Yifan Qiao

Tian Xia, Hanchen Li, Zhifei Li et al.

Modern LLM serving systems increasingly host agentic workloads, whose sessions issue tens of model invocations interleaved with tool calls, accumulating KV cache that can be reused across steps. As requests' total KV cache size easily exceeds GPU HBM capacity, researchers offload them to CPU DRAM. However, tool-call durations span orders of magnitude, and the cost of transferring KV cache between tiers makes it impractical to re-place entries on every call. We observe that agentic programs exhibit a two-phase structure: busy phases of rapid short tool calls and idle phases dominated by long-running calls. Current eviction policies such as LRU fail to capture this property. A binary busy/idle label also falls short because the ratio of busy to idle programs may not match the hardware's GPU-to-CPU capacity ratio. When it does not, one tier sits underutilized while the other is oversubscribed, wasting memory or forcing unnecessary evictions. We present MORI, an agent serving system that solves the above problem. Our key insight is that idleness is a continuous, relative spectrum. MORI ranks all active programs by idleness, assigns the busiest to GPU HBM and the most idle to CPU DRAM, dynamically shifts the partition boundary to match hardware capacity, and enforces admission control at each memory tier. Evaluated on real coding agent workloads collected from Claude Code across four GPU and model pairs, MORI delivers 20--71% higher throughput and 18--43% lower TTFT than the best baseline with offloading.

John Thorpe, Pengzhan Zhao, Jonathan Eyolfson et al.

DNN models across many domains continue to grow in size, resulting in high resource requirements for effective training, and unpalatable (and often unaffordable) costs for organizations and research labs across scales. This paper aims to significantly reduce training costs with effective use of preemptible instances, i.e., those that can be obtained at a much cheaper price while idle, but may be preempted whenever requested by priority users. Doing so, however, requires new forms of resiliency and efficiency to cope with the possibility of frequent preemptions - a failure model that is drastically different from the occasional failures in normal cluster settings that existing checkpointing techniques target. We present Bamboo, a distributed system that tackles these challenges by introducing redundant computations into the training pipeline, i.e., whereby one node performs computations over not only its own layers but also over some layers in its neighbor. Our key insight is that training large models often requires pipeline parallelism where "pipeline bubbles" naturally exist. Bamboo carefully fills redundant computations into these bubbles, providing resilience at a low cost. Across a variety of widely used DNN models, Bamboo outperforms traditional checkpointing by 3.7x in training throughput, and reduces costs by 2.4x compared to a setting where on-demand instances are used.

Yongji Wu, Wenjie Qu, Xueshen Liu et al.

Sparsely-activated Mixture-of-Experts (MoE) architecture has increasingly been adopted to further scale large language models (LLMs). However, frequent failures still pose significant challenges as training scales. The cost of even a single failure is significant, as all GPUs need to idle wait until the failure is resolved, potentially losing considerable training progress as training has to restart from checkpoints. This problem is exacerbated by the growing use of spot instances on public clouds for model training, which despite offering substantial cost savings, introduce frequent preemptions-essentially failures that regularly occur throughout the training process. Existing solutions for efficient fault-tolerant training either lack elasticity or rely on building resiliency into pipeline parallelism, which cannot be applied to MoE models due to the expert parallelism strategy adopted by the MoE architecture. We present Lazarus, a system for resilient and elastic training of MoE models. Lazarus adaptively allocates expert replicas to address the inherent imbalance in expert workload and speeds up training, while a provably optimal expert placement algorithm is developed to maximize the probability of recovery upon failures. Through adaptive expert placement and a flexible token dispatcher, Lazarus can also fully utilize all available nodes after failures, leaving no GPU idle. Our evaluation shows that Lazarus outperforms existing MoE training systems by up to 5.7x under frequent node failures and 3.4x on a real spot instance trace.

Weixuan Ou, Yanzhao Zheng, Shuoshuo Sun et al.

Reinforcement Learning (RL) has been shown to improve the capabilities of large language models (LLMs). However, applying RL to open-domain tasks faces two key challenges: (1) the inherent subjectivity of these tasks prevents the verifiable rewards as required by Reinforcement Learning with Verifiable Rewards (RLVR); (2) Reinforcement Learning from Human Feedback (RLHF) relies on external reward mechanisms. To overcome these limitations, we propose Self-Examining Reinforcement Learning (SERL), a novel self-improving framework where the LLM serves as both Actor and Judge. SERL introduces two synergistic reward mechanisms without any external signals. On the one hand, to improve the Actor's capability, we derive rewards from Copeland-style pairwise comparison judgments across a group of generated responses. On the other hand, a self-consistency reward that encourages coherent judgments is proposed to improve the Judge's reliability. This process refines the Judge's capability, which in turn provides a more robust reward for Actor. Experiments show that our method outperforms existing self-improvement training methods. SERL improves the LC win rate of Qwen3-8B on AlpacaEval 2 from 52.37% to 59.90%. To the best of our knowledge, our method achieves state-of-the-art performance among self-improving approaches. Furthermore, it achieves a performance comparable to significantly larger models like Qwen3-32B, demonstrating superior effectiveness and robustness on open-domain tasks.

Shan Yu, Jiarong Xing, Yifan Qiao et al.

