Junjie Sheng

Junjie Sheng, Lu Wang, Fangkai Yang et al.

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 the 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 propose 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 ($20\%\sim 86\%$) under different levels of safety constraints.

Junjie Sheng, Wenhao Li, Bo Jin et al.

Over-generalization is a thorny issue in cognitive science, where people may become overly cautious due to past experiences. Agents in multi-agent reinforcement learning (MARL) also have been found to suffer relative over-generalization (RO) as people do and stuck to sub-optimal cooperation. Recent methods have shown that assigning reasoning ability to agents can mitigate RO algorithmically and empirically, but there has been a lack of theoretical understanding of RO, let alone designing provably RO-free methods. This paper first proves that RO can be avoided when the MARL method satisfies a consistent reasoning requirement under certain conditions. Then we introduce a novel reasoning framework, called negotiated reasoning, that first builds the connection between reasoning and RO with theoretical justifications. After that, we propose an instantiated algorithm, Stein variational negotiated reasoning (SVNR), which uses Stein variational gradient descent to derive a negotiation policy that provably avoids RO in MARL under maximum entropy policy iteration. The method is further parameterized with neural networks for amortized learning, making computation efficient. Numerical experiments on many RO-challenged environments demonstrate the superiority and efficiency of SVNR compared to state-of-the-art methods in addressing RO.

Jiehao Wu, Zixiao Huang, Wenhao Li et al.

AscendC (Ascend C) operator optimization on Huawei Ascend neural processing units (NPUs) faces a two-fold knowledge bottleneck: unlike the CUDA ecosystem, there are few public reference implementations to learn from, and performance hinges on a coupled two-part artifact - a host-side tiling program that orchestrates data movement and a kernel program that schedules and pipelines instructions. We present AscendOptimizer, an episodic agent that bootstraps this missing expertise by turning execution into experience. On the host side, AscendOptimizer performs profiling-in-the-loop evolutionary search to discover valid and high-performing tiling and data-movement configurations directly from hardware feedback. On the kernel side, it mines transferable optimization motifs by rewinding optimized kernels - systematically de-optimizing them to synthesize instructive "bad-to-good" trajectories - and distills these motifs into a retrievable experience bank for guided rewriting. By alternating host tuning and kernel rewriting in a closed loop, AscendOptimizer steadily expands feasibility and pushes latency down. On a benchmark of 127 real AscendC operators, AscendOptimizer achieves a 1.19x geometric-mean speedup over the open-source baseline, with 49.61% of operators outperforming their references, outperforming strong agent and search baselines.

Haochuan Cui, Junjie Sheng, Bo Jin et al.

With the rapid development of cloud computing, virtual machine scheduling has become one of the most important but challenging issues for the cloud computing community, especially for practical heterogeneous request sequences. By analyzing the impact of request heterogeneity on some popular heuristic schedulers, it can be found that existing scheduling algorithms can not handle the request heterogeneity properly and efficiently. In this paper, a plug-and-play virtual machine scheduling intensifier, called Resource Assigner (ReAssigner), is proposed to enhance the scheduling efficiency of any given scheduler for heterogeneous requests. The key idea of ReAssigner is to pre-assign roles to physical resources and let resources of the same role form a virtual cluster to handle homogeneous requests. ReAssigner can cooperate with arbitrary schedulers by restricting their scheduling space to virtual clusters. With evaluations on the real dataset from Huawei Cloud, the proposed ReAssigner achieves significant scheduling performance improvement compared with some state-of-the-art scheduling methods.

Yujie Zheng, Zhuo Li, Shengtao Zhang et al.

Deploying Large Language Models to data-scarce programming domains poses significant challenges, particularly for kernel synthesis on emerging Domain-Specific Architectures where a "Data Wall" limits available training data. While models excel on data-rich platforms like CUDA, they suffer catastrophic performance drops on data-scarce ecosystems such as NPU programming. To overcome this cold-start barrier without expensive fine-tuning, we introduce EvoKernel, a self-evolving agentic framework that automates the lifecycle of kernel synthesis from initial drafting to continual refining. EvoKernel addresses this by formulating the synthesis process as a memory-based reinforcement learning task. Through a novel value-driven retrieval mechanism, it learns stage-specific Q-values that prioritize experiences based on their contribution to the current objective, whether bootstrapping a feasible draft or iteratively refining latency. Furthermore, by enabling cross-task memory sharing, the agent generalizes insights from simple to complex operators. By building an NPU variant of KernelBench and evaluating on it, EvoKernel improves frontier models' correctness from 11.0% to 83.0% and achieves a median speedup of 3.60x over initial drafts through iterative refinement. This demonstrates that value-guided experience accumulation allows general-purpose models to master the kernel synthesis task on niche hardware ecosystems. Our official page is available at https://evokernel.zhuo.li.

