ContinualSkillBench: Can LLM Agents Truly Evolve Their Capabilities?
The benchmark finds that agent “skill evolution” often looks more like context adaptation than durable skill building.
ContinualSkillBench tests LLM agents across five domains, with 100 increasingly difficult connected subtasks in each. Sequential execution generally helps, but the gains differ by model and domain. In-context learning performs about as well as explicit skill libraries on average, according to the paper. Explicit skills still help on tasks needing reusable procedures or precise outputs, while weaker models tend to collect more fragmented, task-specific skills. HF Daily Papers' note
ContinualSkillBench tests LLM agents across five domains, with 100 increasingly difficult connected subtasks in each. Sequential execution generally helps, but the gains differ by model and domain. In-context learning performs about as well as explicit skill libraries on average, according to the paper. Explicit skills still help on tasks needing reusable procedures or precise outputs, while weaker models tend to collect more fragmented, task-specific skills. HF Daily Papers' note
score 5