Yutong Zhang

Nataliya Nechyporenko, Yutong Zhang, Sean Campbell et al.

What if a robot could rethink its own morphological representation to better meet the demands of diverse tasks? Most robotic systems today treat their physical form as a fixed constraint rather than an adaptive resource, forcing the same rigid geometric representation to serve applications with vastly different computational and precision requirements. We introduce MorphIt, a novel spherical approximation framework that treats morphological representation as a tunable resource. MorphIt enables task-driven morphological adaptation through gradient-based optimization with tunable parameters that provide explicit control over the accuracy-efficiency tradeoff. Unlike existing approaches that rely on either labor-intensive manual specification or inflexible computational methods optimized for visualization rather than robotics, MorphIt generates spherical approximations up to 100x faster while maintaining superior geometric fidelity. Quantitative evaluations demonstrate that MorphIt outperforms baseline approaches (Variational Sphere Set Approximation and Adaptive Medial-Axis Approximation), achieving better mesh approximation with fewer spheres. Through seamless integration with existing robotics infrastructure, MorphIt enables enhanced capabilities in collision detection accuracy, contact-rich interaction simulation, and navigation through confined spaces. By dynamically adapting geometric representations to task requirements, robots can now exploit their physical embodiment as an active resource rather than an inflexible parameter, opening new frontiers for manipulation in environments where physical form must continuously balance precision with computational tractability.

Zhenyu Pan, Yutong Zhang, Jianshu Zhang et al.

Multimodal Large Language Models (MLLMs) already achieve state-of-the-art results across a wide range of tasks and modalities. To push their reasoning ability further, recent studies explore advanced prompting schemes and post-training fine-tuning. Although these techniques improve logical accuracy, they frequently leave the models' outputs burdened with pronounced social biases. Clarifying how reasoning gains interact with bias mitigation-and whether the two objectives inherently trade off-therefore remains an open and pressing research problem. Our study begins by benchmarking three bias-mitigation strategies-supervised fine-uning (SFT), knowledge distillation (KD), and rule-based reinforcement learning (RL)-under identical conditions, establishing their baseline strengths and weaknesses. Building on these results, we vary the proportion of debias-focused and reasoning-centric samples within each paradigm to chart the reasoning-versus-bias trade-off. Our sweeps reveal a consistent sweet spot: a roughly 1:4 mix trained with reinforcement learning cuts stereotype scores by 10% while retaining 88% of the model's original reasoning accuracy, offering concrete guidance for balancing fairness and capability in MLLMs.