arXiv subjects

Cong Feng

Publications and source records attributed to Cong Feng.

At least 19 recordsLinked to original sources

arXiv2026

K/V-Cache Interventions Dissociate Representation Alignment from Persona Expression in Decoder-Only Language Models

We study K/V-cache interventions -- transplanting a target-conditioned K/V trajectory into a source-persona generation -- as a structured surface for persona control in decoder-only language models. Across 13 intervention configurations applied to Llama-3.1-8B for a fixed source-to-target persona pair, we report two consistent dissociations between representation-level alignment and behavioral expression, plus a common failure under position perturbations. First, all layer-band K/V replacements (early, mid, late) achieve strong local V-space alignment (V-gap 0.91, 0.89, 0.84), but only mid-layer replacement (layers 9-20) combines substantial target-marker expression with preserved lexical diversity. Second, full and mid-layer replacement induce comparable alignment (V-gap 0.94 vs. 0.89) yet produce different lexical-diversity profiles (TTR 0.65 vs. 0.77). Third, position perturbations (lag and shuffle) apply distinct operations yet uniformly suppress target-persona expression -- a common behavioral failure rather than a strict dissociation. Representation-level similarity metrics alone are thus not sufficient predictors of downstream persona expression in the regimes we study; the K/V cache emerges as a controllable but structurally constrained intervention surface. Because the transplanted trajectory carries the target's own generated token history, we characterize the intervention as trajectory-level transplantation rather than isolated persona-representation injection; a same-token-sequence control, decoding an identical token sequence under source vs. target conditioning, reproduces the sign and layer localization of the L28 representational shift, indicating the shift is not explained solely by imported token history. These findings characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.

cs.CL↗

arXiv2026

SolarBench: A global solar energy nowcasting benchmark

As the share of solar power grows, nowcasting weather-driven solar variability becomes critical for reliable energy system operation. State-of-the-art approaches increasingly apply deep learning to sky camera and geostationary satellite observations, but fragmented datasets and inconsistent evaluation make it difficult to determine whether reported improvements generalize across climates, cloud regimes, and photovoltaic (PV) systems. Here we introduce SolarBench, an open global benchmark for image-based solar nowcasting. SolarBench harmonizes more than six million sky and satellite images from 11 diverse sites spanning a decade, together with irradiance or PV output and auxiliary atmospheric data. An accompanying toolbox supports reproducible data access, processing, model development, and evaluation. Using SolarBench, we benchmark representative models and reveal a gap between average forecasting accuracy and the ability to capture rapid solar fluctuations. We further quantify predictability across cloud regimes and demonstrate data-efficient adaptation to new PV systems. SolarBench provides an extensible foundation for fair comparison and methodological innovation in solar nowcasting.

cs.CV↗

arXiv2026

SPEAR: Code-Augmented Agentic Prompt Optimization

Automatic prompt engineering (APE) rewrites prompts to improve downstream task performance, but existing APE loops treat the optimizer itself as a fixed pipeline. We port the code-as-action paradigm of CodeAct (Wang et al., 2024a) to APE and propose SPEAR (Sandboxed Prompt Engineer with Active Roll-back), a free-form agentic optimizer with four tools -- evaluate, python, set_prompt, finish -- that decides autonomously how and when to use them. The distinctive tool is the Python sandbox: the optimizer writes and executes arbitrary Python on the current evaluation DataFrame, performing structural error analysis (confusion matrices, error clustering, per group metrics) the agent itself authors. Two guardrails turn the long-horizon agent into a monotone-improving optimizer: auto-rollback on metric regression, and an optional guard metric floor. We evaluate on three industrial LLM-as-judge suites (13 judge tasks across recruiter-intake, conversational-memory, and query-refinement systems) plus seven BBH tasks and GSM8K. SPEAR wins every industrial task on the primary metric ($\kappa$ 0.857 vs 0.359 on tool-selection; F1-macro 0.815 vs 0.763 on filter-relevance; $\kappa$ 0.254 vs 0.218 on the hardest extraction dimension). On BBH-7 SPEAR averages 0.938 accuracy vs GEPA 0.628 and TextGrad 0.484. Ablations show the Python tool is the largest single lever on complex judge tasks ($\Delta \approx +0.79\kappa$ on the 5-class tool-selection judge, $\Delta \approx +0.35\kappa$ on the hardest extraction dimension when removed); its irreplaceable contribution is class-pair confusion aggregation that a long-context LLM cannot extract reliably from the raw eval DataFrame.

