Give an LLM agent a text and a tool: get_nla_explanation(position) β a
natural-language autoencoder
that verbalizes Qwen3.6-27B's layer-42 activation at any token of the text, scored by round-trip
reconstruction (FVE). The agent probes positions, reports what the model was representing, and can
present findings as a widget highlighting tokens with excerpts from their explanations.
Everything β system prompt, context block, tool calls, raw results β is shown below.
The example texts are fully precached and served from this site, so probing them never
touches Hugging Face; custom texts and uncached positions call the two Spaces
(matryoshka,
standard).
Your API keys stay in this browser; nothing is sent anywhere except OpenRouter and, when
needed, the Spaces.