This work characterizes large language models’ chain-of-thought generation as a structured trajectory through representation space. We show that mathematical reasoning traverses functionally ordered, step-specific subspaces that become increasingly separable with layer depth. This structure already exists in base models, while reasoning training primarily accelerates convergence toward termination-related subspaces rather than introducing new representational organization. While early reasoning steps follow similar trajectories, correct and incorrect solutions diverge systematically at late stages. This late-stage divergence enables mid-reasoning prediction of final-answer correctness with ROC–AUC up to 0.87. Furthermore, we introduce trajectory-based steering, an inference-time intervention framework that enables reasoning correction and length control based on derived ideal trajectories. Together, these results establish reasoning trajectories as a geometric lens for interpreting, predicting, and controlling LLM reasoning behavior.

We address the real problem of safe, robust, adaptive resource oversubscription in uncertain environments with our proposed novel technique of chance-constrained imitation learning. Our objective is to enhance resource efficiency while ensuring safety against congestion risk. Traditional supervised or forecasting models are ineffective in learning adaptive oversubscription policies, and conventional online optimization or reinforcement learning is difficult to deploy on real systems. Offline policy learning methods, such as Imitation Learning (IL) can leverage historical resource utilization telemetry data to learn effective policies if we can ensure robustness and safety from the underlying uncertainty in the domain, and thus the data. Our work investigates the nature of this uncertainty, how it can be quantified and proposes a novel chance-constrained IL that implicitly models such uncertainty in a principled manner via additional knowledge in the form of stochastic constraints on the associated risk, to learn provably safe and robust policies. We show empirically a substantial improvement (~ 3-4×) in capacity efficiency and congestion safety in test as well as real deployments.

The advent of abundant image data has catalyzed the advancement of visual control in reinforcement learning (RL) systems, leveraging multiple view- points to capture the same physical states, which could enhance control performance theoretically. However, integrating multi-view data into representation learning remains challenging. In this paper, we introduce SMuCo, an innovative multi-view reinforcement learning algorithm that constructs robust latent representations by optimizing multi- view sequential total correlation. This technique effectively captures task-relevant information and temporal dynamics while filtering out irrelevant data. Our method supports an unlimited number of views and demonstrates superior performance over leading model-free and model-based RL algorithms. Empirical results from the DeepMind Control Suite and the Sapien Basic Manipulation Task confirm SMuCo’s enhanced efficacy, significantly improving task performance across diverse scenarios and views.

Scoring systems are commonly seen for platforms in the era of Big Data. From credit scoring systems in financial services to membership scores in E-commerce shopping platforms, platform managers use such systems to guide users towards the encouraged activity pattern, and manage resources more effectively and efficiently. To establish such scoring systems, several “empirical criteria” are first determined, followed by a dedicated top-down design for each score factor, which usually requires enormous effort to adjust and tune the scoring function in the new application scenario. What's worse, many fresh projects usually have no ground truth or any experience to evaluate a reasonable scoring system, making the designing even harder. To reduce the effort of manual adjustment of the scoring function in every new scoring system, we innovatively study the scoring system from the preset empirical criteria without any ground truth and propose a novel framework to improve the system from scratch. In this paper, we propose a “counter-empirical attacking” mechanism that can generate “attacking” behavior traces and try to break the empirical rules of the scoring system. Then an adversarial “enhancer” is applied to evaluate the scoring system and find the improvement strategy. By training the adversarial learning problem, a proper scoring function can be learned to be robust to the attacking activity traces that are trying to violate the empirical criteria. Extensive experiments have been conducted on two scoring systems, including a shared computing resource platform and a financial credit system. The experimental results have validated the effectiveness of our proposed framework.

Learning from positive and unlabeled data is known as positive-unlabeled (PU) learning in literature and has attracted much attention in recent years. One common approach in PU learning is to sample a set of pseudo-negatives from the unlabeled data using ad-hoc thresholds so that conventional supervised methods can be applied with both positive and negative samples. Owing to the label uncertainty among the unlabeled data, errors of misclassifying unlabeled positive samples as negative samples inevitably appear and may even accumulate during the training processes. Those errors often lead to performance degradation and model instability. To mitigate the impact of label uncertainty and improve the robustness of learning with positive and unlabeled data, we propose a new robust PU learning method with a training strategy motivated by the nature of human learning: easy cases should be learned first. Similar intuition has been utilized in curriculum learning to only use easier cases in the early stage of training before introducing more complex cases. Specifically, we utilize a novel ''hardness'' measure to distinguish unlabeled samples with a high chance of being negative from unlabeled samples with large label noise. An iterative training strategy is then implemented to fine-tune the selection of negative samples during the training process in an iterative manner to include more ''easy'' samples in the early stage of training. Extensive experimental validations over a wide range of learning tasks show that this approach can effectively improve the accuracy and stability of learning with positive and unlabeled data. Our code is available at https://github.com/woriazzc/Robust-PU.

