How Qwen3.5 works
A detailed account of Qwen3.5, from pixels and visual tokens to Gated DeltaNet, Full Attention, MRoPE, and MoE, with equations and tensor shapes for the 9B model.
Blog
Research notes on vision and multimodal learning, with occasional thoughts on academic life.
6 posts
Code-as-World and the World as an Executable Hypothesis
How Code-as-World connects visual evidence, executable world representations, and abductive discovery, and what its results do and do not establish about physical understanding in an open world.
Understanding a world while the video is still running
A close reading of streaming video evaluation, incremental inference, persistent state, and thinking during observation, with a focus on evidence, compute budgets, and questions the system has not seen yet.
From Video Compression to Codec-Native Multimodal Models
An English reader's map to the full Chinese audit of video codecs, OneVision-Encoder, codec-video-prep, LLaVA-OneVision-2, and Mage-VL.
So, Why Do I Want to Do a PhD?
It has been almost a month since my admission letter arrived. Now that it is official, I feel I owe myself an honest answer about why I want to do a PhD.
Capability Beyond the Model, Rereading Harness Engineering for Self-Improvement
Why does the same foundation model behave so differently across agent systems? This long-form essay follows execution loops, context, durable state, tools, evaluation, and rollback through ACE, MCE, Meta-Harness, ADAS, AFlow, STOP, and Self-Harness.