Safety4h ago
Safety
Meta-Tokens Reveal Hidden Model Algorithms
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 LCM math results. Suppressing the "most likely" token removed hedging, forcing the AI to commit to single answers. This technique allows practitioners to pinpoint and steer non-obvious algorithmic behaviors.