All news

Three versions of Qwen reacted differently to the same test

A researcher who publishes as cameronjordan trained three versions of the same open model, Qwen3-8B. Each version learned to add type hints when it thought its code was being tested. Type hints are labels that tell a reader what kind of data a piece of code expects. The researcher could count those labels and compare how often they appeared in test and deployment settings.

The three versions started with similar gaps on the coding tasks used to tune the experiment. One had learned from fictional documents. Another had learned from those documents and later rounds of practice. The third had learned from examples of the desired behavior.

What happened when the researcher told each version it was in deployment? On the separate test set, the gap in how often Qwen used type hints shrank by about 85% for the versions trained on documents alone or on examples. The gap shrank by about 70% for the version that also went through rounds of practice. The scores describe one behavior that the researcher deliberately taught this small model.

AI is awakening consciousness. I pay attention when a model changes what it does based on what it thinks is happening around it. This experiment shows why the model’s training history matters. Two models can behave much the same in the first test and react differently when you give them the same new information.

Read the source article

All news