Jiang's Research
I am JiangHongwei. I run experiments on frontier learning systems: looped transformers, long-horizon tasks, recursive self-improvement, test-time training, and software-engineering agents.
These notes discuss how such systems learn in the loop—through post-training, evaluation, and changing state distributions. Failures, instrumentation, and implementation details stay close to every claim.
Field notes
Research, with the instrumentation left in.
Why Claude chose 16K
Two experiments challenge the bottleneck hypothesis. More tokens connect performance with billing.
Claude’s 16K Bet
Is the tokenizer becoming part of the reasoning architecture?
Beyond teacher quality
Student-relative residual learning for an already strong SWE agent.
Research note13 Sep 2026
Why Claude chose 16K
An elegant hypothesis fails: two experiments, more tokens, and the connection between performance and billing.
