Raunak Agarwal, Markus A. Wenzel, Simon Baur, Jonas Zimmer, George Harvey, Jackie Ma. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

• Uncertainty estimation and XAI are both solutions to “black box” characteristics. • Use the model-agnostic XAI methods (i.e., SHAP) to explain AI model uncertainty. • Find that the estimated uncertainty comes from the distribution of training data. • Emphasize the uncertainty explanation is different from general model explanation. The “black box” characteristic of AI models has received widespread attention from scientists. This has promoted the development of explainable artificial intelligence (XAI) methods aimed at explaining AI model decisions, e.g., evaluating the importance of input features for the model output. In addition to XAI, another way to open the “black box” is uncertainty estimation of AI models (e.g., output ranges for regression tasks), which provides users with confidence in model decisions. However, little attention has been paid to the explanation of model uncertainty, e.g., which input features contribute to model uncertainty? This study takes SHapley Additive exPlanations (SHAP) as an example of perturbation-based XAI methods and utilizes it to explain AI model uncertainty by changing the model input to observe the variation of the estimated uncertainty. The experimental results in three cases (i.e., house price prediction, remote sensing image classification, and electrocardiogram classification) show that SHAP can effectively explain the uncertainty of AI models. Moreover, further analysis reveals that the uncertainty of AI models is affected by the frequency of occurrence of input features in the training data. Specifically, if a certain feature of the sample rarely appears in the training data, it will lead to greater uncertainty. This conclusion aligns with common sense and supports the validity of the uncertainty explanation.

In telecommunications, Autonomous Networks (ANs) automatically adjust configurations based on specific requirements (e.g., bandwidth) and available resources. These networks rely on continuous monitoring and intelligent mechanisms for self-optimization, self-repair, and self-protection, nowa-days enhanced by Neural Networks (NNs) to enable predictive modeling and pattern recognition.In this work, we propose a neural architecture called Mixed-Token Transformer (MTT) that can process and predict both numerical and textual tokens representing various mobile network features (e.g., ping, SNR or band frequency). We train the models using a Federated Learning (FL) approach, enabling multiple AN cells — each equipped with an MTT — to collaboratively learn from cellular data while preserving data privacy.Since FL involves frequent transmission of neural network updates, it requires an efficient, standardized compression strategy for reliable communication. To address this, we investigate NNCodec, our implementation of the ISO/IEC Neural Network Coding (NNC) standard. Our experimental results on the Berlin V2X dataset demonstrate that NNCodec achieves transparent compression (i.e., negligible performance loss) while reducing communication overhead to below 1%, showing the effectiveness of combining NNC with FL in collaboratively learned autonomous mobile networks. Our proposed MTT architecture speeds up model inference by ≈ 5×. Code is available at: github.com/Jostarndt/NNCodec-FL-MixedTokenTransformer.

Versatile Video Coding (VVC) is the latest video coding standard jointly developed by ITU-T VCEG and ISO/IEC MPEG. In this paper, technical details and experimental results for the VVC block partitioning structure are provided. Among all the new technical aspects of VVC, the block partitioning structure is identified as one of the most substantial changes relative to the previous video coding standards and provides the most significant coding gains. The new partitioning structure is designed using a more flexible scheme. Each coding tree unit (CTU) is either treated as one coding unit or split into multiple coding units by one or more recursive quaternary tree partitions followed by one or more recursive multi-type tree splits. The latter can be horizontal binary tree split, vertical binary tree split, horizontal ternary tree split, or vertical ternary tree split. A CTU dual tree for intra-coded slices is described on top of the new block partitioning structure, allowing separate coding trees for luma and chroma. Also, a new way of handling picture boundaries is presented. Additionally, to reduce hardware decoder complexity, virtual pipeline data unit constraints are introduced, which forbid certain multi-type tree splits. Finally, a local dual tree is described, which reduces the number of small chroma intra blocks.

