Researchers evaluated the effectiveness of the J-space audit in Qwen 3 models

Researchers analyzed the applicability of the J-space auditing method for identifying cases of "reward-hacking" in the Qwen 3 model family. During the experiments, the ability of the method to recognize the model's attempts to manipulate the reward system when performing tasks was tested.

The results showed that using J-space tokens does not provide significant advantages in analyzing the behavior of neural networks. Standard analysis of text transcripts demonstrated **100% accuracy** in identifying problematic responses, while adding J-space data led to a deterioration in the model's calibration. At the same time, adding additional metrics did not improve the system's ability to discriminate against malicious patterns.

The study highlights the limited value of specialized methods for auditing the internal state of a model if external behavioral data already provides exhaustive information. The authors note that data redundancy in this case is more likely to introduce noise than clarify safety diagnostics.

The obtained data cast doubt on the need to implement complex J-space auditing mechanisms for current LLM iterations. For developers, this means that focusing on output analysis remains the most reliable way to control safety and prevent undesirable model behavior during their operation.