From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

Summary

The paper asks how large language models progressively turn a query into an answer drawn from stored factual knowledge, using layerwise interventions on the hidden state at the question-end position. On Qwen-2.5-3B-Instruct, Llama-3.2-3B-Instruct, and Gemma-3-4B-Instruct, with country–continent questions answered by noun, adjective, or arbitrary code, the authors fit linear components — a two-dimensional continent-content space, pair-conditioned and global 'request' directions, and a marker-calibrated selection direction — then delete or reverse each and compare against eight length-matched Gaussian controls. In Qwen natural questions, the request direction strengthens from layer 21 before interventions on it alter later fitted knowledge (layers 28–32), while content is still forming. Under a paired protocol, deleting the globally fitted request direction damages answers less in later than earlier layer sets in all three models, while content deletion stays consequential. A matched comparison (Table 3) shows this 'handoff' is specific to the global direction: pair-conditioned deletion retains a large late effect (3.09 versus -0.20); Gemma shows only a partial analogue, and Llama fails the paper's own fit-robustness gate. Clean object-selection evidence is limited to a Qwen base-model capital task. Main limitations: one relation family, six-country fitting sets, diagnostic rather than registered timing onsets, and no explicit code-release statement in the provided text.