cs.CL↗

arXiv2026

Emulating Clinician Cognition via Self-Evolving Deep Clinical Research

Clinical diagnosis is a complex cognitive process, grounded in dynamic cue acquisition and continuous expertise accumulation. Yet most current artificial intelligence (AI) systems are misaligned with this reality, treating diagnosis as single-pass retrospective prediction while lacking auditable mechanisms for governed improvement. We developed DxEvolve, a self-evolving diagnostic agent that bridges these gaps through an interactive deep clinical research workflow. The framework autonomously requisitions examinations and continually externalizes clinical experience from increasing encounter exposure as diagnostic cognition primitives. On the MIMIC-CDM benchmark, DxEvolve improved diagnostic accuracy by 11.2% on average over backbone models and reached 90.4% on a reader-study subset, comparable to the clinician reference (88.8%). DxEvolve improved accuracy on an independent external cohort by 10.2% (categories covered by the source cohort) and 17.1% (uncovered categories) compared to the competitive method. By transforming experience into a governable learning asset, DxEvolve supports an accountable pathway for the continual evolution of clinical AI.

cs.AI↗

arXiv2026

ReviveMoE: Fast Recovery for Hardware Failures in Large-Scale MoE LLM Inference Deployments

As LLM deployments scale over more hardware, the probability of a single failure in a system increases significantly, and cloud operators must consider robust countermeasures to handle these inevitable failures. A common recovery approach is to simply restart the LLM serving instance; however, this is costly in model-as-a-service (MaaS) inference settings, where reloading model weights and recompiling computation graphs can introduce significant delays to incoming requests. We propose ReviveMoE, a method for rapid failure recovery in large-scale LLM deployments without restarting the serving instance. ReviveMoE is designed to support both the traditional LLM architecture, which collocates MoE and attention on the same hardware, and the disaggregated architectures, which separate MoE from attention. Integrated into Huawei Cloud's MaaS, ReviveMoE is built on top of Huawei's xDeepServe serving platform and the XCCL communications library.

cs.DC↗

arXiv2025

Huawei Cloud Model-as-a-Service on the CloudMatrix384 SuperPod

Scaled-out MoE LLMs and scaled-up SuperPods create new systems challenges for production Model-as-a-Service (MaaS), requiring disaggregation, low-latency communication, and decentralized serving. This report presents xDeepServe, the production serving system behind Huawei Cloud's MaaS offering on CloudMatrix384, a 48-server SuperPod with 384 Ascend 910C chips connected by a high-bandwidth UB fabric and global shared memory. It serves models including DeepSeek, Kimi, GLM, Qwen, and MiniMax, among others. xDeepServe is built around Transformerless, a disaggregated execution architecture that decomposes transformer inference into modular units -- attention, feedforward, and MoE -- and supports disaggregated Prefill-Decode and MoE-Attention deployments. To enable disaggregation, we develop XCCL, a memory-semantic communication layer providing microsecond-level point-to-point and scalable all-to-all primitives, and we extend FlowServe with decentralized DP groups and techniques to mitigate stragglers and synchronization variance. In a peak decoding configuration, xDeepServe reaches 2400 tokens/s per Ascend 910C chip at ~50ms time-per-output-token (TPOT).

cs.DC↗

arXiv2025

LLMPrism: Black-box Performance Diagnosis for Production LLM Training Platforms

Large Language Models (LLMs) have brought about revolutionary changes in diverse fields, rendering LLM training of utmost importance for modern enterprises. To meet this demand, multi-tenant large-scale LLM training platforms have been built to offer LLM training services. Nevertheless, due to the complexity and synchronous nature of LLM training process, performance issues occur frequently and can result in substantial resource wastage. The limited visibility from the perspective of platform providers impedes existing profiling methods and poses challenges to the monitoring and diagnosis of the performance of LLM training jobs. For the first time, this paper proposes the utilization of underlying network flow data to reconstruct the training timelines of jobs based on the distinct characteristics in the LLM training procedure. We design LLMPrism, the first black-box performance diagnosis system for LLM training platforms. By progressively recognizing LLM training jobs, identifying their parallelism strategies, and reconstructing the training timelines, LLMPrism achieves non-intrusive, lightweight, and continuous monitoring of LLM training systems. Leveraging this monitoring capability, it further effectively diagnoses potential performance issues. Since Oct. 2024, LLMPrism has been deployed on our large-scale production Platform-X, in which the evaluations and deployment experiences demonstrate that LLMPrism can achieve accurate timeline reconstruction with an error within 0.3% and effectively diagnose various performance issues.

