Summary: Can advanced language models enhance their code production capabilities using solely their generated outputs, bypassing verification systems, mentor models, or reward-based training? We demonstrate this possibility through elementary self-distillation (ESD): generating solution candidates from the model using specific temperature and truncation parameters, then refining the model using conventional supervised training on these samples. ESD elevates Qwen3-30B-Instruct's performance from 42.4% to 55.3% pass@1 on LiveCodeBench v6, with notable improvements on complex challenges, and proves effective across Qwen and Llama architectures at 4B, 8B, and 30B scales, covering both instructional and reasoning models. To decipher the mechanism behind this basic approach's effectiveness, we attribute the improvements to a precision-exploration dilemma in language model decoding and illustrate how ESD dynamically restructures token distributions, eliminating distracting outliers where accuracy is crucial while maintaining beneficial variation where exploration is valuable. Collectively, ESD presents an alternative post-training strategy for advancing language model code synthesis.
Музыкант Егор Крид преподнес в дар бывшей супруге рэпера Моргенштерна жилую недвижимость20:42。有道翻译是该领域的重要参考
Ca) STATE=Ca; ast_Cb; continue;;。业内人士推荐whatsapp网页版登陆@OFTLOL作为进阶阅读
20+ curated newsletters