Baidu (China)

As a prevailing post-processing tool for image retrieval, re-ranking typically leverages high-confidence retrieved samples to refine the initial ranking list. However, a significant challenge persists: existing re-ranking methods are computationally intensive, leading to impractical time costs for real-world applications. To overcome this limitation, we first review current re-ranking techniques and then introduce a novel real-time approach by reformulating the re-ranking process as a highly-parallel Graph Neural Network (GNN). Specifically, we decompose the traditional neighbor-based re-ranking into two distinct stages: retrieving high-quality gallery samples and refining neighbor similarity. We propose that the first stage can be effectively replaced by constructing a k-nearest neighbor (knn) graph, while the second stage can be implemented by propagating messages within this graph. In practice, since the knn graph is sparse, the GNN only considers a limited number of vertices and their connected edges, enabling efficient updates of vertex features. We validate the effectiveness and efficiency of our approach through extensive experiments on five datasets. For example, on the Market-1501 dataset, our method accelerates the re-ranking process to 9.4ms using a single K40m GPU. Additionally, we observe similar acceleration results on the other four retrieval benchmarks, Paris-6k, Oxford-5k, VeRi-776, and University-1652, while maintaining competitive performance.

ABSTRACT When utilizing the finite element method for simulating SAW devices, a large‐scale complex linear system of equations is generated. The efficiency of device simulation primarily depends on the storage requirement and computational speed for solving the linear system. However, the convergence rate of an iterative solver is confined by the ill‐conditioned nature of the coefficient matrix, which is non‐Hermitian and indefinite. In this paper, we propose several approximations for the Schur complement matrix. These approximations can be generated by the finite element method instead of matrix‐matrix multiplication and matrix inversion operations. Based on them, we develop the lower triangular preconditioners for the coefficient matrix and analyze the eigenvalue distribution of the preconditioned matrices as well. Theoretical analysis reveals that the preconditioned matrix is diagonalizable, with its eigenvalues confined to a small rectangular region, thus ensuring rapid convergence of iterative solvers. The theoretical considerations are supported by numerical results.

The integration of dynamic, sparse structures like Mixture-of-Experts (MoE) with parameter-efficient adapters (e.g., LoRA) is a powerful technique for enhancing Large Language Models (LLMs). However, this architectural enhancement comes at a steep cost: despite minimal increases in computational load, the inference latency often skyrockets, leading to decoding speeds slowing by over 2.5 times. Through a fine-grained performance analysis, we pinpoint the primary bottleneck not in the computation itself, but in the severe overhead from fragmented, sequential CUDA kernel launches required for conventional dynamic routing. To address this challenge, we introduce AdaFuse, a framework built on a tight co-design between the algorithm and the underlying hardware system to enable efficient dynamic adapter execution. Departing from conventional layer-wise or block-wise routing, AdaFuse employs a token-level pre-gating strategy, which makes a single, global routing decision for all adapter layers before a token is processed. This ``decide-once, apply-everywhere'' approach effectively staticizes the execution path for each token, creating an opportunity for holistic optimization. We capitalize on this by developing a custom CUDA kernel that performs a fused switching operation, merging the parameters of all selected LoRA adapters into the backbone model in a single, efficient pass. Experimental results on popular open-source LLMs show that AdaFuse achieves accuracy on par with state-of-the-art dynamic adapters while drastically cutting decoding latency by a factor of over 2.4x, thereby bridging the gap between model capability and inference efficiency.

With the rapid advancement of large language models (LLMs), retrieval-augmented generation (RAG) has emerged as a critical approach to supplement the inherent knowledge limitations of LLMs. However, due to the typically large volume of retrieved information, RAG tends to operate with long context lengths. From the perspective of entropy engineering, we identify unconstrained entropy growth and attention dilution due to long retrieval context as significant factors affecting RAG performance. In this paper, we propose the balanced entropy-engineered RAG (BEE-RAG) framework, which improves the adaptability of RAG systems to varying context lengths through the principle of entropy invariance. By leveraging balanced context entropy to reformulate attention dynamics, BEE-RAG separates attention sensitivity from context length, ensuring a stable entropy level. Building upon this, we introduce a zero-shot inference strategy for multi-importance estimation and a parameter-efficient adaptive fine-tuning mechanism to obtain the optimal balancing factor for different settings. Extensive experiments across multiple RAG tasks demonstrate the effectiveness of BEE-RAG.

While large language models (LLMs) demonstrate emerging reasoning capabilities, current inference-time expansion methods incur prohibitive computational costs through exhaustive sampling. Through analyzing decoding trajectories, we observe that most next-token predictions align well with the golden output, except for a few critical tokens that lead to deviations. Inspired by this phenomenon, we propose a novel Hint-Practice Reasoning (HPR) framework that operationalizes this insight through two synergistic components: 1) a hinter (powerful LLM) that provides probabilistic guidance at critical decision points, and 2) a practitioner (efficient smaller model) that executes major reasoning steps. The framework's core innovation lies in Distributional Inconsistency Reduction (DIR), a theoretically-grounded metric that dynamically identifies intervention points by quantifying the divergence between practitioner's reasoning trajectory and hinter's expected distribution in a tree-structured probabilistic space. Through iterative tree updates guided by DIR, HPR reweights promising reasoning paths while deprioritizing low-probability branches. Experiments across arithmetic and commonsense reasoning benchmarks demonstrate HPR's state-of-the-art efficiency-accuracy tradeoffs: it achieves comparable performance to self-consistency and MCTS baselines while decoding only 1/5 tokens, and outperforms existing methods by at most 5.1% absolute accuracy while maintaining similar or lower FLOPs.