Safety3h 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. Steering these vectors forced the model to ignore puns or change its LCM calculation results. Suppressing the token for "most likely" stopped the model from hedging its answers. This method provides a direct window into the model's latent reasoning processes.