Shidu (Ryan) Ren

PACE-Mem: Prefix-Aware Complementary Evidence for Multimodal Agents.

Shidu Ren, Rong-cheng Tu, Hang Zhou, Xiao Luo

PACE-Mem is a training-free retrieval method for multimodal agents with long-term memory. It selects complementary evidence and adapts retrieval depth from a shared selection trajectory, reducing redundancy while retaining relevant context. Across three long-video benchmarks and two agent frameworks, it improves question answering without changing memory construction or answer generation.

FlexiWorld: Learning and Planning via Flexible Action Chunks Across Multiple Time Scales.

Shidu Ren*, Qilin Gu*, Zhenghao Ni*, Junhan Sun, Jiaqi Wang, Damien Scieur, Yunze Liu
* Equal contribution

FlexiWorld is a JEPA-based world model that jointly learns latent dynamics and goal-conditioned action generation across multiple time scales. Flexible action chunks support different planning granularities without retraining, achieving 89.29% mean success across four control benchmarks and goal distances with actor-residual planning (ARCEM).

NeurIPS 2026 - LCFM

ICM-Bench: Person-Level Identity Reasoning in Multimodal Agents with Long-Term Memory.

Shidu Ren, Yunze Liu, Xing Liu, Chi-Hao Wu, Enmin Zhou, Junxiao Shen

ICM-Bench evaluates whether multimodal agents can remember recurring people and reason about their identities and relationships across long video memories. It contains 839 synthetic clips spanning 141 minutes and 1,217 open-ended questions across three identity-centric memory tasks.

IEEE VLSI 2026

A Split-Inference Intracortical Interface IC for Battery-Free mm-Scale Magnetoelectrically Powered Brain Implants.

Mustafa Kanchwala*, Jianxiong Xu*, Mohammad Abdolrazzaghi*, Wonjune Kim, Gerard O'Leary, Yu Huang, Junyu Ma, Jose Sales Filho, Sudip Nag, Hanfeng Cai, Qiaosong Deng, Weian Deng, Anush Mutyala, Theeban Kumaresan, Shidu Ren, Chae Lim, Mandana Movahed, Homeira Moradi, George Eleftheriades, Taufik A. Valiante, Jacob T. Robinson, Roman Genov

We present a split-inference intracortical interface IC for battery-free, millimeter-scale brain implants, with validation across motor reach-to-grasp and inner-speech decoding. The work connects low-power neural decoding, spike-sorting pipelines, and system-level evaluation, showing neural output data-rate reductions of 1,066× and 16,732× while preserving decoding performance.