dailyai.report

Researchers used J-lens on Qwen3.6-27B to identify meta-tokens that trigger specific internal computations. Manipulating these vectors forced the model to ignore puns or change mathematical answers for LCM problems. Suppressing the "most likely" token stopped the model from hedging. This discovery allows practitioners to steer internal logic by targeting specific algorithmic activations.