Wei Lin

Zihan Lin, Xiaohan Wang, Jie Cao et al.

Large language models (LLMs) transcend passive generation and act as goal-directed agents by invoking external tools. Reinforcement learning (RL) offers a principled framework for optimizing these emergent tool-use policies, yet the prevailing paradigm relies exclusively on sparse outcome rewards and lacks consideration of the particularity of tool-use tasks, inflating policy-gradient variance and resulting in inefficient training. To better understand and address these challenges, we first establish a theoretical link between policy entropy and training stability of tool-use tasks, which reveals that structured, low-entropy tokens are primary determinants of rewards. Motivated by this insight, we propose \textbf{Res}haped \textbf{T}oken-level policy gradients (\textbf{ResT}) for tool-use tasks. ResT reshapes the policy gradient through entropy-informed token reweighting, progressively upweighting reasoning tokens as training proceeds. This entropy-aware scheme enables a smooth shift from structural correctness to semantic reasoning and stabilizes convergence in multi-turn tool-use tasks. Evaluation on BFCL and API-Bank shows that ResT achieves state-of-the-art results, outperforming prior methods by up to $8.76\%$. When fine-tuned on a 4B base LLM, ResT further surpasses GPT-4o by $4.11\%$ on single-turn tasks and $1.50\%$ on multi-turn base tasks.

Yiwu Yao, Yuchao Li, Chengyu Wang et al.

The intensive computation of Automatic Speech Recognition (ASR) models obstructs them from being deployed on mobile devices. In this paper, we present a novel quantized Winograd optimization pipeline, which combines the quantization and fast convolution to achieve efficient inference acceleration on mobile devices for ASR models. To avoid the information loss due to the combination of quantization and Winograd convolution, a Range-Scaled Quantization (RSQ) training method is proposed to expand the quantized numerical range and to distill knowledge from high-precision values. Moreover, an improved Conv1D equipped DFSMN (ConvDFSMN) model is designed for mobile deployment. We conduct extensive experiments on both ConvDFSMN and Wav2letter models. Results demonstrate the models can be effectively optimized with the proposed pipeline. Especially, Wav2letter achieves 1.48* speedup with an approximate 0.07% WER decrease on ARMv7-based mobile devices.

Haiyue Zhu, Yiting Li, Fengjun Bai et al.

Data-efficient domain adaptation with only a few labelled data is desired for many robotic applications, e.g., in grasping detection, the inference skill learned from a grasping dataset is not universal enough to directly apply on various other daily/industrial applications. This paper presents an approach enabling the easy domain adaptation through a novel grasping detection network with confidence-driven semi-supervised learning, where these two components deeply interact with each other. The proposed grasping detection network specially provides a prediction uncertainty estimation mechanism by leveraging on Feature Pyramid Network (FPN), and the mean-teacher semi-supervised learning utilizes such uncertainty information to emphasizing the consistency loss only for those unlabelled data with high confidence, which we referred it as the confidence-driven mean teacher. This approach largely prevents the student model to learn the incorrect/harmful information from the consistency loss, which speeds up the learning progress and improves the model accuracy. Our results show that the proposed network can achieve high success rate on the Cornell grasping dataset, and for domain adaptation with very limited data, the confidence-driven mean teacher outperforms the original mean teacher and direct training by more than 10% in evaluation loss especially for avoiding the overfitting and model diverging.