cs.SE↗

arXiv2025

L4: Diagnosing Large-scale LLM Training Failures via Automated Log Analysis

As Large Language Models (LLMs) show their capabilities across various applications, training customized LLMs has become essential for modern enterprises. However, due to the complexity of LLM training, which requires massive computational resources and extensive training time, failures are inevitable during the training process. These failures result in considerable waste of resource and time, highlighting the critical need for effective and efficient failure diagnosis to reduce the cost of LLM training. In this paper, we present the first empirical study on the failure reports of 428 LLM training failures in our production Platform-X between May 2023 and April 2024. Our study reveals that hardware and user faults are the predominant root causes, and current diagnosis processes rely heavily on training logs. Unfortunately, existing log-based diagnostic methods fall short in handling LLM training logs. Considering the unique features of LLM training, we identify three distinct patterns of LLM training logs: cross-job, spatial, and temporal patterns. We then introduce our Log-based Large-scale LLM training failure diagnosis framework, L4, which can automatically extract failure-indicating information (i.e., log events, nodes, stages, and iterations) from extensive training logs, thereby reducing manual effort and facilitating failure recovery. Experimental results on real-world datasets show that L4 outperforms existing approaches in identifying failure-indicating logs and localizing faulty nodes. Furthermore, L4 has been applied in Platform-X and demonstrated its effectiveness in enabling accurate and efficient failure diagnosis.

cs.SE↗

arXiv2024

Exploring the Use of Autonomous Unmanned Vehicles for Supporting Power Grid Operations

This paper explores the use of autonomous unmanned vehicles to support power grid operations. With built-in batteries and the capability to carry additional battery energy storage, the rising number of autonomous vehicles can represent a substantial amount of capacity that is currently underutilized in the power grid. Unlike traditional electric vehicles that require drivers, the operations of autonomous vehicles can be performed without human intervention. To guide idle vehicles to autonomously support power grids, we propose a tractable optimization-based method to effectively integrate these "mobile batteries" into grid operations. During real-time operations, the vehicles are strategically routed to target locations to maintain power balance and reduce operating costs. Numerical studies have confirmed both the validity and the scalability of the proposed algorithm to efficiently integrate autonomous vehicles into routine power system operations.

eess.SY↗

arXiv2024

Demystifying and Extracting Fault-indicating Information from Logs for Failure Diagnosis

Logs are imperative in the maintenance of online service systems, which often encompass important information for effective failure mitigation. While existing anomaly detection methodologies facilitate the identification of anomalous logs within extensive runtime data, manual investigation of log messages by engineers remains essential to comprehend faults, which is labor-intensive and error-prone. Upon examining the log-based troubleshooting practices at CloudA, we find that engineers typically prioritize two categories of log information for diagnosis. These include fault-indicating descriptions, which record abnormal system events, and fault-indicating parameters, which specify the associated entities. Motivated by this finding, we propose an approach to automatically extract such faultindicating information from logs for fault diagnosis, named LoFI. LoFI comprises two key stages. In the first stage, LoFI performs coarse-grained filtering to collect logs related to the faults based on semantic similarity. In the second stage, LoFI leverages a pre-trained language model with a novel prompt-based tuning method to extract fine-grained information of interest from the collected logs. We evaluate LoFI on logs collected from Apache Spark and an industrial dataset from CloudA. The experimental results demonstrate that LoFI outperforms all baseline methods by a significant margin, achieving an absolute improvement of 25.8~37.9 in F1 over the best baseline method, ChatGPT. This highlights the effectiveness of LoFI in recognizing fault-indicating information. Furthermore, the successful deployment of LoFI at CloudA and user studies validate the utility of our method. The code and data are available at https://github.com/Jun-jie-Huang/LoFI.

cs.SE↗

arXiv2024

Multi-Modal Multi-Granularity Tokenizer for Chu Bamboo Slip Scripts

This study presents a multi-modal multi-granularity tokenizer specifically designed for analyzing ancient Chinese scripts, focusing on the Chu bamboo slip (CBS) script used during the Spring and Autumn and Warring States period (771-256 BCE) in Ancient China. Considering the complex hierarchical structure of ancient Chinese scripts, where a single character may be a combination of multiple sub-characters, our tokenizer first adopts character detection to locate character boundaries, and then conducts character recognition at both the character and sub-character levels. Moreover, to support the academic community, we have also assembled the first large-scale dataset of CBSs with over 100K annotated character image scans. On the part-of-speech tagging task built on our dataset, using our tokenizer gives a 5.5% relative improvement in F1-score compared to mainstream sub-word tokenizers. Our work not only aids in further investigations of the specific script but also has the potential to advance research on other forms of ancient Chinese scripts.

