ML//code generation

Code generation is the work of LLMs specialized or fine-tuned for writing, editing and reasoning about code. From Codex (2021) to DeepSeek-Coder and Qwen2.5-Coder, it is among the most commercially impactful applications of LLMs.

Code generation is the work of LLMs specialized or fine-tuned for writing, editing and reasoning about code. From Codex (2021) to DeepSeek-Coder and Qwen2.5-Coder, it is among the most commercially impactful applications of LLMs.

The field moved from research to products: Copilot, Cursor, Devin, Claude Code. Code is the domain where models improved fastest, and the reason is checkability: a program can be compiled, run and tested, so a system can generate candidates, keep those a verifier accepts and feed the errors back, and training can reward what passes. That loop of generate, test and repair is the core of coding agents, with the caveat that tests sample behaviour and prove nothing beyond the cases they cover.

Its benchmarks are HumanEval (saturated), SWE-Bench (agent-level), LiveCodeBench and Codeforces.