SkillOpt-Lite Lifts GPT-5.4-nano Past GPT-5.5 on Spreadsheets

Something worth pausing on from a fresh arXiv preprint: a minimalist optimization loop reportedly gets more capability out of a small model than the same loop gets out of the larger one it is measured against. On LiveMath, the authors report a 25.4 point gain for GPT-5.4-nano and an 8.8 point gain for GPT-5.5. On SpreadsheetBench, the tuned nano lands at 0.7758 accuracy, above GPT-5.5's 0.7620.

The pitch, from Yifei Shen, Bo Li, and Xinjie Zhang, is that skill optimization for coding agents can be treated as a zeroth-order search problem, and that three components do the heavy lifting: file-system-based trajectory exploration, consensus attribute mining, and independent validation gating. They frame it as a minimal viable pipeline, with each component justified by theory or empirical necessity, and describe it as designed to plug into production coding agents like VSCode Copilot.

The reason this matters if you are not writing agent papers: the default reflex for the last couple of years has been that if you want a better agent, you reach for a better model. A result that says a small model plus a tight optimization loop can outrun a bigger model on a couple of concrete tasks flips the calculus. If it holds up outside these benchmarks, the cost curve for capable coding agents bends downward, since you run the cheap model and spend your compute on skill search rather than on parameters.

The honest caveat is that this is a preprint, benchmarks are not products, and the reported wins are on two specific tasks. What the paper does not give you is a picture of how the approach behaves on longer-horizon agent runs, on messier real developer workloads, or on model families beyond the two GPT-5.4 and GPT-5.5 variants tested. Minimal on the reported benchmarks is not the same as robust in daily engineering use.

For teams shipping agentic products on tight budgets, though, the direction is the part worth watching. If skill-level optimization at the code layer generalizes, the interesting lever moves from which model tier you buy toward how well you can search over the agent's own scaffolding.