cs.CL↗

arXiv2024

FaultProfIT: Hierarchical Fault Profiling of Incident Tickets in Large-scale Cloud Systems

Postmortem analysis is essential in the management of incidents within cloud systems, which provides valuable insights to improve system's reliability and robustness. At CloudA, fault pattern profiling is performed during the postmortem phase, which involves the classification of incidents' faults into unique categories, referred to as fault pattern. By aggregating and analyzing these fault patterns, engineers can discern common faults, vulnerable components and emerging fault trends. However, this process is currently conducted by manual labeling, which has inherent drawbacks. On the one hand, the sheer volume of incidents means only the most severe ones are analyzed, causing a skewed overview of fault patterns. On the other hand, the complexity of the task demands extensive domain knowledge, which leads to errors and inconsistencies. To address these limitations, we propose an automated approach, named FaultProfIT, for Fault pattern Profiling of Incident Tickets. It leverages hierarchy-guided contrastive learning to train a hierarchy-aware incident encoder and predicts fault patterns with enhanced incident representations. We evaluate FaultProfIT using the production incidents from CloudA. The results demonstrate that FaultProfIT outperforms state-of-the-art methods. Our ablation study and analysis also verify the effectiveness of hierarchy-guided contrastive learning. Additionally, we have deployed FaultProfIT at CloudA for six months. To date, FaultProfIT has analyzed 10,000+ incidents from 30+ cloud services, successfully revealing several fault trends that have informed system improvements.

cs.SE↗

arXiv2023

Practical Anomaly Detection over Multivariate Monitoring Metrics for Online Services

As modern software systems continue to grow in terms of complexity and volume, anomaly detection on multivariate monitoring metrics, which profile systems' health status, becomes more and more critical and challenging. In particular, the dependency between different metrics and their historical patterns plays a critical role in pursuing prompt and accurate anomaly detection. Existing approaches fall short of industrial needs for being unable to capture such information efficiently. To fill this significant gap, in this paper, we propose CMAnomaly, an anomaly detection framework on multivariate monitoring metrics based on collaborative machine. The proposed collaborative machine is a mechanism to capture the pairwise interactions along with feature and temporal dimensions with linear time complexity. Cost-effective models can then be employed to leverage both the dependency between monitoring metrics and their historical patterns for anomaly detection. The proposed framework is extensively evaluated with both public data and industrial data collected from a large-scale online service system of Huawei Cloud. The experimental results demonstrate that compared with state-of-the-art baseline models, CMAnomaly achieves an average F1 score of 0.9494, outperforming baselines by 6.77% to 10.68%, and runs 10X to 20X faster. Furthermore, we also share our experience of deploying CMAnomaly in Huawei Cloud.

cs.SE↗

arXiv2023

Prism: Revealing Hidden Functional Clusters from Massive Instances in Cloud Systems

Ensuring the reliability of cloud systems is critical for both cloud vendors and customers. Cloud systems often rely on virtualization techniques to create instances of hardware resources, such as virtual machines. However, virtualization hinders the observability of cloud systems, making it challenging to diagnose platform-level issues. To improve system observability, we propose to infer functional clusters of instances, i.e., groups of instances having similar functionalities. We first conduct a pilot study on a large-scale cloud system, i.e., Huawei Cloud, demonstrating that instances having similar functionalities share similar communication and resource usage patterns. Motivated by these findings, we formulate the identification of functional clusters as a clustering problem and propose a non-intrusive solution called Prism. Prism adopts a coarse-to-fine clustering strategy. It first partitions instances into coarse-grained chunks based on communication patterns. Within each chunk, Prism further groups instances with similar resource usage patterns to produce fine-grained functional clusters. Such a design reduces noises in the data and allows Prism to process massive instances efficiently. We evaluate Prism on two datasets collected from the real-world production environment of Huawei Cloud. Our experiments show that Prism achieves a v-measure of ~0.95, surpassing existing state-of-the-art solutions. Additionally, we illustrate the integration of Prism within monitoring systems for enhanced cloud reliability through two real-world use cases.

cs.SE↗

arXiv2023

Identifying Performance Issues in Cloud Service Systems Based on Relational-Temporal Features

