Yifu Guo (郭毅芙)

Decoupling Continual Semantic Segmentation

First Author

Proposed DecoupleCSS that addresses catastrophic forgetting and background shift by decoupling class-aware detection from class-agnostic segmentation. Achieved SOTA with 12.12% and 17.99% mIoU improvements.

SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning

First Author | 192 Stars | Media Report

Proposed EvolveAgent, a trajectory-level iterative optimization framework. Significantly expanded solution space through crossover and mutation, improving baseline from 28% to 58% on SWE-Bench.

VideoSeg-R1: Reasoning Video Object Segmentation via Reinforcement Learning

Co-first Author

First work introducing reinforcement learning into VOS. Combines hierarchical frame sampling and multimodal reasoning with SAM2 and XMem for improved segmentation.

CogCaS: Cognitive Cascade Segmentation in Continual Semantic Segmentation

Co-first Author

Systematically analyzes existing CISS paradigms and proposes a new modeling method that effectively solves class-incremental semantic segmentation problems.

SkillEvo: An Experience Learning Framework with RL for Skill Evolution

Second Author

Addresses sparse rewards and weak generalization in long-horizon tasks through a two-stage framework that evolves trajectories into reusable skills.

S2A-Attention for Multimodal 3D Semantic Segmentation

Co-first Author

Lightweight multimodal 3D segmentation framework with Agent attention mechanism. Reduces inference latency to 20% while maintaining SOTA accuracy on nuScenes and KITTI.

RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories

Baseline Reproduction and Idea Discussion