Researchers published a slimmed-down framework for teaching AI agents to improve their own skills — and claim it outperforms the heavier system it replaces.
The paper, from EvolvingLMMs-Lab, introduces SkillOpt-Lite, a stripped-back approach to "skill optimization" — the process by which autonomous agents revise the procedures they use to complete tasks. The authors ground the method in zeroth-order optimization theory and three stated principles: file-system-based trajectory exploration, consensus attribute mining, and independent validation gating. On the LiveMath benchmark, SkillOpt-Lite improved scores by 8.8 points on what the paper designates GPT-5.5 and by 25.4 points on GPT-5.4-nano — model designations used by the paper's authors that have not been independently confirmed against any publicly documented release lineup. The team also tested a broader "HarnessOpt" variant on SpreadsheetBench, where the nano-class model reached 0.7758 accuracy against 0.7620 for the larger model running a standard pipeline.
The result worth watching is not the benchmark numbers themselves but the direction: a cheaper, smaller model beating a larger one because the scaffolding around it was optimized rather than the model weights. That flips the usual "just use a bigger model" instinct and has real cost implications for anyone running agents at scale.
The team says SkillOpt-Lite plugs into production tools like VSCode Copilot and requires, as advertised, a single line of configuration. Whether that claim holds outside controlled benchmark conditions is the question every "one-line solution" paper eventually has to answer.