News
New blog post: Closing the loop, on how a simple rule beats an agent-written weight edit.
MemShift, on separating detection from repair in LLM-derived personal memory, is on arXiv.
ToolBottleneckBench, a benchmark for tool agents under controlled pressures, is on arXiv.
Joined Initial. S in Tokyo as an AI Agent Development Engineer.
Started my Master’s at The University of Osaka, advised by Prof. Chuan Xiao.