Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration, such as evaluation batch size, GPU count, and GPU version, can introduce significant differences in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9\% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision—while critical for reproducibility—is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

While recent depth completion methods have achieved remarkable results filling in relatively dense depth maps (e.g., projected 64-line LiDAR on KITTI or 500 sampled points on NYUv2) with RGB guidance, their performance on very sparse input (e.g., 4-line LiDAR or 32 depth point measurements) is unverified. These sparser regimes present new challenges, as a 4-line LiDAR increases the distance between pixels without depth and their nearest depth point sixfold from 5 pixels to 30 pixels compared to 64 lines. Ob-serving that existing methods struggle with sparse and variable distribution depth maps, we propose an Affinity-Based Shift Correction (ASC) module that iteratively aligns depth predictions to input depth based on predicted affinities between image pixels and depth points. Our framework enables each depth point to adaptively influence and improve predictions across the image, leading to largely improved results for fewer-line, fewer-point, and variable sparsity settings. Further, we show improved performance in domain transfer from KITTI to nuScenes andfrom random sampling to irregular point distributions. Our correction module can easily be added to any depth completion or RGB-only depth estimation model, notably allowing the latter to perform both completion and estimation with a single model.

We tackle the task of Azimuth (angular dimension) super-resolution for Frequency Modulated Continuous Wave (FMCW) multiple-input multiple-output (MIMO) radar. FMCW MIMO radar is widely used in autonomous driving alongside Lidar and RGB cameras. However, compared to Lidar, MIMO radar is usually of low resolution due to hardware size restrictions. For example, achieving 1° azimuth resolution requires at least 100 receivers, but a single MIMO device usually supports at most 12 receivers. Having limitations on the number of receivers is problematic since a high-resolution measurement of azimuth angle is essential for estimating the location and velocity of objects. To improve the azimuth resolution of MIMO radar, we propose a light, yet efficient, Analog-to-Digital super-resolution model (ADC-SR) that predicts or hallucinates additional radar signals using signals from only a few receivers. Compared with the baseline models that are applied to processed radar Range-Azimuth-Doppler (RAD) maps, we show that our ADC-SR method that processes raw ADC signals achieves comparable performance with 98% (50 times) fewer parameters. We also propose a hybrid super-resolution model (Hybrid-SR) combining our ADC-SR with a standard RAD super-resolution model, and show that performance can be improved by a large margin. Experiments on our Pitt-Radar dataset and the RADIal dataset validate the importance of leveraging raw radar ADC signals. To assess the value of our super-resolution model for autonomous driving, we also perform object detection on the results of our super-resolution model and find that our super-resolution model improves detection performance by around 4% in mAP. The Pitt-Radar and the code will be released at the link.

We aim to improve the performance of the vehicle detection model with Lidar-Radar fusion and data augmentation. The recent works for Lidar-Radar fusion such as MVDNet or ST-MVDNet, have been proposed to have effective performance in detecting vehicles, and address the issue regarding missing modality. However, there are few works applying some global data augmentations such as rotation, translation, and scaling which are common for Lidar-only model. In order to further improve the previous Lidar-Radar fusion model, we propose a model named ST-MVDNet++ by leveraging the self-training teacher-student framework with integrating more common data augmentations such as global rotation, translation, and scaling. To ensure the data augmentations are consistent and matched across Lidar and Radar, we apply the augmentations on bird-eye-view coordinates. We also introduce the student-only augmentation for robust training of the student model with the consistency loss from teacher model. We demonstrate that our leveraging of global consistent Lidar-Radar augmentation improve the previous works by 1 ∼ 2% in all of the experimental settings.

