# SkillNEO Turns AI Agents’ Partial Progress Into Better Skills

> A study reports that letting weaker skill versions receive more edits improved results on 86 benchmark tasks, with gains also carrying over to stronger models.

Oossa · 2026-10-16 · https://oossa.com/en/skillneo-turns-ai-agents-partial-progress-into-better-skills

Event date: 2026-10-14

A new arXiv paper introduces SkillNEO, a method for improving the reusable instructions—called skills—that AI agents follow to complete tasks. The authors report that it raised GPT-5.4 Mini’s average reward from 49.1% to 66.4% across 86 SkillsBench tasks in eight areas, including software engineering and office work.

The method addresses a problem with keeping only the best-scoring version of a skill: a version that scores worse may still contain a useful repair that needs more edits before it pays off. SkillNEO chooses separately which version to use and which versions to keep developing. By comparing related versions on the same task, it looks for steps that worked and failures that still need fixing. It can also return to an earlier version and revise it using evidence gathered later.

## Why keep a weaker version?

In the study, GPT-5.4 Mini carried out tasks, suggested revisions and edited the skills. The authors evaluated the final results on the same task instances used during skill evolution, rather than describing a separate test on new tasks. They report gains in all eight domains. The resulting score was within the range of the original-skill results reported for GPT-5.5 and GPT-5.6 Luna.

The authors also tried the modified skills with stronger models, without further editing those skills. On 41 tasks with modified skills, reuse improved Luna’s score by 10.5 percentage points. GPT-5.5 scored 10.6 points above its published reference using the original skills. The paper says some versions that scored lower during development eventually became the versions selected for use.

## What the results do—and don’t—show

The findings suggest that a less capable model’s attempts can produce useful instructions for a stronger agent, even when the smaller model itself still misses some task requirements. But the evaluation covers a specific set of 86 benchmark tasks, and the final evaluation reused the task instances involved in skill development. The paper therefore does not establish how well the method works on unfamiliar tasks. The supplied paper summary does not give pricing or access details for the models.

## The facts

- SkillNEO was tested on 86 SkillsBench tasks spanning eight domains.
- The authors report GPT-5.4 Mini’s average reward rising from 49.1% to 66.4%.
- On 41 tasks with modified skills, reuse improved Luna’s score by 10.5 percentage points.
- GPT-5.5 scored 10.6 percentage points above its published original-skill reference.

## Why it matters

For teams building task-focused AI agents, the results offer evidence that a cheaper or weaker model could help refine reusable instructions later used by a stronger model. The reported gains are limited to the tested benchmark tasks, so it remains unknown whether they carry over to new tasks.

## Sources & references

1. [SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills](https://arxiv.org/abs/2610.20926) – arXiv, 2026-10-16

Last updated: 2026-10-16