Cloud systems are susceptible to performance issues, which may cause service-level agreement violations and financial losses. In current practice, crucial metrics are monitored periodically to provide insight into the operational status of components. Identifying performance issues is often formulated as an anomaly detection problem, which is tackled by analyzing each metric independently. However, this approach overlooks the complex dependencies existing among cloud components. Some graph neural network-based methods take both temporal and relational information into account, however, the correlation violations in the metrics that serve as indicators of underlying performance issues are difficult for them to identify. Furthermore, a large volume of components in a cloud system results in a vast array of noisy metrics. This complexity renders it impractical for engineers to fully comprehend the correlations, making it challenging to identify performance issues accurately. To address these limitations, we propose Identifying Performance Issues based on Relational-Temporal Features (ISOLATE ), a learning-based approach that leverages both the relational and temporal features of metrics to identify performance issues. In particular, it adopts a graph neural network with attention to characterizing the relations among metrics and extracts long-term and multi-scale temporal patterns using a GRU and a convolution network, respectively. The learned graph attention weights can be further used to localize the correlation-violated metrics. Moreover, to relieve the impact of noisy data, ISOLATE utilizes a positive unlabeled learning strategy that tags pseudo-labels based on a small portion of confirmed negative examples. Extensive evaluation on both public and industrial datasets shows that ISOLATE outperforms all baseline models with 0.945 F1-score and 0.920 Hit rate@3.

cs.LG↗

arXiv2023

Log-based Anomaly Detection based on EVT Theory with feedback

System logs play a critical role in maintaining the reliability of software systems. Fruitful studies have explored automatic log-based anomaly detection and achieved notable accuracy on benchmark datasets. However, when applied to large-scale cloud systems, these solutions face limitations due to high resource consumption and lack of adaptability to evolving logs. In this paper, we present an accurate, lightweight, and adaptive log-based anomaly detection framework, referred to as SeaLog. Our method introduces a Trie-based Detection Agent (TDA) that employs a lightweight, dynamically-growing trie structure for real-time anomaly detection. To enhance TDA's accuracy in response to evolving log data, we enable it to receive feedback from experts. Interestingly, our findings suggest that contemporary large language models, such as ChatGPT, can provide feedback with a level of consistency comparable to human experts, which can potentially reduce manual verification efforts. We extensively evaluate SeaLog on two public datasets and an industrial dataset. The results show that SeaLog outperforms all baseline methods in terms of effectiveness, runs 2X to 10X faster and only consumes 5% to 41% of the memory resource.

cs.SE↗

arXiv2019

Transferrable Operative Difficulty Assessment in Robot-assisted Teleoperation: A Domain Adaptation Approach

Providing an accurate and efficient assessment of operative difficulty is important for designing robot-assisted teleoperation interfaces that are easy and natural for human operators to use. In this paper, we aim to develop a data-driven approach to numerically characterize the operative difficulty demand of complex teleoperation. In effort to provide an entirely task-independent assessment, we consider using only data collected from the human user including: (1) physiological response, and (2) movement kinematics. By leveraging an unsupervised domain adaptation technique, our approach learns the user information that defines task difficulty in a well-known source, namely, a Fitt's target reaching task, and generalizes that knowledge to a more complex human motor control scenario, namely, the teleoperation of a robotic system. Our approach consists of two main parts: (1) The first part accounts for the inherent variances of user physiological and kinematic response between these cross-domain motor control scenarios that are vastly different. (2) A stacked two-layer learner is designed to improve the overall modeling performance, yielding a 96.6% accuracy in predicting the known difficulty of a Fitts' reaching task when using movement kinematic features. We then validate the effectiveness of our model by investigating teleoperated robotic needle steering as a case study. Compared with a standard NASA TLX user survey, our results indicate significant differences in the difficulty demand for various choices of needle steering control algorithms, p<0.05, as well as the difficulty of steering the needle to different targets, p<0.05. The results highlight the potential of our approach to be used as a design tool to create more intuitive and natural teleoperation interfaces in robot-assisted systems.

cs.RO↗

arXiv2018

Reinforcement Learning based Dynamic Model Selection for Short-Term Load Forecasting

With the growing prevalence of smart grid technology, short-term load forecasting (STLF) becomes particularly important in power system operations. There is a large collection of methods developed for STLF, but selecting a suitable method under varying conditions is still challenging. This paper develops a novel reinforcement learning based dynamic model selection (DMS) method for STLF. A forecasting model pool is first built, including ten state-of-the-art machine learning based forecasting models. Then a Q-learning agent learns the optimal policy of selecting the best forecasting model for the next time step, based on the model performance. The optimal DMS policy is applied to select the best model at each time step with a moving window. Numerical simulations on two-year load and weather data show that the Q-learning algorithm converges fast, resulting in effective and efficient DMS. The developed STLF model with Q-learning based DMS improves the forecasting accuracy by approximately 50%, compared to the state-of-the-art machine learning based STLF models.

cs.LG↗