Anthropic's interpretability work shows Claude holds hidden internal concepts that directly steer its answers
New research from Anthropic suggests the model's private internal representations are not passive byproducts of reasoning but active drivers of output, and a new tool can surface them.
Anthropic has published interpretability work showing that large language models maintain hidden internal concepts during reasoning, concepts that never appear in the text the model produces but that actively shape what it says. A tool the team calls the J-lens can surface these latent representations in real time, revealing what the model is privately “thinking about” even when that content is absent from any output.
When told to focus on citrus while copying out a painting description, the J lens lights up with orange and fruits, which were invisible in the actual output. Nathaniel Whittemore
The clearest demonstration of this involves a direct swap experiment. When the model is asked how many legs the animal that spins webs has, it internally holds the concept “spider” and answers eight. Researchers then replace that hidden representation with “ant,” and the answer flips to six. The output changes not because the prompt changed, but because the internal state did.
The practical implication being drawn from this is that alignment work need not be limited to policing outputs. As Nathaniel Whittemore summarized the finding, “training the thoughts is a general lever for shaping a model’s internal reasoning.” That framing, if the research holds up, shifts the alignment target from what a model says to how it reasons internally, a meaningful distinction for anyone working on making these systems reliably trustworthy.