We address the task of domain adaptation in object detection, where there is an obvious domain gap between a domain with annotations (source) and a domain of interest without annotations (target). As a popular semi-supervised learning method, the teacher-student framework (a student model is supervised by the pseudo labels from a teacher model) has also yielded a large accuracy gain in cross-domain object detection. However, it suffers from the domain shift and generates many low-quality pseudo labels (e.g., false positives), which leads to sub-optimal performance. To mitigate this problem, we propose a teacher-student framework named Adaptive Teacher (AT) which leverages domain adversarial learning and weak-strong data augmentation to address the domain gap. Specifically, we employ feature-level adversarial training in the student model, allowing features derived from the source and target domains to share similar distributions. This process ensures the student model produces domain-invariant features. Furthermore, we apply weak-strong augmentation and mutual learning between the teacher model (taking data from the target domain) and the student model (taking data from both domains). This enables the teacher model to learn the knowledge from the student model without being biased to the source domain. We show that AT demonstrates superiority over existing approaches and even Oracle (fully-supervised) models by a large margin. For example, we achieve 50.9% (49.3%) mAP on Foggy Cityscape (Cli-part1K), which is 9.2% (5.2%) and 8.2% (11.0%) higher than previous state-of-the-art and Oracle, respectively.

Fusion of multiple sensor modalities such as camera, Lidar, and Radar, which are commonly found on autonomous vehicles, not only allows for accurate detection but also robustifies perception against adverse weather conditions and individual sensor failures. Due to inherent sensor characteristics, Radar performs well under extreme weather conditions (snow, rain, fog) that significantly degrade camera and Lidar. Recently, a few works have developed vehicle detection methods fusing Lidar and Radar signals, i.e., MVD-Net. However, these models are typically developed under the assumption that the models always have access to two error-free sensor streams. If one of the sensors is unavailable or missing, the model may fail catastrophically. To mitigate this problem, we propose the Self-Training Multimodal Vehicle Detection Network (ST-MVDNet) which leverages a Teacher-Student mutual learning framework and a simulated sensor noise model used in strong data augmentation for Lidar and Radar. We show that by (1) enforcing output consistency between a Teacher network and a Student network and by (2) introducing missing modalities (strong augmentations) during training, our learned model breaks away from the error-free sensor assumption. This consistency enforcement enables the Student model to handle missing data properly and improve the Teacher model by updating it with the Student model's exponential moving average. Our experiments demonstrate that our proposed learning framework for multi-modal detection is able to better handle missing sensor data during inference. Furthermore, our method achieves new state-of-the-art performance (5% gain) on the Oxford Radar Robotcar dataset under various evaluation settings.

We address the problem of estimating the 3D pose of a network of cameras for large-environment wide-baseline scenarios, e.g., cameras for construction sites, sports stadiums, and public spaces. This task is challenging since detecting and matching the same 3D keypoint observed from two very different camera views is difficult, making standard structure-from-motion (SfM) pipelines inapplicable. In such circumstances, treating people in the scene as "keypoints" and associating them across different camera views can be an alternative method for obtaining correspondences. Based on this intuition, we propose a method that uses ideas from person re-identification (re-ID) for wide-baseline camera calibration. Our method first employs a re-ID method to associate human bounding boxes across cameras, then converts bounding box correspondences to point correspondences, and finally solves for camera pose using multi-view geometry and bundle adjustment. Since our method does not require specialized calibration targets except for visible people, it applies to situations where frequent calibration updates are required. We perform extensive experiments on datasets captured from scenes of different sizes (80m2, 350m2, 600m2), camera settings (indoor and outdoor), and human activities (walking, playing basketball, construction). Experiment results show that our method achieves similar performance to standard SfM methods relying on manually labeled point correspondences.

We address the task of Re-Identification (Re-ID) in multi-target multi-camera (MTMC) tracking where we track multiple pedestrians using multiple overlapping uncalibrated (unknown pose) cameras. Since the videos are temporally synchronized and spatially overlapping, we can see a person from multiple views and associate their trajectory across cameras. In order to find the correct association between pedestrians visible from multiple views during the same time window, we extract a visual feature from a tracklet (sequence of pedestrian images) that encodes its similarity and dissimilarity to all other candidate tracklets. We propose a inter-tracklet (person to person) attention mechanism that learns a representation for a target tracklet while taking into account other tracklets across multiple views. Furthermore, to encode the gait and motion of a person, we introduce second intra-tracklet (person-specific) attention module with position embeddings. This second module employs a transformer encoder to learn a feature from a sequence of features over one tracklet. Experimental results on WILDTRACK and our new dataset ‘ConstructSite’ confirm the superiority of our model over state-of-the-art ReID methods (5% and 10% performance gain respectively) in the context of uncalibrated MTMC tracking. While our model is designed for overlapping cameras, we also obtain state-of-the-art results on two other benchmark datasets (MARS and DukeMTMC) with non-overlapping cameras.

