MiniCPM4: Ultra-Efficient LLMs on End Devices

This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Sufficient evaluation results show that MiniCPM4 outperforms open-source models of similar size across multiple benchmarks, highlighting both its efficiency and effectiveness. Notably, MiniCPM4-8B demonstrates significant speed improvements over Qwen3-8B when processing long sequences. Through further adaptation, MiniCPM4 successfully powers diverse applications, including trustworthy survey generation and tool use with model context protocol, clearly showcasing its broad usability.

ToLeaP: Rethinking Development of Tool Learning with Large Language Models

Tool learning, which enables large language models (LLMs) to utilize external tools effectively, has garnered increasing attention for its potential to revolutionize productivity across industries. Despite rapid development in tool learning, key challenges and opportunities remain understudied, limiting deeper insights and future advancements. In this paper, we investigate the tool learning ability of 41 prevalent LLMs by reproducing 33 benchmarks and enabling one-click evaluation for seven of them, forming a Tool Learning Platform named ToLeaP. We also collect 21 out of 33 potential training datasets to facilitate future exploration. After analyzing over 3,000 bad cases of 41 LLMs based on ToLeaP, we identify four main critical challenges: (1) benchmark limitations induce both the neglect and lack of (2) autonomous learning, (3) generalization, and (4) long-horizon task-solving capabilities of LLMs. To aid future advancements, we take a step further toward exploring potential directions, namely (1) real-world benchmark construction, (2) compatibility-aware autonomous learning, (3) rationale learning by thinking, and (4) identifying and recalling key clues. The preliminary experiments demonstrate their effectiveness, highlighting the need for further research and exploration.

DPNet: Dynamic Pooling Network for Tiny Object Detection

In unmanned aerial systems, especially in complex environments, accurately detecting tiny objects is crucial. Resizing images is a common strategy to improve detection accuracy, particularly for small objects. However, simply enlarging images significantly increases computational costs and the number of negative samples, severely degrading detection performance and limiting its applicability. This paper proposes a Dynamic Pooling Network (DPNet) for tiny object detection to mitigate these issues. DPNet employs a flexible down-sampling strategy by introducing a factor (df) to relax the fixed downsampling process of the feature map to an adjustable one. Furthermore, we design a lightweight predictor to predict df for each input image, which is used to decrease the resolution of feature maps in the backbone. Thus, we achieve input-aware downsampling. We also design an Adaptive Normalization Module (ANM) to make a unified detector compatible with different dfs. A guidance loss supervises the predictor's training. DPNet dynamically allocates computing resources to trade off between detection accuracy and efficiency. Experiments on the TinyCOCO and TinyPerson datasets show that DPNet can save over 35% and 25% GFLOPs, respectively, while maintaining comparable detection performance. The code will be made publicly available.

Dance of Fireworks: An Interactive Broadcast Gymnastics Training System Based on Pose Estimation

This study introduces Dance of Fireworks, an interactive system designed to combat sedentary health risks by enhancing engagement in radio calisthenics. Leveraging mobile device cameras and lightweight pose estimation (PoseNet/TensorFlow Lite), the system extracts body keypoints, computes joint angles, and compares them with standardized motions to deliver real-time corrective feedback. To incentivize participation, it dynamically maps users' movements (such as joint angles and velocity) to customizable fireworks animations, rewarding improved accuracy with richer visual effects. Experiments involving 136 participants demonstrated a significant reduction in average joint angle errors from 21.3 degrees to 9.8 degrees (p < 0.01) over four sessions, with 93.4 percent of users affirming its exercise-promoting efficacy and 85.4 percent praising its entertainment value. The system operates without predefined motion templates or specialised hardware, enabling seamless integration into office environments. Future enhancements will focus on improving pose recognition accuracy, reducing latency, and adding features such as multiplayer interaction and music synchronisation. This work presents a cost-effective, engaging solution to promote physical activity in sedentary populations.

LexPro-1.0 Technical Report

In this report, we introduce our first-generation reasoning model, LexPro-1.0, a large language model designed for the highly specialized Chinese legal domain, offering comprehensive capabilities to meet diverse realistic needs. Existing legal LLMs face two primary challenges. Firstly, their design and evaluation are predominantly driven by computer science perspectives, leading to insufficient incorporation of legal expertise and logic, which is crucial for high-precision legal applications, such as handling complex prosecutorial tasks. Secondly, these models often underperform due to a lack of comprehensive training data from the legal domain, limiting their ability to effectively address real-world legal scenarios. To address this, we first compile millions of legal documents covering over 20 types of crimes from 31 provinces in China for model training. From the extensive dataset, we further select high-quality for supervised fine-tuning, ensuring enhanced relevance and precision. The model further undergoes large-scale reinforcement learning without additional supervision, emphasizing the enhancement of its reasoning capabilities and explainability. To validate its effectiveness in complex legal applications, we also conduct human evaluations with legal experts. We develop fine-tuned models based on DeepSeek-R1-Distilled versions, available in three dense configurations: 14B, 32B, and 70B.

