Fireworks reproduces J-Lens readout on open models

Fireworks AI reproduced and verified the silent signal readout of the Jacobian Lens (J-Lens) proposed by Anthropic in open models. This work applies the J-Lens probe to Kimi K3 and Qwen3.5-9B, reading out vocabulary related to the final answer from the hidden states before the model generates the first token, i.e., the "silent signal." In Kimi's arithmetic and citrus contrast groups, under the same sentence output, J-Lens read out arithmetic words (7, 9, answer, seven) and citrus words (orange, lemon, lime) respectively, with no cross-talk between the two groups; among the 20 vocabulary lists from 10 examples, each list had at least one word enter the top 10 of the readout results. The related results were released on August 12, 2026, and there are currently no further developments.

AI-generated from public reports

1. Aug 12· 2 reports

   Fireworks reproduces the open model J-Lens readout

   Source:Fireworks AI