From Parameters to Answers: How LLMs Retrieve and U... | AI Research

From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge
This research investigates how Large Language Models (LLMs) transition from processing a user's question to retrieving the specific internal knowledge required to generate an answer. By analyzing the hidden states of models like Qwen, Llama, and Gemma, the authors aim to uncover the "mechanical" process of how a model identifies what information it needs and how it eventually consolidates that information into a final response. To see meta in practice, How to Make Cinematic Commercials walks through a concrete example.

Distinguishing Routing from Content

A central finding of the paper is that a model’s internal hidden states perform two distinct roles: "content" and "routing." Content refers to the actual information used to form an answer, while routing acts as a guide, directing the model’s internal computation toward where that information is stored. The researchers identified two types of routing: parameter routing, which points to knowledge embedded in the model's weights, and hidden-state routing, which points to information already retrieved and stored within the model's current processing layers.

The Process of Knowledge Formation

The study reveals that these components emerge in a specific order as a model processes a question. In the early layers, the model begins to build "answer-supporting content." At this stage, routing has little impact. As the computation progresses, parameter routing strengthens, actively guiding the model to pull relevant facts from its pre-trained parameters. Once this knowledge is retrieved, it is integrated into the hidden state. Finally, hidden-state routing takes over, directing the model to focus on this newly formed content to finalize the answer. The same ai evaluation question is explored in Wrong Prediction, Right Answer, which adds a research perspective.

The Routing-Content Handoff

The researchers identified a critical "handoff" point in the model's computation. By intervening on the model's internal states—specifically by redirecting parameter routing toward incorrect information—they found that the model is easily influenced in its earlier layers. However, as the computation reaches later layers, the model becomes increasingly resistant to these routing changes, relying instead on the content it has already consolidated. This suggests that the model’s causal dependence shifts from the "search" mechanism (routing) to the "result" (content) as it nears the end of its reasoning process.

Key Takeaways

The study highlights that LLMs do not simply "know" an answer from the start; they engage in a structured, multi-layer process of retrieval and consolidation. By separating these functional roles, the authors demonstrate that the model’s internal state is not a static representation of knowledge, but a dynamic workspace where information is actively located, retrieved, and refined. These results provide a clearer picture of the causal mechanisms that allow LLMs to transform a simple query into a specific, fact-based answer. The ai search story also surfaces in Qwen Developers Open-Source Local-First Search Layer..., adding another angle. as detailed in the full paper on Arxiv