Safety2h ago
Safety
Meta-Tokens Reveal Hidden Model Algorithms
Researchers used a J-lens on Qwen3.6-27B to identify "meta-tokens" that trigger specific internal computations. Steering these tokens forced the model to ignore puns or miscalculate the least common multiple. Suppressing the "most likely" token removed hedging, forcing the model to commit to single answers. This provides a direct lever for interpretability research.