Oversubscription is a common practice for improving cloud resource utilization. It allows the cloud service provider to sell more resources than the physical limit, assuming not all users would fully utilize the resources simultaneously. However, how to design an oversubscription policy that improves utilization while satisfying some safety constraints remains an open problem. Existing methods and industrial practices are over-conservative, ignoring the coordination of diverse resource usage patterns and probabilistic constraints. To address these two limitations, this paper formulates the oversubscription for cloud as a chance-constrained optimization problem and proposes an effective Chance-Constrained Multi-Agent Reinforcement Learning (C2MARL) method to solve this problem. Specifically, C2MARL reduces the number of constraints by considering their upper bounds and leverages a multi-agent reinforcement learning paradigm to learn a safe and optimal coordination policy. We evaluate our C2MARL on an internal cloud platform and public cloud datasets. Experiments show that our C2MARL outperforms existing methods in improving utilization () under different levels of safety constraints.

Capacity management has always been a great challenge for cloud platforms due to massive, heterogeneous on-demand instances running at different times. To better plan the capacity for the whole platform, a class of cloud computing instances have been released to collect computing demands beforehand. To use such instances, users are allowed to submit jobs to run for a pre-specified uninterrupted duration in a flexible range of time in the future with a discount compared to the normal on-demand instances. Proactively scheduling those pre-collected job requests considering the capacity status over the platform can greatly help balance the computing workloads along time. In this work, we formulate the scheduling problem for these pre-collected job requests under uncertain available capacity as a Prediction + Optimization problem with uncertainty in constraints, and propose an effective algorithm called Controlling under Uncertain Constraints (CUC), where the predicted capacity guides the optimization of job scheduling and job scheduling results are leveraged to improve the prediction of capacity through Bayesian optimization. The proposed formulation and solution are commonly applicable for proactively scheduling problems in cloud computing. Our extensive experiments on three public, industrial datasets shows that CUC has great potential for supporting high reliability in cloud platforms.

In cloud systems, an accurate capacity planning is very important for cloud provider to improve service availability. Traditional methods simply predicting "when the available resources is exhausted" are not effective due to customer demand fragmentation and platform allocation constraints. In this paper, we propose a novel prediction approach which proactively predicts the level of resource allocation failures from the perspective of low capacity status. By jointly considering the data from different sources in both time series form and static form, the proposed approach can make accurate LCS predictions in a complex and dynamic cloud environment, and thereby improve the service availability of cloud systems. The proposed approach is evaluated by real-world datasets collected from a large scale public cloud platform, and the results confirm its effectiveness.

In large-scale online service systems, incidents occur frequently due to a variety of causes, from updates of software and hardware to changes in operation environment. These incidents could significantly degrade system’s availability and customers’ satisfaction. Some incidents are linked because they are duplicate or inter-related. The linked incidents can greatly help on-call engineers find mitigation solutions and identify the root causes. In this work, we investigate the incidents and their links in a representative real-world incident management (IcM) system. Based on the identified indicators of linked incidents, we further propose LiDAR (Linked Incident identification with DAta-driven Representation), a deep learning based approach to incident linking. More specifically, we incorporate the textual description of incidents and structural information extracted from historical linked incidents to identify possible links among a large number of incidents. To show the effectiveness of our method, we apply our method to a real-world IcM system and find that our method outperforms other state-of-the-art methods.

With the rapid growth of cloud service systems and their increasing complexity, service failures become unavoidable. Outages, which are critical service failures, could dramatically degrade system availability and impact user experience. To minimize service downtime and ensure high system availability, we develop an intelligent outage management approach, called AirAlert, which can forecast the occurrence of outages before they actually happen and diagnose the root cause after they indeed occur. AirAlert works as a global watcher for the entire cloud system, which collects all alerting signals, detects dependency among signals and proactively predicts outages that may happen anywhere in the whole cloud system. We analyze the relationships between outages and alerting signals by leveraging Bayesian network and predict outages using a robust gradient boosting tree based classification method. The proposed outage management approach is evaluated using the outage dataset collected from a Microsoft cloud system and the results confirm the effectiveness of the proposed approach.