This paper asks a simple question: when a large language model answers, when is it routing the question to the right internal pathway and when is it pulling up stored knowledge? The authors study how query-routing information and target knowledge change over the model’s layers as the model forms an answer. They do this by intervening in the model’s internal state at the end of the question and watching what changes.

The team worked with three models called Qwen, Llama, and Gemma. They used a controlled task: country-to-continent questions that could produce different answer types (a noun, an adjective, or a piece of code). To probe internal processing they fit simple directions in the hidden state that represent request information. For example, a “pair-conditioned request direction” marks which country is being asked about in single-country questions. A “global request direction” marks whether the first or second country is requested in paired questions. They also tested separate candidate contents that might already be present in the hidden state. Interventions mean they changed those fitted directions or contents at specific layers and then observed effects on later behavior.

Their measurements show a sequence of events inside the models. In a frozen Qwen reanalysis, the pair-conditioned direction grew stronger before changing later fitted knowledge — in other words, a causal window opened while answer-supporting content was still forming. The three models did not behave the same. Gemma showed a partly overlapping mid-layer profile where routing and content both mattered. Llama, under the same gating choices, did not show a sustained routing-effect window. In the paired-question setup, dependence on the global request direction tended to fall from earlier to later layers, while dependence on the fitted content stayed strong. A matched Qwen comparison showed the pair-conditioned direction still had a later effect, suggesting that the operational handoff involves the global fitted direction specifically, not every piece of request information. The paper frames these results as separating early readability, natural strength, causal steering, and later content dependence.