Wenhao Li, Wenwu Li, Chuyun Shen et al.

We present TextAtari, a benchmark for evaluating language agents on very long-horizon decision-making tasks spanning up to 100,000 steps. By translating the visual state representations of classic Atari games into rich textual descriptions, TextAtari creates a challenging test bed that bridges sequential decision-making with natural language processing. The benchmark includes nearly 100 distinct tasks with varying complexity, action spaces, and planning horizons, all rendered as text through an unsupervised representation learning framework (AtariARI). We evaluate three open-source large language models (Qwen2.5-7B, Gemma-7B, and Llama3.1-8B) across three agent frameworks (zero-shot, few-shot chain-of-thought, and reflection reasoning) to assess how different forms of prior knowledge affect performance on these long-horizon challenges. Four scenarios-Basic, Obscured, Manual Augmentation, and Reference-based-investigate the impact of semantic understanding, instruction comprehension, and expert demonstrations on agent decision-making. Our results reveal significant performance gaps between language agents and human players in extensive planning tasks, highlighting challenges in sequential reasoning, state tracking, and strategic planning across tens of thousands of steps. TextAtari provides standardized evaluation protocols, baseline implementations, and a framework for advancing research at the intersection of language models and planning. Our code is available at https://github.com/Lww007/Text-Atari-Agents.

Junjie Sheng, Zixiao Huang, Chuyun Shen et al.

The formidable capacity for zero- or few-shot decision-making in language agents encourages us to pose a compelling question: Can language agents be alternatives to PPO agents in traditional sequential decision-making tasks? To investigate this, we first take environments collected in OpenAI Gym as our testbeds and ground them to textual environments that construct the TextGym simulator. This allows for straightforward and efficient comparisons between PPO agents and language agents, given the widespread adoption of OpenAI Gym. To ensure a fair and effective benchmarking, we introduce $5$ levels of scenario for accurate domain-knowledge controlling and a unified RL-inspired framework for language agents. Additionally, we propose an innovative explore-exploit-guided language (EXE) agent to solve tasks within TextGym. Through numerical experiments and ablation studies, we extract valuable insights into the decision-making capabilities of language agents and make a preliminary evaluation of their potential to be alternatives to PPO in classical sequential decision-making problems. This paper sheds light on the performance of language agents and paves the way for future research in this exciting domain. Our code is publicly available at~\url{https://github.com/mail-ecnu/Text-Gym-Agents}.

Junjie Sheng, Jiehao Wu, Haochuan Cui et al.

Recent advancements in reinforcement learning (RL) have shown promise for optimizing virtual machine scheduling (VMS) in small-scale clusters. The utilization of RL to large-scale cloud computing scenarios remains notably constrained. This paper introduces a scalable RL framework, called Cluster Value Decomposition Reinforcement Learning (CVD-RL), to surmount the scalability hurdles inherent in large-scale VMS. The CVD-RL framework innovatively combines a decomposition operator with a look-ahead operator to adeptly manage representation complexities, while complemented by a Top-$k$ filter operator that refines exploration efficiency. Different from existing approaches limited to clusters of $10$ or fewer physical machines (PMs), CVD-RL extends its applicability to environments encompassing up to $50$ PMs. Furthermore, the CVD-RL framework demonstrates generalization capabilities that surpass contemporary SOTA methodologies across a variety of scenarios in empirical studies. This breakthrough not only showcases the framework's exceptional scalability and performance but also represents a significant leap in the application of RL for VMS within complex, large-scale cloud infrastructures. The code is available at https://anonymous.4open.science/r/marl4sche-D0FE.

Zixiao Huang, Lifeng Guo, Wenhao Li et al.

