TL-GRPO: Turn-Level RL for Reasoning-Guided Iterative Optimization

Large language models have demonstrated strong reasoning capabilities in complex tasks through tool integration, which is typically framed as a Markov Decision Process and optimized with trajectory-level RL algorithms such as GRPO. However, a common class of reasoning tasks, iterative optimization, presents distinct challenges: the agent interacts with the same underlying environment state across turns, and the value of a trajectory is determined by the best turn-level reward rather than cumulative returns. Existing GRPO-based methods cannot perform fine-grained, turn-level optimization in such settings, while black-box optimization methods discard prior knowledge and reasoning capabilities. To address this gap, we propose Turn-Level GRPO (TL-GRPO), a lightweight RL algorithm that performs turn-level group sampling for fine-grained optimization. We evaluate TL-GRPO on analog circuit sizing (ACS), a challenging scientific optimization task requiring multiple simulations and domain expertise. Results show that TL-GRPO outperforms standard GRPO and Bayesian optimization methods across various specifications. Furthermore, our 30B model trained with TL-GRPO achieves state-of-the-art performance on ACS tasks under same simulation budget, demonstrating both strong generalization and practical utility.

What You See is What You Grasp: User-Friendly Grasping Guided by Near-eye-tracking

This work presents a next-generation human-robot interface that can infer and realize the user's manipulation intention via sight only. Specifically, we develop a system that integrates near-eye-tracking and robotic manipulation to enable user-specified actions (e.g., grasp, pick-and-place, etc), where visual information is merged with human attention to create a mapping for desired robot actions. To enable sight guided manipulation, a head-mounted near-eye-tracking device is developed to track the eyeball movements in real-time, so that the user's visual attention can be identified. To improve the grasping performance, a transformer based grasp model is then developed. Stacked transformer blocks are used to extract hierarchical features where the volumes of channels are expanded at each stage while squeezing the resolution of feature maps. Experimental validation demonstrates that the eye-tracking system yields low gaze estimation error and the grasping system yields promising results on multiple grasping datasets. This work is a proof of concept for gaze interaction-based assistive robot, which holds great promise to help the elder or upper limb disabilities in their daily lives. A demo video is available at \url{https://www.youtube.com/watch?v=yuZ1hukYUrM}.