让下一篇重要论文，不再错过。

解析强到弱在线策略蒸馏与GRPO的计算效率对比

Unpacking the Compute Efficiency of Strong-to-Weak OPD against GRPO

Zeyu Leo Liu, Yang Li, Bo Pang, Emre Can Acikgoz, Xuan-Phi Nguyen, Anurag Koul, Srijan Bansal, Erik Nijkamp 等 10 位作者

AI 速览：系统比较反向KL在线策略蒸馏（OPD）与GRPO在达到目标准确率时所需训练FLOPs的差异，并揭示教师来源对OPD效率的关键影响。登录阅读完整解读 →

On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher's per-token supervision. Does this dense signal reduce the training FLOPs needed to reach a target accuracy compared with reinforcement learning? We systematically compare vanilla reverse-KL OPD with Group Relative Policy Optimization (GRPO) across task difficulty, student and teacher size, domain, and optimization choices, measuring held-out accuracy against training FLOPs. With an external post-trained teacher, OPD generally loses to GRPO in accuracy per training FLOP at matched settings, even when the teacher is further trained with GRPO, and often suffers response-length inflation. Yet a different teacher changes the picture: when the teacher is the student's own GRPO checkpoint, OPD recovers that checkpoint's accuracy on math with $50\times$ fewer distillation FLOPs than GRPO needs to reach the same accuracy. Following the same teacher setting, we consistently observe this high efficiency with a larger student, on code, and in a second model family. On a single domain, however, the GRPO checkpoint is already the desired model, so distilling it adds to total compute. Across domains, separate GRPO checkpoints hold different capabilities, and no single checkpoint serves every domain. Multi-teacher OPD (DeepSeek-AI, 2026) consolidates per-domain GRPO experts into a single student that matches each expert's accuracy, with the distillation stage costing only 5.2% (on 2 domains) and 8.0% (on 4 domains) of the specialist training FLOPs. Counting specialist training and distillation together, this pipeline reaches joint GRPO's performance on the 2-domain mix with 42.0% fewer total FLOPs.

PS-DLO: Property- and Spatially Aware Deformable Linear Object Manipulation Policy

Mengdi Li, Boxian Yao, Xiyu Zhu, Yunfan Lou, Yifan Yang, Yuanpeng He, Xuan Han, Tao Liu 等 9 位作者

Deformable linear objects (DLO) pose a coupled perception and control challenge: their slender geometry makes precise localization difficult, while their deformations depend on object-specific physical properties. To address this challenge, we propose PS-DLO, a property- and spatially aware DLO manipulation policy. First, contrastive image-text learning maps the initial observation to categorical descriptions of DLO properties, providing explicit physical and structural priors for action generation. Second, SAM3 extracts mask-guided features from future observations, which are distilled into the policy's action expert to encode task-relevant DLO regions and their subsequent configurations. We evaluate PS-DLO on nine real-world tasks across single-arm and bimanual platforms, training a shared multi-task policy for each embodiment. PS-DLO achieves an average success rate of 68.9%, exceeding the strongest baseline by 20.0 percentage points. In generalization experiments, PS-DLO achieves higher success rates than the baseline in six of seven visual-shift conditions on butterfly-knot untying. These results demonstrate that combining object-property priors with future spatial supervision improves multi-task DLO manipulation and robustness to visual variation.

尚无 AI 解读，展示原文摘要