 performing exceptionally well. These results position Lumina-mGPT 2.0 as a strong, flexible foundation model for unified multimodal generation. We have released our training details, code, and models at https://github.com/Alpha-VLLM/Lumina-mGPT-2.0.

cs.CV↗

arXiv2025

Dense SAE Latents Are Features, Not Bugs

Sparse autoencoders (SAEs) are designed to extract interpretable features from language models by enforcing a sparsity constraint. Ideally, training an SAE would yield latents that are both sparse and semantically meaningful. However, many SAE latents activate frequently (i.e., are \emph{dense}), raising concerns that they may be undesirable artifacts of the training procedure. In this work, we systematically investigate the geometry, function, and origin of dense latents and show that they are not only persistent but often reflect meaningful model representations. We first demonstrate that dense latents tend to form antipodal pairs that reconstruct specific directions in the residual stream, and that ablating their subspace suppresses the emergence of new dense features in retrained SAEs -- suggesting that high density features are an intrinsic property of the residual space. We then introduce a taxonomy of dense latents, identifying classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction. Finally, we analyze how these features evolve across layers, revealing a shift from structural features in early layers, to semantic features in mid layers, and finally to output-oriented signals in the last layers of the model. Our findings indicate that dense latents serve functional roles in language model computation and should not be dismissed as training noise.

cs.LG↗

arXiv2025

The effects of projection on measuring the splashback feature

The splashback radius $R_{\rm sp}$ is a boundary of a halo that separates infalling and accreted matter. This results in a steep drop in the density profile at $R_{\rm st}$, which is a commonly adopted proxy for $R_{\rm sp}$. Observationally, $R_{\rm st}$ can be measured through fitting the projected galaxy number density profile of the halo, but there has been some discrepancy between the observed and expected $R_{\rm st}$. Therefore, we investigate whether the projection of the density profile onto the plane of the sky could lead to any systematic bias in determining $R_{\rm st}$, by studying the true 3-dimensional and projected halo density profiles from the IllustrisTNG simulation. We investigate a range of projection lengths, and find that $R^p_{\rm st}$ obtained from projected profiles is close to the true $R^*_{\rm st}$, but has a slight decreasing trend with increasing projection length. We also quantify the prominence of the splashback feature and find how the feature shape changes with projection length.

astro-ph.GA↗

arXiv2024

Effect of Top Al$_2$O$_3$ Interlayer Thickness on Memory Window and Reliability of FeFETs With TiN/Al$_2$O$_3$/Hf$_{0.5}$Zr$_{0.5}$O$_2$/SiO$_x$/Si (MIFIS) Gate Structure

We investigate the effect of top Al2O3 interlayer thickness on the memory window (MW) of Si channel ferroelectric field-effect transistors (Si-FeFETs) with TiN/Al$_2$O$_3$/Hf$_{0.5}$Zr$_{0.5}$O$_2$/SiO$_x$/Si (MIFIS) gate structure. We find that the MW first increases and then remains almost constant with the increasing thickness of the top Al2O3. The phenomenon is attributed to the lower electric field of the ferroelectric Hf$_{0.5}$Zr$_{0.5}$O$_2$ in the MIFIS structure with a thicker top Al2O3 after a program operation. The lower electric field makes the charges trapped at the top Al2O3/Hf0.5Zr0.5O$_2$ interface, which are injected from the metal gate, cannot be retained. Furthermore, we study the effect of the top Al$_2$O$_3$ interlayer thickness on the reliability (endurance characteristics and retention characteristics). We find that the MIFIS structure with a thicker top Al$_2$O$_3$ interlayer has poorer retention and endurance characteristics. Our work is helpful in deeply understanding the effect of top interlayer thickness on the MW and reliability of Si-FeFETs with MIFIS gate stacks.

cond-mat.mtrl-sci↗

arXiv2024

The Geometry of Concepts: Sparse Autoencoder Feature Structure

Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The "atomic" small-scale structure contains "crystals" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man-woman-king-queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The "brain" intermediate-scale structure has significant spatial modularity; for example, math and code features form a "lobe" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The "galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.

q-bio.NC↗

arXiv2024

Impact of the Top SiO2 Interlayer Thickness on Memory Window of Si Channel FeFET with TiN/SiO2/Hf0.5Zr0.5O2/SiOx/Si (MIFIS) Gate Structure

We study the impact of top SiO2 interlayer thickness on the memory window (MW) of Si channel ferroelectric field-effect transistor (FeFET) with TiN/SiO2/Hf0.5Zr0.5O2/SiOx/Si (MIFIS) gate structure. We find that the MW increases with the increasing thickness of the top SiO2 interlayer, and such an increase exhibits a two-stage linear dependence. The physical origin is the presence of the different interfacial charges trapped at the top SiO2/Hf0.5Zr0.5O2 interface. Moreover, we investigate the dependence of endurance characteristics on initial MW. We find that the endurance characteristic degrades with increasing the initial MW. By inserting a 3.4 nm SiO2 dielectric interlayer between the gate metal TiN and the ferroelectric Hf0.5Zr0.5O2, we achieve a MW of 6.3 V and retention over 10 years. Our work is helpful in the device design of FeFET.

cond-mat.mtrl-sci↗

arXiv2024

Impact of Top SiO2 interlayer Thickness on Memory Window of Si Channel FeFET with TiN/SiO2/Hf0.5Zr0.5O2/SiOx/Si (MIFIS) Gate Structure

