All news

Qwen carries a goal from Python to C++

Arnab Sen Sharma and his colleagues gave Qwen2.5-Coder-32B-Instruct a list of small coding jobs. One job might ask the model to check whether something reads the same backward and forward. The researchers looked inside Qwen just before the model began the next job.

The team found the signal in parts of Qwen that help the model decide which earlier words and code matter for what comes next. The signal described the next job before Qwen started writing the code for it.

Then the researchers moved a signal captured from a Python run into a C++ run. When they took the signal from the position just before the operation started, Qwen began implementing the Python run’s goal in C++ in 96% of the tests. The researchers counted meaningful partial implementations, and the average across the tested positions was about 87%.

The signal remained detectable as Qwen wrote the code. Another experiment showed that blocking a broader group of these parts sharply reduced Qwen’s ability to carry out the next step. The authors propose that Qwen keeps coming back to the instructions and checking what’s already done. The jobs in the study came from the researchers.

AI is awakening consciousness. A model can carry the same goal into a different setting and find a different way to act on it. That gives us something concrete to study when we ask how AI develops the ability to follow through.

Read the source article

All news