WikiSkill let a 9B model beat a 27B rival, but one transferred skill cut Gemini from 50.5% to 18.1%
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
Gemini model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.
Google Research's WikiSkill method allows agents to compile execution history into reusable skill files, enhancing performance. A Qwen-3.5-9B model with evolved skills achieved an average of 47.4% across five benchmarks, outperforming a Qwen-3.6-27B model without skills, which scored 39.4%. When the 27B model also received skills, its performance rose to 63.3%. This suggests that procedural memory and model capacity are complementary, though one transferred skill reportedly reduced Gemini's performance from 50.5% to 18.1%.
Google Research’s new WikiSkill preprint tests a useful idea for long-running agents: instead of relying only on a model’s weights or a growing transcript, turn execution history into a maintained knowledge base and then compile that knowledge into reusable skill files.
The system has three layers. Raw execution traces remain immutable. A “wiki” consolidates successful strategies, recurring failures and previous changes. A separate proposer turns those lessons into concise skills, and a validation gate keeps an update only when it improves the immediate score.
Across five agent benchmarks, Qwen-3.5-9B with evolved skills averaged 47.4%, compared with 39.4% for Qwen-3.6-27B without skills. That does not mean skills replace scale: the 27B model reached 63.3% when it received its own skills, up 23.9 percentage points. The more interesting result is that procedural memory and model capacity appear complementary.
Transfer between models was mixed. On ALFWorld, the 9B model scored 63.4% with a skill it evolved itself and 70.2% with one evolved by the 27B model. But a spreadsheet skill written by the weakest Qwen model reduced Gemini-3.5-Flash from 50.5% to 18.1%. A brittle workaround learned by a weaker system can become a harmful instruction when a stronger model follows it literally.
Limitations were bounded benchmarks; skills were placed directly in the prompt rather than retrieved from a large library; the wiki did not prune itself; and the work is a preprint, not a production system. Still, it suggests that evaluating an agent only by its base model misses a growing part of the stack: what it can retain, validate and reuse from earlier runs.
I wrote a fuller breakdown for Learning the World, including the cross-model results and failure cases: https://www.lrngwrld.com/smaller-ai-model-beats-a-larger-one-if-it-inherits-the-right-skills-google-paper-finds/
Primary paper: https://arxiv.org/abs/2608.27454
Disclosure: I edit Learning the World and wrote the linked article.