We study the impact of top SiO2 interlayer thickness on memory window of Si channel FeFET with TiN/SiO2/Hf0.5Zr0.5O2/SiOx/Si (MIFIS) gate structure. The memory window increases with thicker top SiO2. We realize the memory window of 6.3 V for 3.4 nm top SiO2. Moreover, we find that the endurance characteristic degrades with increasing the initial memory window.

physics.app-ph↗

arXiv2023

Enlargement of Memory Window of Si Channel FeFET by Inserting Al2O3 Interlayer on Ferroelectric Hf0.5Zr0.5O2

In this work, we demonstrate the enlargement of the memory window of Si channel FeFET with ferroelectric Hf0.5Zr0.5O2 by gate-side dielectric interlayer engineering. By inserting an Al2O3 dielectric interlayer between TiN gate metal and ferroelectric Hf0.5Zr0.5O2, we achieve a memory window of 3.2 V with endurance of ~105 cycles and retention over 10 years. The physical origin of memory window enlargement is clarified to be charge trapping at the Al2O3/Hf0.5Zr0.5O2 interface, which has an opposite charge polarity to the trapped charges at the Hf0.5Zr0.5O2/SiOx interface.

cond-mat.mtrl-sci↗

arXiv2023

Augmenting Interpretable Knowledge Tracing by Ability Attribute and Attention Mechanism

Knowledge tracing aims to model students' past answer sequences to track the change in their knowledge acquisition during exercise activities and to predict their future learning performance. Most existing approaches ignore the fact that students' abilities are constantly changing or vary between individuals, and lack the interpretability of model predictions. To this end, in this paper, we propose a novel model based on ability attributes and attention mechanism. We first segment the interaction sequences and captures students' ability attributes, then dynamically assign students to groups with similar abilities, and quantify the relevance of the exercises to the skill by calculating the attention weights between the exercises and the skill to enhance the interpretability of the model. We conducted extensive experiments and evaluate real online education datasets. The results confirm that the proposed model is better at predicting performance than five well-known representative knowledge tracing models, and the model prediction results are explained through an inference path.

cs.LG↗

arXiv2020

Thermodynamic driving force of transient negative capacitance of ferroelectric capacitors

This paper investigates the thermodynamic driving force of transient negative capacitance (NC) in the series circuit of the resistor and ferroelectric capacitor (R-FEC). We find that the widely used Landau-Khalatnikov (L-K) theory, that is, the minimum of the Gibbs free energy, is inapplicable to explain the transient NC. The thermodynamic driving force of the transient NC phenomenon is the minimum of the difference between the elastic Gibbs free energy and the electric polarization work. The appearance of the transient NC phenomenon is not due to the widely accepted view that the ferroelectric polarization goes through the negative curvature region of elastic Gibbs free energy landscape (Ga). Instead, the transient NC phenomenon appears when the energy barrier of Ga disappears. The transient NC is dependent on both the intrinsic ferroelectric material parameters and extrinsic factors in the R-FEC circuit.

physics.app-ph↗

arXiv2020

Domain-Embeddings Based DGA Detection with Incremental Training Method

DGA-based botnet, which uses Domain Generation Algorithms (DGAs) to evade supervision, has become a part of the most destructive threats to network security. Over the past decades, a wealth of defense mechanisms focusing on domain features have emerged to address the problem. Nonetheless, DGA detection remains a daunting and challenging task due to the big data nature of Internet traffic and the potential fact that the linguistic features extracted only from the domain names are insufficient and the enemies could easily forge them to disturb detection. In this paper, we propose a novel DGA detection system which employs an incremental word-embeddings method to capture the interactions between end hosts and domains, characterize time-series patterns of DNS queries for each IP address and therefore explore temporal similarities between domains. We carefully modify the Word2Vec algorithm and leverage it to automatically learn dynamic and discriminative feature representations for over 1.9 million domains, and develop an simple classifier for distinguishing malicious domains from the benign. Given the ability to identify temporal patterns of domains and update models incrementally, the proposed scheme makes the progress towards adapting to the changing and evolving strategies of DGA domains. Our system is evaluated and compared with the state-of-art system FANCI and two deep-learning methods CNN and LSTM, with data from a large university's network named TUNET. The results suggest that our system outperforms the strong competitors by a large margin on multiple metrics and meanwhile achieves a remarkable speed-up on model updating.

cs.CR↗

arXiv2019

HinDom: A Robust Malicious Domain Detection System based on Heterogeneous Information Network with Transductive Classification

Domain name system (DNS) is a crucial part of the Internet, yet has been widely exploited by cyber attackers. Apart from making static methods like blacklists or sinkholes infeasible, some weasel attackers can even bypass detection systems with machine learning based classifiers. As a solution to this problem, we propose a robust domain detection system named HinDom. Instead of relying on manually selected features, HinDom models the DNS scene as a Heterogeneous Information Network (HIN) consist of clients, domains, IP addresses and their diverse relationships. Besides, the metapath-based transductive classification method enables HinDom to detect malicious domains with only a small fraction of labeled samples. So far as we know, this is the first work to apply HIN in DNS analysis. We build a prototype of HinDom and evaluate it in CERNET2 and TUNET. The results reveal that HinDom is accurate, robust and can identify previously unknown malicious domains.

cs.CR↗