Graph combinatorial optimization (GCO) problems are central to domains like logistics and bioinformatics. While traditional solvers dominate, large language models (LLMs) offer new possibilities for structured reasoning, yet struggle with complex GCO tasks requiring rigorous combinatorial analysis and multi-step deduction, often producing hallucinated steps. We first formalize the Optimal Thoughts Design (OTD) problem, which provides a structured guidance for producing high-quality intermediate reasoning steps. Building on this formulation, we introduce GraphThought, a novel framework that generates effective reasoning sequences through either heuristic-guided forward search or solver-aligned backward reasoning. By fine-tuning LLMs on these structured thought sequences, we develop Llama-GT, an 8B-parameter model that achieves state-of-the-art performance on the GraphArena benchmark, outperforming significantly larger models like DeepSeek-V3. Our results demonstrate that when scaffolded with structured reasoning priors, principled thought generation can significantly enhance LLM performance on GCO tasks without requiring increased model scale.

JieHao Wu, Ziwei Wang, Junjie Sheng et al.

In cloud services, virtual machine (VM) scheduling is a typical Online Dynamic Multidimensional Bin Packing (ODMBP) problem, characterized by large-scale complexity and fluctuating demands. Traditional optimization methods struggle to adapt to real-time changes, domain-expert-designed heuristic approaches suffer from rigid strategies, and existing learning-based methods often lack generalizability and interpretability. To address these limitations, this paper proposes a hierarchical language agent framework named MiCo, which provides a large language model (LLM)-driven heuristic design paradigm for solving ODMBP. Specifically, ODMBP is formulated as a Semi-Markov Decision Process with Options (SMDP-Option), enabling dynamic scheduling through a two-stage architecture, i.e., Option Miner and Option Composer. Option Miner utilizes LLMs to discover diverse and useful non-context-aware strategies by interacting with constructed environments. Option Composer employs LLMs to discover a composing strategy that integrates the non-context-aware strategies with the contextual ones. Extensive experiments on real-world enterprise datasets demonstrate that MiCo achieves a 96.9\% competitive ratio in large-scale scenarios involving more than 10,000 virtual machines. It maintains high performance even under nonstationary request flows and diverse configurations, thus validating its effectiveness in complex and large-scale cloud environments.

Xidong Yang, Wenhao Li, Junjie Sheng et al.

Reinforcement learning (RL) has driven breakthroughs in AI, from game-play to scientific discovery and AI alignment. However, its broader applicability remains limited by challenges such as low data efficiency and poor generalizability. Recent advances suggest that large language models, with their rich world knowledge and reasoning capabilities, could complement RL by enabling semantic state modeling and task-agnostic planning. In this work, we propose the Agentic Episodic Control (AEC), a novel architecture that integrates RL with LLMs to enhance decision-making. The AEC can leverage a large language model (LLM) to map the observations into language-grounded embeddings, which further can be stored in an episodic memory for rapid retrieval of high-value experiences. Simultaneously, a World-Graph working memory module is utilized to capture structured environmental dynamics in order to enhance relational reasoning. Furthermore, a lightweight critical state detector dynamically arbitrates between the episodic memory recall and the world-model-guided exploration. On the whole, by combining the trial-and-error learning scheme with LLM-derived semantic priors, the proposed AEC can improve both data efficiency and generalizability in reinforcement learning. In experiments on BabyAI-Text benchmark tasks, AEC demonstrates substantial improvements over existing baselines, especially on complex and generalization tasks like FindObj, where it outperforms the best baseline by up to 76%. The proposed AEC framework bridges the strengths of numeric reinforcement learning and symbolic reasoning, which provides a pathway toward more adaptable and sample-efficient agents.

Moyi Yang, Junjie Sheng, Xiangfeng Wang et al.

Fairness has been taken as a critical metric in machine learning models, which is considered as an important component of trustworthy machine learning. In this paper, we focus on obtaining fairness for popular link prediction tasks, which are measured by dyadic fairness. A novel pre-processing methodology is proposed to establish dyadic fairness through data repairing based on optimal transport theory. With the well-established theoretical connection between the dyadic fairness for graph link prediction and a conditional distribution alignment problem, the dyadic repairing scheme can be equivalently transformed into a conditional distribution alignment problem. Furthermore, an optimal transport-based dyadic fairness algorithm called DyadicOT is obtained by efficiently solving the alignment problem, satisfying flexibility and unambiguity requirements. The proposed DyadicOT algorithm shows superior results in obtaining fairness compared to other fairness methods on two benchmark graph datasets.

