Yu Zhang, Songwei Liu, Chenqian Yan, Linsheng, Beichen Ning, Fangmin Chen, Xing Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The multi-objective multigraph Shortest Path Problem (SPP) is intractable, necessitating efficient solution approaches. To address general multi-objective multigraph SPPs, this article introduces a Multi-Objective Multi-Graph A* (MOMGA*) algorithm and develops a learning-based heuristic function to expedite the search. MOMGA* generalises the Airport Multi-Objective A* (AMOA*), which was designed for a specific application on multigraphs, and further modifies its path selection and expansion procedures. Theoretical analysis demonstrates that the modifications in MOMGA* yield advantages over AMOA*, including higher search efficiency, more effective use of admissible heuristics for accelerating search, and seamless integration with likely-admissible heuristics without sacrificing solution quality. The admissibility proof of MOMGA* is also provided. The developed heuristic function is likely-admissible. It embraces node embedding techniques to extract node characteristics, based on which shortest path costs (heuristics) for every two nodes are estimated through neural networks. In particular, we present an extensive review of walk-based shallow embedding methods and experimentally validate their superior ability in capturing the characteristics of nodes for accurately predicting heuristics. Evaluation based on randomly generated multi-objective multigraphs confirms: (i) MOMGA* comprehensively outperforms AMOA*, consistent with the theoretical analysis; (ii) walk-based sampling for node embeddings is key to preserving distance-related information in graphs; (iii) the proposed likely-admissible heuristics, even learnt with a limited amount of training data, can empower MOMGA* to efficiently obtain a collection of optimal and near-optimal solutions; and (iv) a good balance between optimality and tractability in MOMGA* is controllable by tuning the predictive accuracy of learning heuristics.

Large Language Models (LLMs) have revolutionized natural language processing tasks. However, their practical application is constrained by substantial memory and computational demands. Post-training quantization (PTQ) is considered an effective method to accelerate LLM inference. Despite its growing popularity in LLM model compression, PTQ deployment faces two major challenges. First, low-bit quantization leads to performance degradation. Second, restricted by the limited integer computing unit type on GPUs, quantized matrix operations with different precisions cannot be effectively accelerated. To address these issues, we introduce a novel arbitrary-bit quantization algorithm and inference framework, ABQ-LLM. It achieves superior performance across various quantization settings and enables efficient arbitrary-precision quantized inference on the GPU. ABQ-LLM introduces several key innovations: (1) a distribution correction method for transformer blocks to mitigate distribution differences caused by full quantization of weights and activations, improving performance at low bit-widths. (2) the bit balance strategy to counteract performance degradation from asymmetric distribution issues at very low bit-widths (e.g., 2-bit). (3) an innovative quantization acceleration framework that reconstructs the quantization matrix multiplication of arbitrary precision combinations based on BTC (Binary TensorCore) equivalents, gets rid of the limitations of INT4/INT8 computing units. ABQ-LLM can convert each component bit width gain into actual acceleration gain, maximizing performance under mixed precision(e.g., W6A6, W2A8). Based on W2*A8 quantization configuration on LLaMA-7B model, it achieved a WikiText2 perplexity of 7.59 (2.17⬇ vs 9.76 in AffineQuant). Compared to SmoothQuant, we realized 1.6x acceleration improvement and 2.7x memory compression gain.

This paper investigates heuristic initialisation for shortest path search on multi-objective multigraphs using genetic algorithms. An initialised solution comprises a node path and an edge path. The state-of-the-art initialisation method encodes node paths into random-key node priority sequences and assigns random values to represent the indices of traversed parallel edges, where the node priority is a combination of hop count and a random value. However, this method has three limitations. (i) Hop counts are computed using Dijkstra’s algorithm, incurring non-negligible computational cost. (ii) The upper bound of the random values added to hop counts is defined by a hyperparameter, requiring tedious parameter tuning. (iii) Heuristic information about edges is ignored, despite its potential to improve solution quality. To address these issues, this paper first eliminates the randomisation hyperparameter and constrains randomness to the range [0, 1). Second, a heuristic edge initialisation method is proposed. Finally, to reduce the computational burden of Dijkstra’s algorithm, a learning-based node priority generation method is devised using neural networks trained on node embeddings. Experiments on benchmark multi-objective multigraphs show that (i) combining hop counts with randomness in [0, 1) and heuristic edge initialisation outperforms the state-of-the-art method, and (ii) with limited training data, the learning-based node priorities surpass entirely random node priorities and are comparable to state-of-the-art performance, highlighting their potential in large-scale and dynamic scenarios.

Recent years have witnessed a few attempts of vision transformers for single image super-resolution (SISR). Since the high resolution of intermediate features in SISR models increases memory and computational requirements, efficient SISR transformers are more favored. Based on some popular transformer backbone, many methods have explored reasonable schemes to reduce the computational complexity of the self-attention module while achieving impressive performance. However, these methods only focus on the performance on the training platform (e.g., Pytorch/Tensorflow) without further optimization for the deployment platform (e.g., TensorRT). Therefore, they inevitably contain some redundant operators, posing challenges for subsequent deployment in real-world applications. In this paper, we propose a deployment-friendly transformer unit, namely UFONE (i.e., UnFolding ONce is Enough), to alleviate these problems. In each UFONE, we introduce an Inner-patch Transformer Layer (ITL) to efficiently reconstruct the local structural information from patches and a Spatial-Aware Layer (SAL) to exploit the long-range dependencies between patches. Based on UFONE, we propose a Deployment-friendly Inner-patch Transformer Network (DITN) for the SISR task, which can achieve favorable performance with low latency and memory usage on both training and deployment platforms. Furthermore, to further boost the deployment efficiency of the proposed DITN on TensorRT, we also provide an efficient substitution for layer normalization and propose a fusion optimization strategy for specific operators. Extensive experiments show that our models can achieve competitive results in terms of qualitative and quantitative performance with high deployment efficiency.

Routing and scheduling of airport ground movements poses a critical issue for efficient surface operations. For real-world applications, multiple objectives should be considered, leading to a multigraph representation of the search space. Meanwhile, intermediate holding is often needed to a) release availability of scarce resources such as runways and gates for more cost-effective routing and scheduling, b) provide additional solutions in the speed profile database that can be used during routing and scheduling, and c) keep airport ground movements functional during disruptive events that may paralyse part of the taxiway network. This paper presents an extended multiobjective memetic algorithm upon the multigraph model to search for desirable solutions with intermediate holding. The performance of the proposed algorithm is examined with problem instances of different airport layouts. The results demonstrate prominent savings in both time and fuel costs compared with solutions without intermediate holding.

To solve general multi-objective multigraph shortest path problems, this paper proposes an algorithm (MOMGA*) that incorporates an online likely-admissible learning-based heuristic function to accelerate the solution-finding process. MOMGA* is an extended and generalised version of the airport multi-objective A* (AMOA*) algorithm that is tailored for a specific application problem. The online heuristic function is added and developed using artificial neural networks that estimate the costs between two nodes based on their metrics. To implement this metric-based prediction, a graph embedding technique is adopted to learn node feature representations. Results on a range of benchmark multi-objective multigraphs show that (i) in the absence of heuristic information, MOMGA* can deliver the same Pareto optimal solutions as AMOA* does, while requiring less computational time, and (ii) empowered by the likely-admissible learning- based heuristics, MOMGA* is able to provide a set of optimal and near-optimal solutions and strike a good balance between optimality and tractability.