Can open models carry readable silent signals before they speak? Reproducing J-Lens Readouts on Kimi K3 & Qwen3.5-9B

In 30 Sekunden

Was passiert ist

Researchers applied J-Lens probes to open models Kimi K3 and Qwen3.5-9B, detecting task-relevant vocabulary in hidden states before models generate visible output tokens.

Warum es zählt

Matters for engineers building interpretability tools, debugging model behavior, or understanding what models compute internally during inference.

Achtung

Method relies on fitted probes trained on calibration data; results depend on tokenizer choices and may not generalize across all model architectures or tasks.