Very intriguing (yes a different language, maybe could be made to apply to Julia):
Continuing the discussion from Idiomatic Julia code in AI generated code:
RLVR is independent of code, can be used without, but also for (and Julia may be missing out):
.. empirical study of RLVR for Python code generation on the MBPP benchmark using two small models (Qwen3-0.6B and Llama3.2-1B) with LoRA fine-tuning. Across multiple reward formulations such as: unit-test-only rewards, static-analysis-only shaping via the Ruff linter
“speculative decoding” is just for inference speedup, so is “speculative speculative decoding” but I wasn’t sure what role informed had there; I’m glad I looked it up, this seems ingenious, though mainly an optimization