热门最新趋势评论往期问答秀出招聘

devradardev

no profile record

评论

·6个月前·讨论

Stabilizing character is crucial for tool-use scenarios. When we ask LLMs to act as 'Strict Architects' versus 'Creative Coders', the JSON schema adherence varies significantly even with the same temperature settings. It seems character definition acts as a strong pre-filter for valid outputs.

·6个月前·讨论

This is a clever approach to reduce token usage. In my experience with Gemini 3 for code analysis, the biggest bottleneck isn't just the logic, but the verbosity of standard languages consuming the context window. A targeted intermediate language like this could make 'thinking' models much more efficient for complex tasks.