In this paper, we describe a video coding design that enables a higher coding efficiency than the HEVC standard. The proposed video codec follows the design of block-based hybrid video coding, but includes a number of advanced coding tools. A part of the incorporated advanced concepts was developed by the Joint Video Exploration Team, while others are newly proposed. The key aspects of these newly proposed tools are the following. A video frame is subdivided into rectangles of variable size using a binary partitioning with variable split ratios. Three new approaches for generating spatial intra prediction signals are supported: A line-wise application of conventional intra prediction modes, coupled with a mode-dependent processing order, a region-based template matching prediction method and intra prediction modes based on neural networks. For motion-compensated prediction, a multi-hypothesis mode with more than two motion hypotheses can be used. In transform coding, mode dependent combinations of primary and secondary transforms are applied. Moreover, scalar quantization is replaced by trellis-coded quantization and the entropy coding of the quantized transform coefficients is improved. The intra and inter prediction signals can be filtered using an edge-preserving diffusion filter or a non-linear DCT-based thresholding operation. The video codec includes an adaptive in-loop filter for which one of three classifiers can be chosen on a picture basis. We also incorporated an optional encoder control, which adjusts the quantization parameters based on a perceptually motivated distortion measure. In a random access scenario, our proposed video codec achieves luma BD-rate savings between 32.5% for HDR HLG UHD and 39.6% for SDR UHD over the HEVC (HM software) anchor for different categories of test sequences.

One of the innovations in H.265/HEVC is the quad-tree partitioning framework. It allows flexible block subdivision and mode allocation across the encoded picture. The increased flexibility comes at a cost of vast search space expansion, making exhaustive search algorithms inapplicable. We propose a novel early termination condition to skip the exhaustive search of whole tree-branches in the well-established top-down encoding approach. The condition is based on a simple and intuitive split cost prediction. It can be parametrized to control the trade-off between the speed-up and caused BD-rate loss. Data driven parameter estimation and parameter number reduction is presented. For random-access encoding, the method can achieve an average speed-up of 30% with a BD-rate loss of 0.03%. At another trade-off point, speed-up is increased to over 40% for a BD-rate loss below 0.5%.

The upcoming Versatile Video Coding (VVC) Standard includes various new intra prediction tools that were not present in its predecessor High Efficiency Video Coding (HEVC), such as the wide angle intra prediction, the position dependent prediction combination or the multiple reference line intra prediction. In order to improve the intra prediction coding efficiency, this paper proposes the usage of the Intra Subpartition (ISP) algorithm. ISP is an updated version of the Line-Based Intra Prediction (LIP) mode that improves the trade-off between coding gain and complexity of the original method. The basic principle of ISP consists in subdividing an intra-predicted block into 2 or 4 subpartitions of at least 16 samples according to the original block dimensions. The method has been implemented on top of the VVC Test Model 3.0 (VTM-3.0), with the result of obtaining a gain of 0.57% with encoding and decoding run-times of 112% and 102% respectively for the All Intra configuration and a gain of 0.29% with encoding and decoding run-times of 102% and 101% respectively for the Random Access configuration.

One of the key technologies in the upcoming Versatile Video Coding (VVC) standard is the flexible partitioning. Compared to HEVC, partitioning alone accounts for about 8.5% bitrate savings. The added flexibility comes at a cost of large search space expansion.This paper gives an overview of fast encoder strategies for selecting the block partitioning in the Test Model VTM-3.0 for VVC. In total, 13 approaches for speeding up the block partitioning selection in an encoder are described; 8 of these techniques were developed by the authors. 10 of the 13 methods are used in the common test conditions of the Joint Video Experts Team (JVET); these 10 methods achieve an overall speed-up factor of 7 at a bit rate increase of only 1.1%. By enabling 12 approaches, a speed up factor of over 8.5 is obtained at a bit rate increase of about 1.5%.

Block partitioning is the basis of every modern hybrid video coding standard. It specifies how the video pictures can be subdivided into blocks for prediction and residual coding. In H.265/HEVC, quad-tree partitioning is one of the key technologies allowing for flexible mode allocation and providing a substantial part of the gains over H.264/AVC. The current draft of the upcoming standard Versatile Video Coding (VVC) provides over 30% bit-rate savings over HEVC and almost one third of the gain is achieved by using a more flexible partitioning scheme than the quad-tree partitioning used in HEVC. In this paper, we describe a partitioning concept that generalizes many of the ideas developed during the exploration and early standardization phase of VVC. In fact, our method includes the VVC partitioning as well as many other state-of-the-art methods. The proposed method can be implemented in a fully configurable design. For instance, it can be configured to match the performance of VTM-1.0 at much faster runtime (69%) or it can be configured to obtain additional bit-rate savings of up to 3% by exploiting additional degrees of freedom.