Qwen model's J-lens reveals meta-tokens that surface hidden algorithms

Researchers from the AI Alignment Forum, including agam_bhatia, camilablank, and Neel Nanda, have applied the J-lens interpretability technique to Qwen3.6-27B to uncover 'meta-tokens'—tokens that surface the model's internal algorithms. J-lens, originally introduced by Anthropic, accesses a cognitive space in the residual stream. By taking the Jacobian with respect to the penultimate layer (averaged over 25 sequences of 128 tokens from the Pile), the team found that certain Chinese tokens appear at specific layers and correspond to algorithmic processing. For instance, '什么意思' ('what does it mean') appears when the model encounters ambiguous text; steering it away causes the model to treat a pun literally. Similarly, 'gcd' fires during LCM problems, and swapping the GCD vector from 9 to 3 changes the model's answer for the LCM of 27 and 90 from 270 to 810.

Beyond these case studies, the team identified a hedging meta-token '大概率' ('most likely'), which appears before the model offers multiple possibilities. Suppressing it forces the model to commit to a single answer—e.g., changing 'There are several logical places John could have gone' to 'he went to the stationery store.' The work is a proof-of-concept that J-lens can directly indicate algorithms, not just intermediate variables. The authors note that current J-lens is limited to single-token concepts, but multi-token extensions could unlock even deeper insights into model reasoning. This approach offers a new path for AI safety and interpretability by making internal algorithms transparent and steerable.