Detecting small cracks in concrete is difficult due to the complexity and thinness of cracking patterns, which requires the development of refined vision-based segmentation algorithms that can accurately characterize the details of crack defects. While existing methods are good at generally outlining cracks, due to inherent differences in shape distributions between common objects and cracks, their predictions often have disconnected segments and inaccuracy along boundaries. To this end, we develop a refinement framework using reinforcement learning (RL) that can better recognize details specific to cracks. Our method uses an RL agent to iteratively improve per-pixel crack predictions of a general segmentation model. We find that in addition to connecting gaps in predictions, the RL agent is also able to detect cracks that are missed in the original predictions. It does so by using the originally detected regions as crack priors to branch out from. Refining outputs of a commonly used per-pixel segmentation model, our method outperforms the current state-of-the-art approaches for crack segmentation. Our experiments also demonstrate that our method generalizes well to a similar task of vessel segmentation.

Person re-identification (re-ID) aims to recognize instances of the same person contained in multiple images taken across different cameras. Existing methods for re-ID tend to rely heavily on the assumption that both query and gallery images of the same person have the same clothing. Unfortunately, this assumption may not hold for datasets captured over long periods of time. To tackle the re-ID problem in the context of clothing changes, we propose a novel representation learning method which is able to generate a shape-based feature representation that is invariant to clothing. We call our model the Clothing Agnostic Shape Extraction Network (CASE-Net). CASE-Net learns a representation of a person that depends primarily on shape via adversarial learning and feature disentanglement. Quantitative and qualitative results across 5 datasets (Div-Market, Market1501, three large-scale datasets under clothing changes) show our approach makes significant improvements over prior state-of-the-art approaches.

Person re-identification (re-ID) solves the task of matching images across cameras and is among the research topics in vision community. Since query images in real-world scenarios might suffer from resolution loss, how to solve the resolution mismatch problem during person re-ID becomes a practical problem. Instead of applying separate image super-resolution models, we propose a novel network architecture of Resolution Adaptation and re-Identification Network (RAIN) to solve cross-resolution person re-ID. Advancing the strategy of adversarial learning, we aim at extracting resolution-invariant representations for re-ID, while the proposed model is learned in an end-to-end training fashion. Our experiments confirm that the use of our model can recognize low-resolution query images, even if the resolution is not seen during training. Moreover, the extension of our model for semi-supervised re-ID further confirms the scalability of our proposed method for real-world scenarios and applications.

Few-shot learning (FSL) requires one to learn from object categories with a small amount of training data (as novel classes), while the remaining categories (as base classes) contain a sufficient amount of data for training. It is often desirable to transfer knowledge from the base classes and derive dominant features efficiently for the novel samples. In this work, we propose a sampling method that de-correlates an image based on maximum entropy reinforcement learning, and extracts varying sequences of patches on every forward-pass with discriminative information observed. This can be viewed as a form of "learned" data augmentation in the sense that we search for different sequences of patches within an image and performs classification with aggregation of the extracted features, resulting in improved FSL performances. In addition, our positive and negative sampling policies along with a newly defined reward function would favorably improve the effectiveness of our model. Our experiments on two benchmark datasets confirm the effectiveness of our framework and its superiority over recent FSL approaches.

Audio-visual event localization requires one to identify the event which is both visible and audible in a video (either at a frame or video level). To address this task, we propose a deep neural network named Audio-Visual sequence-to-sequence dual network (AVSDN). By jointly taking both audio and visual features at each time segment as inputs, our proposed model learns global and local event information in a sequence to sequence manner, which can be realized in either fully supervised or weakly supervised settings. Empirical results confirm that our proposed method performs favorably against recent deep learning approaches in both settings.

