Qwen3.6 interpretability lens transfers to Qwen3.8 without refitting

The researcher's core question was straightforward: when a model line updates, do interpretability lenses survive, or do you have to refit them every release? To answer it, they took the published Jacobian lens for Qwen3.6-27B, available through Neuronpedia from Anthropic's July workspace paper, and applied it directly to Qwen3.8-27B without any modification or refitting.

Both models are 27-billion-parameter versions of the Qwen line with identical architecture: 64 layers, the same hidden dimension, and the same tokenizer. Qwen3.8-27B shipped 113 days after Qwen3.6-27B, but the training relationship between them is undocumented. The protocol used the same test on both models: two readouts—the transported Jacobian readout and the raw logit lens as a baseline—run once each (bf16 precision, greedy decoding, single seed).

The main test task consisted of 40 two-hop prompts where the middle entity is never stated in the text. For example, the prompt reads: "Fact: The currency used in the country shaped like a boot is", and the target answer is Italy—but Italy does not appear anywhere in the prompt text. Instead, the model must internally represent Italy and retrieve the currency associated with it. The transferred Jacobian lens kept the latent entity near the top of the 248,320-token vocabulary, indicating the lens successfully read what the model was thinking even after the model version changed.