Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge.To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents.In this work, we show that these efficiency gains do not transfer uniformly to judge-centric inference.Across GSM8K, MMLU, and HumanEval, we find that reuse strategies that are effective for execution agents can severely perturb judge behavior: end-task accuracy may appear stable, yet the judge's selection becomes highly inconsistent with dense prefill.We quantify this risk using Judge Consistency Rate (JCR) and provide diagnostics showing that reuse systematically weakens cross-candidate attention, especially for later candidate blocks.Our ablation further demonstrates that explicit crosscandidate interaction is crucial for preserving dense-prefill decisions.Overall, our results identify a previously overlooked failure mode of KV cache reuse and highlight judge-centric inference as a distinct regime that demands dedicated, risk-aware system design.1

Semantic segmentation with dense pixel-wise annotation has achieved excellent performance thanks to deep learning. However, the generalization of semantic segmentation in the wild remains challenging. In this paper, we address the problem of unsupervised domain adaptation (UDA) in semantic segmentation. Motivated by the fact that source and target domain have invariant semantic structures, we propose to exploit such invariance across domains by leveraging co-occurring patterns between pairwise pixels in the output of structured semantic segmentation. This is different from most existing approaches that attempt to adapt domains based on individual pixel-wise information in image, feature, or output level. Specifically, we perform domain adaptation on the affinity relationship between adjacent pixels termed affinity space of source and target domain. To this end, we develop two affinity space adaptation strategies: affinity space cleaning and adversarial affinity space alignment. Extensive experiments demonstrate that the proposed method achieves superior performance against some state-of-the-art methods on several challenging benchmarks for semantic segmentation across domains. The code is available at https://github.com/idealwei/ASANet.

Most database systems rely on complex multi-layer and compatibility-oriented storage stacks, which results in sub-optimal database management system (DBMS) performance and significant write amplification. A heavy storage stack can be tolerated in the slow disk era because its storage overhead is completely overlapped by hardware delay. However, with advances in storage technologies, emerging NVMe devices have reached the same level of latency as software, which in turn has caused the storage stack to become a new bottleneck. On the other hand, NVMe devices not only improve I/O efficiency but also introduce distinctive hardware features that require software modifications to take advantage of themTo fully exploit the hardware potential of NVMe devices, we propose a lightweight native storage stack called Lightstack to minimize the software overhead. The core of Lightstack is an efficient table storage engine, LATTE, which abstracts the essential data service of the database's 2D table. LATTE is designed from the ground up to use NVMe devices efficiently. It directly accesses NVMe devices to reduce single I/O latency and utilizes a parallel scheduling strategy to leverage multiple deep I/O queues and CPU cores. Besides, undo logging on heterogeneous storage is proposed to mitigate the write amplification further. We also implement a working prototype and evaluate it with standard benchmarks on the Intel Optane DC P4800X NVMe SSD and the DC P3608 Series NVMe SSD. Experimental results show that LATTE has up to 3.6-6.5× the throughput of MySQL's InnoDB and MyRocks engines, with latency as low as 28% in the same hardware environment.

Learning robust and discriminative representations is essential for 3D object retrieval. In this paper, we present an improved Multi-view Convolutional Neural Network (MVCNN) for view-based 3D object representation learning. Our technical contributions are divided into two aspects. First, we propose to employ Group-view Similarity Learning (GSL) over the multi-view representations before the aggregation operation (i.e., max-pooling in MVCNN). We assume that the similarity information among the view groups of different 3D objects can provide an important cue but has been neglected more or less by previous methods. To enhance it, we add a branch to the original MVCNN architecture and learn to maintain such group-view similarity relationships. Second, we utilize an end-to-end metric learning loss function to improve the representation learning process. In particular, we propose an improved Triplet-Center Loss (TCL) named Adaptive Margin based Triplet-Center Loss (AMTCL). The original TCL assumes a fixed and common margin to control the relative distance relationship between a sample to its corresponding class center and to the nearest negative center. Though TCL has demonstrated its great capacity on the 3D object retrieval task, however, when considering the distinguishability between samples of one class and samples of another class, we assume that it would be more appropriate that the margin takes different values based on the distinguishability of samples of different classes. Therefore we propose to adaptively and dynamically adjust the margin hyperparameter based on the normalized confusion matrix which is obtained on the training set during the training process. Extensive experiments on several public 3D shape benchmarks show that our method, GSL + AMTCL, can learn more suitable representations for 3D object retrieval, obtaining superior performance against state-of-the-art methods.

Log replication is a key component in highly available database systems. In order to guarantee data consistency and reliability, it is common for modern database systems to utilize Paxos protocol, which is responsible for replicating transactional logs from one primary node to multiple backups. However, the Paxos replication needs to store and synchronize some additional metadata, such as committed log sequence number (commit point), to guarantee the consistency of the database. This increases the overhead of storage and network, which would have a negative impact on the throughput in the update intensive work load. In this paper, we present an implementation of log replication and database recovery methods, which adopts the idea of piggybacking, i.e., commit point can be embedded in the commit logs. This practice not only retains virtues of Paxos replication, but also reduces disk and network IO effectively. We implemented and evaluated our approach in a main memory database system. Our experiments show that the piggybacking method can offer 1.3× higher throughput than typical log replication with synchronization mechanism.