Serving large language models (LLMs) is expensive, especially for providers hosting many models, making cost reduction essential. The unique workload patterns of serving multiple LLMs (i.e., multi-LLM serving) create new opportunities and challenges for this task. The long-tail popularity of models and their long idle periods present opportunities to improve utilization through GPU sharing. However, existing GPU sharing systems lack the ability to adjust their resource allocation and sharing policies at runtime, making them ineffective at meeting latency service-level objectives (SLOs) under rapidly fluctuating workloads. This paper presents Prism, a multi-LLM serving system that unleashes the full potential of GPU sharing to achieve both cost efficiency and SLO attainment. At its core, Prism tackles a key limitation of existing systems$\unicode{x2014}$the lack of $\textit{cross-model memory coordination}$, which is essential for flexibly sharing GPU memory across models under dynamic workloads. Prism achieves this with two key designs. First, it supports on-demand memory allocation by dynamically mapping physical to virtual memory pages, allowing flexible memory redistribution among models that space- and time-share a GPU. Second, it improves memory efficiency through a two-level scheduling policy that dynamically adjusts sharing strategies based on models' runtime demands. Evaluations on real-world traces show that Prism achieves more than $2\times$ cost savings and $3.3\times$ SLO attainment compared to state-of-the-art systems.

Hongyin Zhang, Shiyuan Zhang, Junxi Jin et al.

Vision-Language-Action (VLA) models based on flow matching have shown excellent performance in general-purpose robotic manipulation tasks. However, the action accuracy of these models on complex downstream tasks is unsatisfactory. One important reason is that these models rely solely on the post-training paradigm of imitation learning, which makes it difficult to have a deeper understanding of the distribution properties of data quality, which is exactly what Reinforcement Learning (RL) excels at. In this paper, we theoretically propose an offline RL post-training objective for VLA flow models and induce an efficient and feasible offline RL fine-tuning algorithm -- Adaptive Reinforced Flow Matching (ARFM). By introducing an adaptively adjusted scaling factor in the VLA flow model loss, we construct a principled bias-variance trade-off objective function to optimally control the impact of RL signal on flow loss. ARFM adaptively balances RL advantage preservation and flow loss gradient variance control, resulting in a more stable and efficient fine-tuning process. Extensive simulation and real-world experimental results show that ARFM exhibits excellent generalization, robustness, few-shot learning, and continuous learning performance.

John Thorpe, Yifan Qiao, Jonathan Eyolfson et al.

A graph neural network (GNN) enables deep learning on structured graph data. There are two major GNN training obstacles: 1) it relies on high-end servers with many GPUs which are expensive to purchase and maintain, and 2) limited memory on GPUs cannot scale to today's billion-edge graphs. This paper presents Dorylus: a distributed system for training GNNs. Uniquely, Dorylus can take advantage of serverless computing to increase scalability at a low cost. The key insight guiding our design is computation separation. Computation separation makes it possible to construct a deep, bounded-asynchronous pipeline where graph and tensor parallel tasks can fully overlap, effectively hiding the network latency incurred by Lambdas. With the help of thousands of Lambda threads, Dorylus scales GNN training to billion-edge graphs. Currently, for large graphs, CPU servers offer the best performance-per-dollar over GPU servers. Just using Lambdas on top of CPU servers offers up to 2.75x more performance-per-dollar than training only with CPU servers. Concretely, Dorylus is 1.22x faster and 4.83x cheaper than GPU servers for massive sparse graphs. Dorylus is up to 3.8x faster and 10.7x cheaper compared to existing sampling-based systems.

Yingrui Yang, Yifan Qiao, Jinjin Shao et al.

Although considerable efforts have been devoted to transformer-based ranking models for document search, the relevance-efficiency tradeoff remains a critical problem for ad-hoc ranking. To overcome this challenge, this paper presents BECR (BERT-based Composite Re-Ranking), a composite re-ranking scheme that combines deep contextual token interactions and traditional lexical term-matching features. In particular, BECR exploits a token encoding mechanism to decompose the query representations into pre-computable uni-grams and skip-n-grams. By applying token encoding on top of a dual-encoder architecture, BECR separates the attentions between a query and a document while capturing the contextual semantics of a query. In contrast to previous approaches, this framework does not perform expensive BERT computations during online inference. Thus, it is significantly faster, yet still able to achieve high competitiveness in ad-hoc ranking relevance. Finally, an in-depth comparison between BECR and other start-of-the-art neural ranking baselines is described using the TREC datasets, thereby further demonstrating the enhanced relevance and efficiency of BECR.

Yifan Qiao, Chenyan Xiong, Zhenghao Liu et al.

This paper studies the performances and behaviors of BERT in ranking tasks. We explore several different ways to leverage the pre-trained BERT and fine-tune it on two ranking tasks: MS MARCO passage reranking and TREC Web Track ad hoc document ranking. Experimental results on MS MARCO demonstrate the strong effectiveness of BERT in question-answering focused passage ranking tasks, as well as the fact that BERT is a strong interaction-based seq2seq matching model. Experimental results on TREC show the gaps between the BERT pre-trained on surrounding contexts and the needs of ad hoc document ranking. Analyses illustrate how BERT allocates its attentions between query-document tokens in its Transformer layers, how it prefers semantic matches between paraphrase tokens, and how that differs with the soft match patterns learned by a click-trained neural ranker.