SDA-3 tl;dr
A short explanation of SDA-3 as a method for mapping LLM response structure without claiming access to hidden reasoning.
Tag
A short explanation of SDA-3 as a method for mapping LLM response structure without claiming access to hidden reasoning.
This post introduces SDA-3, a protocol for inferring the structure of an LLM’s embedding space through observable outputs, without relying on access to internal weights or hidden states.