Junjie Sheng, Shengliang Cai, Haochuan Cui et al.

A novel simulator called VMAgent is introduced to help RL researchers better explore new methods, especially for virtual machine scheduling. VMAgent is inspired by practical virtual machine (VM) scheduling tasks and provides an efficient simulation platform that can reflect the real situations of cloud computing. Three scenarios (fading, recovering, and expansion) are concluded from practical cloud computing and corresponds to many reinforcement learning challenges (high dimensional state and action spaces, high non-stationarity, and life-long demand). VMAgent provides flexible configurations for RL researchers to design their customized scheduling environments considering different problem features. From the VM scheduling perspective, VMAgent also helps to explore better learning-based scheduling solutions.

Wenhao Li, Xiangfeng Wang, Bo Jin et al.

Non-stationarity is one thorny issue in cooperative multi-agent reinforcement learning (MARL). One of the reasons is the policy changes of agents during the learning process. Some existing works have discussed various consequences caused by non-stationarity with several kinds of measurement indicators. This makes the objectives or goals of existing algorithms are inevitably inconsistent and disparate. In this paper, we introduce a novel notion, the $δ$-measurement, to explicitly measure the non-stationarity of a policy sequence, which can be further proved to be bounded by the KL-divergence of consecutive joint policies. A straightforward but highly non-trivial way is to control the joint policies' divergence, which is difficult to estimate accurately by imposing the trust-region constraint on the joint policy. Although it has lower computational complexity to decompose the joint policy and impose trust-region constraints on the factorized policies, simple policy factorization like mean-field approximation will lead to more considerable policy divergence, which can be considered as the trust-region decomposition dilemma. We model the joint policy as a pairwise Markov random field and propose a trust-region decomposition network (TRD-Net) based on message passing to estimate the joint policy divergence more accurately. The Multi-Agent Mirror descent policy algorithm with Trust region decomposition, called MAMT, is established by adjusting the trust-region of the local policies adaptively in an end-to-end manner. MAMT can approximately constrain the consecutive joint policies' divergence to satisfy $δ$-stationarity and alleviate the non-stationarity problem. Our method can bring noticeable and stable performance improvement compared with baselines in cooperative tasks of different complexity.

Wenhao Li, Xiangfeng Wang, Bo Jin et al.

When solving a complex task, humans will spontaneously form teams and to complete different parts of the whole task, respectively. Meanwhile, the cooperation between teammates will improve efficiency. However, for current cooperative MARL methods, the cooperation team is constructed through either heuristics or end-to-end blackbox optimization. In order to improve the efficiency of cooperation and exploration, we propose a structured diversification emergence MARL framework named {\sc{Rochico}} based on reinforced organization control and hierarchical consensus learning. {\sc{Rochico}} first learns an adaptive grouping policy through the organization control module, which is established by independent multi-agent reinforcement learning. Further, the hierarchical consensus module based on the hierarchical intentions with consensus constraint is introduced after team formation. Simultaneously, utilizing the hierarchical consensus module and a self-supervised intrinsic reward enhanced decision module, the proposed cooperative MARL algorithm {\sc{Rochico}} can output the final diversified multi-agent cooperative policy. All three modules are organically combined to promote the structured diversification emergence. Comparative experiments on four large-scale cooperation tasks show that {\sc{Rochico}} is significantly better than the current SOTA algorithms in terms of exploration efficiency and cooperation strength.

Junjie Sheng, Xiangfeng Wang, Bo Jin et al.

This work explores the large-scale multi-agent communication mechanism under a multi-agent reinforcement learning (MARL) setting. We summarize the general categories of topology for communication structures in MARL literature, which are often manually specified. Then we propose a novel framework termed as Learning Structured Communication (LSC) by using a more flexible and efficient communication topology. Our framework allows for adaptive agent grouping to form different hierarchical formations over episodes, which is generated by an auxiliary task combined with a hierarchical routing protocol. Given each formed topology, a hierarchical graph neural network is learned to enable effective message information generation and propagation among inter- and intra-group communications. In contrast to existing communication mechanisms, our method has an explicit while learnable design for hierarchical communication. Experiments on challenging tasks show the proposed LSC enjoys high communication efficiency, scalability, and global cooperation capability.