> ## Content Index
> Fetch the complete content index at: https://www.lrngwrld.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google researchers’ WikiSkill helps a 9B model beat a 27B rival
- URL: https://www.lrngwrld.com/smaller-ai-model-beats-a-larger-one-if-it-inherits-the-right-skills-google-paper-finds/
- Published: 2026-08-30T09:04:09.000Z
- Updated: 2026-08-30T09:22:39.000Z
- Description: The smaller Qwen model scored 47.4% across five agent benchmarks after evolving reusable skills, compared with 39.4% for the larger model without them.
- Author: Michael de Waal-Montgomery
- Tags: News

A 9-billion-parameter AI model equipped with reusable skills outperformed a 27-billion-parameter model without them across five agent benchmarks in a new Google Research paper, suggesting that accumulated procedural knowledge can sometimes outweigh model scale.

In the [WikiSkill preprint](https://arxiv.org/abs/2608.27454?ref=lrngwrld.com), submitted to arXiv on 27 August, researchers Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan and Tu Vu describe a framework that allows AI agents to turn experience into a persistent knowledge base and then compile that knowledge into reusable skill files. The system changes the instructions available to an agent, not the model’s underlying weights.

The headline comparison came from two Qwen models. Qwen-3.5-9B averaged 47.4% after evolving skills with WikiSkill, while Qwen-3.6-27B averaged 39.4% without skills. The smaller model therefore finished eight percentage points ahead of a model with three times as many parameters under those test conditions.

WikiSkill separates an agent’s memory into three layers. An immutable raw layer stores execution traces. A wiki turns those traces into a cumulative record of successful strategies, recurring failures and previous skill changes. A skills layer contains the concise procedural instructions used during future tasks.

After each batch of tasks, a wiki-maintainer agent consolidates the results and a skill-proposer agent suggests an update. A validation gate keeps the change only when it improves performance. Rejected skill edits are rolled back, but the wiki retains the lesson so that later proposals do not have to rediscover the same failure.

The researchers tested five models from the Qwen, Gemma and Gemini families on benchmarks covering mathematical reasoning, web search, spreadsheet work, long-context document questions and interactive embodied tasks. WikiSkill improved most model-and-benchmark combinations and outperformed the three other skill-evolution systems used in the paper’s comparison.

Larger models generally gained more from the process. WikiSkill lifted the average score of Qwen-3.5-4B by 12.3 percentage points, Qwen-3.5-9B by 17.5 points and Qwen-3.6-27B by 23.9 points. The 27-billion-parameter model rose from 39.4% without skills to 63.3% with them, showing that skill accumulation and model scale were complementary rather than substitutes.

Some skills also transferred between models. On the ALFWorld benchmark, Qwen-3.5-9B scored 63.4% with a skill it evolved itself and 70.2% with one evolved by Qwen-3.6-27B. In another test, a skill written by the smallest Qwen model helped Gemma-4-31B beat its self-evolved result on LiveMath.

Transfer was not consistently beneficial. A spreadsheet skill evolved by Qwen-3.5-4B reduced Gemini-3.5-Flash’s score from 50.5% to 18.1%. The authors argue that a weaker model can encode brittle workarounds or unnecessary tool calls that a stronger model then follows too literally.

The paper also stops short of demonstrating a production-ready memory system. Skills were inserted directly into prompts rather than selected from a large library, the wiki did not prune itself, and the validation gate rejected any update that failed to improve the immediate score. The tests were bounded benchmarks rather than long-running real-world jobs, and the work is an arXiv preprint rather than a product announcement.

Within those limits, the results sharpen an important distinction. Model size determines what an agent may be capable of doing, but persistent skills determine whether it can reuse what earlier runs have already learned. On these tests, a smaller system with a good playbook beat a larger system starting from scratch.

✦ Takeaways by Learning The World AI Show Hide 
- Google Research’s WikiSkill helped a 9-billion-parameter model average 47.4% across five agent benchmarks, beating a 27-billion-parameter model without skills at 39.4%.
- The framework converts raw execution traces into a persistent wiki, then uses that accumulated knowledge to propose and validate reusable skill files.
- Skills transferred across models and families, but weak-model skills sometimes damaged stronger systems, underscoring that the results remain benchmark-dependent.

AI-generated from this article and reviewed by the editor.