Fireworks AI applied the Jacobian Lens (J‑Lens) – a trained probe that reads a model’s hidden state – to two open models, Kimi K3 and Qwen3.5‑9B. The probe recovered vocabulary related to the task in the top‑10 predictions of the internal state before the models actually wrote those words. In Kimi, all 20 tested word lists showed at least one match; in Qwen, 19 of the 20 lists matched. The study also showed the pattern holds while the model processes the prompt, not just during generation.
Why it matters
It lets researchers peek at a model’s reasoning before it speaks, opening new ways to study and improve AI behavior.
