I post-trained Qwen3-Coder to fix bugs using an actual debugger. The result:
Solve rate: 70% → 89%
Median turns to fix: 46 → 19 (-59%)
Instead of just reading code or print-debugging, it:
- reasons from execution
- inspects live variables and call stacks
- sets breakpoints, steps, and evaluates expressions
@lu__jasper I would expect one step per batch to be better because:
1. less off-policy (async lag already adds staleness)
2. each update averages over more samples, so the gradient is less noisy