Person re-identification (re-ID) aims at matching images of the same identity across camera views. Due to varying distances between cameras and persons of interest, resolution mismatch can be expected, which would degrade person re-ID performance in real-world scenarios. To overcome this problem, we propose a novel generative adversarial network to address cross-resolution person re-ID, allowing query images with varying resolutions. By advancing adversarial learning techniques, our proposed model learns resolution-invariant image representations while being able to recover the missing details in low-resolution input images. The resulting features can be jointly applied for improving person re-ID performance due to preserving resolution invariance and recovering re-ID oriented discriminative details. Our experiments on five benchmark datasets confirm the effectiveness of our approach and its superiority over the state-of-the-art methods, especially when the input resolutions are unseen during training.

Person re-identification (re-ID) aims at recognizing the same person from images taken across different cameras. To address this challenging task, existing re-ID models typically rely on a large amount of labeled training data, which is not practical for real-world applications. To alleviate this limitation, researchers now targets at cross-dataset re-ID which focuses on generalizing the discriminative ability to the unlabeled target domain when given a labeled source domain dataset. To achieve this goal, our proposed Pose Disentanglement and Adaptation Network (PDA-Net) aims at learning deep image representation with pose and domain information properly disentangled. With the learned cross-domain pose invariant feature space, our proposed PDA-Net is able to perform pose disentanglement across domains without supervision in identities, and the resulting features can be applied to cross-dataset re-ID. Both of our qualitative and quantitative results on two benchmark datasets confirm the effectiveness of our approach and its superiority over the state-of-the-art cross-dataset Re-ID approaches.

Video summarization still remains a challenging task. Due to sufficient video data on the Internet, such task draws significant attention in the vision community and benefits a wide range of applications, e.g., video retrieval, search, etc. To effectively perform video summarization by deriving the keyframes which represent the given input video, we propose a novel framework named Hierarchical Multi-Attention Network (H-MAN) which comprises the shot-level reconstruction model and multi-head attention model. While our designed attention model is two-stage hierarchical structure for producing various attention maps, we are among the first to utilize the multi-attention mechanism in the video summarization task, which brings improved performance. The quantitative and qualitative results demonstrate the effectiveness of our model, which performs favorably against state-of-the-art approaches.

Deep reinforcement learning has shown its success in game playing. However, 2.5D fighting games would be a challenging task to handle due to ambiguity in visual appearances like height or depth of the characters. Moreover, actions in such games typically involve particular sequential action orders, which also makes the network design very difficult. Based on the network of Asynchronous Advantage Actor-Critic (A3C), we create an OpenAI-gym-like gaming environment with the game of Little Fighter 2 (LF2), and present a novel A3C+ network for learning RL agents. The introduced model includes a Recurrent Info network, which utilizes game-related info features with recurrent layers to observe combo skills for fighting. In the experiments, we consider LF2 in different settings, which successfully demonstrates the use of our proposed model for learning 2.5D fighting games.

Cyber security has grown up to be a hot issue in recent years. How to identify potential malware becomes a challenging task. To tackle this challenge, we adopt deep learning approaches and perform flow detection on real data. However, real data often encounters an issue of imbalanced data distribution which will lead to a gradient dilution issue. When training a neural network, this problem will not only result in a bias toward the majority class but show the inability to learn from the minority classes. In this paper, we propose a Tree-Shaped Deep Neural Network (TSDNN) which classifies the data in a layer-wise manner. To better learn from the minority classes, we propose a Quantity Dependent Backpropagation (QDBP) algorithm which incorporates the knowledge of the disparity between classes. We evaluate our method on an imbalanced data set. Experimental result demonstrates that our approach outperforms the state-of-the-art methods and justifies that the proposed method is able to overcome the difficulty of imbalanced learning. We also conduct a partial flow experiment which shows the feasibility of realtime detection and a zero-shot learning experiment which justifies the generalization capability of deep learning in cyber security.