Scaling Coding Agents via Atomic Skills

Key points

Current LLM coding agents are trained mainly on composite benchmarks such as bug fixing, which often leads to task-specific overfitting and limited generalization.
To address this, the authors propose a new scaling paradigm that shifts the focus from task-level optimization to mastery of atomic skills.
They first formalize five fundamental atomic skills that act as basis vectors for complex software engineering tasks: code localization, code editing, unit test generation, issue reproduction, and code review.
Compared with composite coding tasks, these atomic skills are more generalizable and composable.
The next step is to scale coding agents by performing joint RL on the atomic skills.
In this way, the atomic skills improve consistently without negative interference or trade-offs between them.

External research summaries. These are not HDATF publications or measured product results.