Global Semantic-Guided Sub-image Feature Weight Allocation in High-Resolution Large Vision-Language Models

As the demand for high-resolution image processing in Large Vision-Language Models (LVLMs) grows, sub-image partitioning has become a popular approach for mitigating visual information loss associated with fixed-resolution processing. However, existing partitioning methods uniformly process sub-images, resulting in suboptimal image understanding. In this work, we reveal that the sub-images with higher semantic relevance to the entire image encapsulate richer visual information for preserving the model's visual understanding ability. Therefore, we propose the Global Semantic-guided Weight Allocator (GSWA) module, which dynamically allocates weights to sub-images based on their relative information density, emulating human visual attention mechanisms. This approach enables the model to focus on more informative regions, overcoming the limitations of uniform treatment. We integrate GSWA into the InternVL2-2B framework to create SleighVL, a lightweight yet high-performing model. Extensive experiments demonstrate that SleighVL outperforms models with comparable parameters and remains competitive with larger models. Our work provides a promising direction for more efficient and contextually aware high-resolution image processing in LVLMs, advancing multimodal system development.

Collaborative Evolving Strategy for Automatic Data-Centric Development

Artificial Intelligence (AI) significantly influences many fields, largely thanks to the vast amounts of high-quality data for machine learning models. The emphasis is now on a data-centric AI strategy, prioritizing data development over model design progress. Automating this process is crucial. In this paper, we serve as the first work to introduce the automatic data-centric development (AD^2) task and outline its core challenges, which require domain-experts-like task scheduling and implementation capability, largely unexplored by previous work. By leveraging the strong complex problem-solving capabilities of large language models (LLMs), we propose an LLM-based autonomous agent, equipped with a strategy named Collaborative Knowledge-STudying-Enhanced Evolution by Retrieval (Co-STEER), to simultaneously address all the challenges. Specifically, our proposed Co-STEER agent enriches its domain knowledge through our proposed evolving strategy and develops both its scheduling and implementation skills by accumulating and retrieving domain-specific practical experience. With an improved schedule, the capability for implementation accelerates. Simultaneously, as implementation feedback becomes more thorough, the scheduling accuracy increases. These two capabilities evolve together through practical feedback, enabling a collaborative evolution process. Extensive experimental results demonstrate that our Co-STEER agent breaks new ground in AD^2 research, possesses strong evolvable schedule and implementation ability, and demonstrates the significant effectiveness of its components. Our Co-STEER paves the way for AD^2 advancements.

RD2Bench: Toward Data-Centric Automatic R&D

The progress of humanity is driven by those successful discoveries accompanied by countless failed experiments. Researchers often seek the potential research directions by reading and then verifying them through experiments. The process imposes a significant burden on researchers. In the past decade, the data-driven black-box deep learning method demonstrates its effectiveness in a wide range of real-world scenarios, which exacerbates the experimental burden of researchers and thus renders the potential successful discoveries veiled. Therefore, automating such a research and development (R&D) process is an urgent need. In this paper, we serve as the first effort to formalize the goal by proposing a Real-world Data-centric automatic R&D Benchmark, namely RD2Bench. RD2Bench benchmarks all the operations in data-centric automatic R&D (D-CARD) as a whole to navigate future work toward our goal directly. We focuses on evaluating the interaction and synergistic effects of various model capabilities and aiding to select the well-performed trustworthy models. Although RD2Bench is very challenging to the state-of-the-art (SOTA) large language model (LLM) named GPT-4, indicating ample research opportunities and more research efforts, LLMs possess promising potential to bring more significant development to D-CARD: They are able to implement some simple methods without adopting any additional techniques. We appeal to future work to take developing techniques for tackling automatic R&D into consideration, thus bringing the opportunities of the potential revolutionary upgrade to human productivity.

Variational Stochastic Gradient Descent for Deep Neural Networks

Optimizing deep neural networks is one of the main tasks in successful deep learning. Current state-of-the-art optimizers are adaptive gradient-based optimization methods such as Adam. Recently, there has been an increasing interest in formulating gradient-based optimizers in a probabilistic framework for better estimation of gradients and modeling uncertainties. Here, we propose to combine both approaches, resulting in the Variational Stochastic Gradient Descent (VSGD) optimizer. We model gradient updates as a probabilistic model and utilize stochastic variational inference (SVI) to derive an efficient and effective update rule. Further, we show how our VSGD method relates to other adaptive gradient-based optimizers like Adam. Lastly, we carry out experiments on two image classification datasets and four deep neural network architectures, where we show that VSGD outperforms Adam and SGD.

A Critical Look at Classic Test-Time Adaptation Methods in Semantic Segmentation

Test-time adaptation (TTA) aims to adapt a model, initially trained on training data, to potential distribution shifts in the test data. Most existing TTA studies, however, focus on classification tasks, leaving a notable gap in the exploration of TTA for semantic segmentation. This pronounced emphasis on classification might lead numerous newcomers and engineers to mistakenly assume that classic TTA methods designed for classification can be directly applied to segmentation. Nonetheless, this assumption remains unverified, posing an open question. To address this, we conduct a systematic, empirical study to disclose the unique challenges of segmentation TTA, and to determine whether classic TTA strategies can effectively address this task. Our comprehensive results have led to three key observations. First, the classic batch norm updating strategy, commonly used in classification TTA, only brings slight performance improvement, and in some cases it might even adversely affect the results. Even with the application of advanced distribution estimation techniques like batch renormalization, the problem remains unresolved. Second, the teacher-student scheme does enhance training stability for segmentation TTA in the presence of noisy pseudo-labels. However, it cannot directly result in performance improvement compared to the original model without TTA. Third, segmentation TTA suffers a severe long-tailed imbalance problem, which is substantially more complex than that in TTA for classification. This long-tailed challenge significantly affects segmentation TTA performance, even when the accuracy of pseudo-labels is high. In light of these observations, we conclude that TTA for segmentation presents significant challenges, and simply using classic TTA methods cannot address this problem well.