Show HN: Qwen Meetup Presentation, Function Calling Harness, 6.75% to 100%(autobe.dev)
autobe.dev
Show HN: Qwen Meetup Presentation, Function Calling Harness, 6.75% to 100%
https://autobe.dev/blog/function-calling-harness-qwen-meetup-korea/
1 comments
Next subject: function calling for domains where the final answer is unverifiable.
Legal rulings, business decisions: you can't verify the conclusion. But you can verify the process. Model every mandatory reasoning step as typed schema properties, put the verdict last.
CoT becomes 100% structurally guaranteed. Streaming + property order validation catches the AI skipping steps in real time, mid-generation. DeepPartial<T> fills corrections progressively as output streams in.
If you can verify the process, you can harness any domain.
Legal rulings, business decisions: you can't verify the conclusion. But you can verify the process. Model every mandatory reasoning step as typed schema properties, put the verdict last.
CoT becomes 100% structurally guaranteed. Streaming + property order validation catches the AI skipping steps in real time, mid-generation. DeepPartial<T> fills corrections progressively as output streams in.
If you can verify the process, you can harness any domain.
The talk was about how I got function calling to work reliably on deeply recursive union types — the stuff the industry generally says doesn't work. With qwen3-coder-next, first-try success rate was 6.75%. And the entire Qwen 3.5 model family was hitting 0% on union types due to a consistent double-stringify bug. Both ended up at 100%.
Slides are also downloadable in the article — speaker notes are written inside as slide notes if you'd like the full narrative behind each slide.