Qi Jia, Ye Shen, Xiujie Song, Kaiwei Zhang, Shibo Wang, Dun Pei, Xiangyang Zhu, Guangtao Zhai. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent advancements in large language models (LLMs) have shifted focus toward scaling inference-time compute-improving performance without retraining the model.A common approach is to sample multiple outputs in parallel, and select one of these as the final output.While existing work has focused on English and specific domains, we study how to robustly scale inference-time compute in a multilingual, multi-task setting: spanning open-ended generations, math and translation tasks, for open models at 8B and 111B scale, across seven languages.Our findings highlight the need for tailored sampling and selection strategies.We propose novel solutions tailored for this multi-faceted inference scenario, demonstrating notable gains across languages and tasks.Our methods achieve an average +6.8 jump in win-rates for 8B models on m-ArenaHard-v2.0prompts in non-English languages against proprietary models like Gemini.At larger scale, our 111B model shows a +9.0 improvement with just five samples compared to single-sample decoding.These results emphasize the importance of language-and taskaware approaches to democratize inferencetime improvements.

In complex road environments, fog and low-light conditions degrade visible image quality, limiting detection systems. The integration of infrared (IR) imaging in these scenarios offers superior penetration capabilities and serves as a valuable complement to visible light. However, existing muiltispectral detectors struggle with modality imbalance and insufficient feature complementarity. To address these issues, a novel multispectral detection framework AT-DETR is proposed in this paper, with a two-stage training strategy and two essential modules: the adaptive feature alignment and fusion (AFAF) module and the three-branch dynamic parameter enhancement (TDPE) module. The AFAF module adaptively assigns fusion weights to balance the contribution of different modalities, while the TDPE module enhances the quality of degraded images through lightweight convolutional parameter estimation. The two-stage training strategy allows the network to initially prioritize feature extraction and subsequently optimize scene quality. The experimental results demonstrate that AT-DETR can improve [email protected] by 7.1/8.6 on the combined dataset of the$\mathrm{M}^{3} \text{FD}$and MSRS respectively, compared with using a single-source detector. Furthermore, its strong generalization capabilities are validated through tests on the FLIR_Aligned dataset. Implementation code and pre-trained weights can be accessed at https://github.com/JianDaoCX/ATDETR.

INTRODUCTION: Influenza continuously evolves to escape population immunity, which makes formulating a vaccine challenging. Antigenic differences between vaccine strains and circulating strains can affect vaccine effectiveness (VE). Quantifying the antigenic difference between vaccine strains and circulating strains can aid interpretation of VE, and several antigenic distance metrics have been discussed in the literature. Here, we compare how the predicted breadth of vaccine-induced antibody response varies when different metrics are used to calculate antigenic distance. METHODS: We analyzed data from a seasonal influenza vaccine cohort that collected serum samples from 2013/14 - 2017/18 at three study sites. The data include pre- and post-vaccination HAI titers to the vaccine strains and a panel of heterologous strains. We used that data to calculate four different antigenic distance measures between assay strains and vaccine strains: difference in year of isolation (temporal), p-Epitope (sequence), Grantham's distance (biophysical), and antigenic cartography distance (serological). We analyzed agreement between the four metrics using Spearman's correlation and intraclass correlation. We then fit Bayesian generalized additive mixed-effects models to predict the effect of antigenic distance on post-vaccination titer after controlling for confounders and analyzed the pairwise difference in predictions between metrics. RESULTS: The four antigenic distance metrics had low or moderate correlation for influenza subtypes A(H1N1), B/Victoria, and B/Yamagata. A(H3N2) distances were highly correlated. We found that after accounting for pre-vaccination titer, study site, and repeated measurements across individuals, the predicted post-vaccination titers conditional on antigenic distance and subtype were nearly identical across antigenic distance metrics, with A(H3N2) showing the only notable deviation between metrics, despite higher agreement for that subtype. DISCUSSION: Despite moderate correlation among metrics, we found that different antigenic distance metrics generated similar predictions about breadth of vaccine response. Costly titer assays for antigenic cartography may not be needed when simpler sequence-based metrics suffice for quantifying vaccine breadth.

Taking human-robot collaborative assembly as an example, the methods based on contact forces can improve the assembly efficiency of industrial robots with large components in industrial manufacturing. However, due to the large size, high payload, assembly accuracy and dynamic changes in grip position, accurately estimating the contact forces between the payload and the operator becomes challenging when handling these large components. In this paper, a two-stage method is proposed for payload dynamic parameter identification. The parameter identification equation in the sensor coordinate system is initially established. Furthermore, the identification model of recursive restricted total least squares (RRTLS) based on total least squares (TLS) is constructed to achieve low-consumption online identification. According to the assembly requirements and payload characteristics, the posture coordinate system is designed for safety, including the feasible workspace for the robot. Subsequently, the static identification postures and dynamic excitation trajectory are planned to obtain static values and dynamic inertial parameters. In the end, a high-payload human-robot collaborative assembly system is built to validate the proposed method. Experimental results show that compared with the existing methods, the proposed approach can effectively identify and compensate the payload, leading to more accurate external force sensing.

The video surveillance system based on face analysis has played an increasingly important role in the security industry. Compared with identification methods of other physical characteristics, face verification method is easy to be accepted by people. In the video surveillance scene, it is common to capture multiple faces belonging to a same person. We cannot get a good result of face recognition if we use all the images without considering image quality. In order to solve this problem, we propose a face deduplication system which is combined with face detection and face quality evaluation to obtain the highest quality face image of a person. The experimental results in this paper also show that our method can effectively detect the faces and select the high-quality face images, so as to improve the accuracy of face recognition.

With the large amount of distributed generation and diversity load access to the power grid, the changes of distribution network structure and operating characteristics put forward new requirements for its control. Improving the safety and stability of the distribution network has been the important goal of the current distribution network construction. In order to realize the interaction between source and load, an interruptible load model based on the optimal price of electricity is established by analyzing the relationship between the price and the interruptible load. The optimal power flow calculation method of power system with distributed power supply is proposed combined with this model, and its feasibility is verified by simulation.

Accurate mass segmentation on mammograms is a critical step in computer-aided diagnosis (CAD) systems. It is also a challenging task since some of the mass lesions are embedded in normal tissues and possess poor contrast or ambiguous margins. Besides, the shapes and densities of masses in mammograms are various. In this paper, a hybrid method combining a random walks algorithm and Chan-Vese (CV) active contour is proposed for automatic mass segmentation on mammograms. The data set used in this study consists of 1095 mass regions of interest (ROIs). First, the original ROI is preprocessed to suppress noise and surrounding tissues. Based on the preprocessed ROI, a set of seed points is generated for initial random walks segmentation. Afterward, an initial contour of mass and two probability matrices are produced by the initial random walks segmentation. These two probability matrices are used to modify the energy function of the CV model for prevention of contour leaking. Lastly, the final segmentation result is derived by the modified CV model, during which the probability matrices are updated by inserting several rounds of random walks. The proposed method is tested and compared with other four methods. The segmentation results are evaluated based on four evaluation metrics. Experimental results indicate that the proposed method produces more accurate mass segmentation results than the other four methods.