Large Language Models (LLMs) demonstrate impressive capabilities, yet their outputs often suffer from misalignment with human preferences due to the inadequacy of weak supervision and a lack of fine-grained control. Training-time alignment methods like Reinforcement Learning from Human Feedback (RLHF) face prohibitive costs in expert supervision and inherent scalability limitations, offering limited dynamic control during inference. Consequently, there is an urgent need for scalable and adaptable alignment mechanisms. To address this, we propose W2S-AlignTree, a pioneering plug-and-play inference-time alignment framework that synergistically combines Monte Carlo Tree Search (MCTS) with the Weak-to-Strong Generalization paradigm for the first time. W2S-AlignTree formulates LLM alignment as an optimal heuristic search problem within a generative search tree. By leveraging weak model's real-time, step-level signals as alignment proxies and introducing an Entropy-Aware exploration mechanism, W2S-AlignTree enables fine-grained guidance during strong model's generation without modifying its parameters. The approach dynamically balances exploration and exploitation in high-dimensional generation search trees. Experiments across controlled sentiment generation, summarization, and instruction-following show that W2S-AlignTree consistently outperforms strong baselines. Notably, W2S-AlignTree raises the performance of Llama3-8B from 1.89 to 2.19, a relative improvement of 15.9% on the summarization task.

Training Large Language Models (LLMs) is prohibitively expensive, creating a critical scaling gap where insights from small-scale experiments often fail to transfer to resource-intensive production systems, thereby hindering efficient innovation. To bridge this, we introduce Farseer, a novel and refined scaling law offering enhanced predictive accuracy across scales. By systematically constructing a model loss surface $L(N,D)$, Farseer achieves a significantly better fit to empirical data than prior laws (e.g., \Chinchilla's law). Our methodology yields accurate, robust, and highly generalizable predictions, demonstrating excellent extrapolation capabilities, outperforming Chinchilla's law, whose extrapolation error is 433\% higher. This allows for the reliable evaluation of competing training strategies across all $(N,D)$ settings, enabling conclusions from small-scale ablation studies to be confidently extrapolated to predict large-scale performance. Furthermore, Farseer provides new insights into optimal compute allocation, better reflecting the nuanced demands of modern LLM training. To validate our approach, we trained an extensive suite of approximately 1,000 LLMs across diverse scales and configurations, consuming roughly 3 million NVIDIA H100 GPU hours. To foster further research, we are comprehensively open-sourcing all code, data, results (https://github.com/Farseer-Scaling-Law/Farseer), all training logs (https://wandb.ai/billzid/Farseer?nw=nwuserbillzid), all models used in scaling law fitting (https://huggingface.co/Farseer-Scaling-Law).

Abstract The detection of road potholes plays a crucial role in ensuring passenger comfort and the structural safety of vehicles. To address the challenges of pothole detection in complex road environments, this paper proposes a model focusing on shape features (pothole detection you only look once, PD‐YOLO). The model aims to overcome the limitations of multi‐scale feature learning caused by the use of fixed convolutional kernels in the baseline model, by constructing a feature extraction module that better adapts to variations in the shape of potholes. Subsequently, a cross‐stage partial network was designed using a one‐time aggregation method, simplifying the model while enabling the network to fuse information between feature maps at different stages. Additionally, a dynamic sparse attention mechanism is introduced to select relevant features, reducing redundancy and suppressing background noise. Experiments conducted on the VOC2007 and GRDDC2020_Pothole datasets reveal that compared to the baseline model YOLOv8, PD‐YOLO achieves improvements of 3.9% and 2.8% in mean average precision, with a frame rate of approximately 290 frames per second, effectively meeting the accuracy and real‐time requirements for pothole detection. The code and dataset for this paper are located at: https://github.com/woyijiankou/PD‐YOLO .

Abstract To tackle challenges in road multi‐object detection, such as object occlusion, small object detection, and multi‐scale object detection difficulties, a new YOLOv8n‐RSFM structure is proposed. The key improvement of this structure lies in the introduction of the transformer decoder head, which optimizes the matching between the ground truth and predicted boxes, thereby effectively addressing issues of object overlap and multi‐scale detection. Additionally, a small object detection layer is incorporated to retain crucial information beneficial for detecting small objects, significantly improving the detection accuracy for small targets. To enhance learning capacity and reduce redundant computations, the FasterNet backbone is employed to replace CSPDarknet53, thus accelerating the training process. Finally, the INNER‐MPDIoU loss function is introduced to replace the original algorithm's complete IoU to accelerate the convergence and obtain more accurate regression results. A series of experiments were conducted on different datasets. The experimental results show that the proposed model YOLOv8N‐RSFM outperforms the original model YOLOv8n in small target detection. On the VisDrone, TinyPerson, and VSCrowd datasets, the mean accuracy percentage improved by 7.9%, 12.3%, and 4.5%, respectively.