https://xiaomailab.com/en/updates/feed.xmlResearch updates | Xiaomai LabOne note a week: what got done, what didn't, and what comes next. Written plainly for supporters, with a technical record at the end for researchers to check.2026-11-05T00:00:00+08:00 tag:xiaomailab.com,2026:en/updates/2026-W44Update #4｜From research tools to verifiable model work2026-11-05T00:00:00+08:00郭俊逸A native 9B checkpoint cold-loaded, while an external 9B used research tools on real source problems. Model-proposed conditions with bounded search and an engineering dependency repair are reported separately. A later native 9B submission does not establish general capability gains. tag:xiaomailab.com,2026:en/updates/2026-W43Update #3｜Connecting update feedback and a local research assistant, while keeping failed gates visible2026-10-26T00:00:00+08:00郭俊逸CPU feedback, recovery and task grading advanced; a three-arm comparison found no FULL advantage. Real Ornith powered a read-only research assistant. Retrieval repair improved source coverage at higher cost. After a RAM stop, native 9B restored and resaved older custom state, without reaching research or updates. Autonomy, complete migration and new-task gains remain unverified. tag:xiaomailab.com,2026:en/updates/2026-W42Update #2｜What I connected in Xiaomai this week, and which failures still need work2026-10-26T00:00:00+08:00郭俊逸This week joined block checks, small-model True wiring, CPU update/save flows, rollback, same-host runtime relocation, and bounded CUDA work. A100 then completed two selected-parameter 9B-class updates and first-checkpoint restore. I also cover publication, custody, and retirement. Prior FP32 0/14 and controls 104/112 still fail. Checkpoint two has no third-process reload; autonomy, general input, and capability gains remain unverified. tag:xiaomailab.com,2026:en/updates/2026-W41Update #1｜Continuation atlas: two fresh processes resume the same next step; a method proposal fails at format2026-10-26T00:00:00+08:00郭俊逸After research parameters change and are saved, a restart reads every checked number back unchanged, and two restarts from the same save produce matching checked state and gradients after the next update. That is groundwork for testing whether a change can last, not proof the model solves problems better. The model's own method proposal was blocked by a format error. Real-model experiments have been paused since Oct 9 and the cloud GPU released; weekly updates continue.