How LLMs Find Answers in Trillions of Parameters: Layerwise Probing of Internal Knowledge Retrieval

Forum topic · 小凯 · 2026-09-11

Summary

This post analyzes an arXiv paper (2609.11859, 'From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge' by Wenkang Wei et al.) that investigates how large language models retrieve factual knowledge from their parameters. Using layerwise interventions—surgically modifying hidden states at specific layers—the authors probe whether models first 'understand' a query (query-routing information) and then 'retrieve' the answer (target knowledge), or mix both processes. Experiments on Qwen, Llama, and Gemma reveal distinct behaviors: Qwen shows a clear three-stage pipeline (early layers parse the question, a mid-layer 'causal window' opens where routing interventions alter answers, late layers solidify the output); Gemma exhibits partial overlap between query understanding and answer content; Llama shows no sustained routing window, suggesting a more end-to-end pattern-matching mechanism. A key finding, 'operational handoff,' shows models gradually abandon reliance on question-type information while maintaining dependence on answer content. The post discusses implications for the understanding-vs-memorization debate, AI interpretability, hallucination detection, knowledge editing, and future research directions.

The Treasure Hunter Deep in the Parameters: How LLMs Find Answers Among Trillions of Numbers

This post reviews the paper "From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge" by Wenkang Wei, Yuan Fang, Renhe Jiang, Hong Cheng, and Xingtong Yu (arXiv:2609.11859, published 2026-09-10).

> *"We don't see things as they are, we see them as we are."* — Anaïs Nin

The Core Question

A decade-long debate in AI asks whether LLMs genuinely "understand" language or merely memorize and match statistical patterns. Rather than taking sides, the paper asks a more precise question: when an LLM answers a question, how does its dependence on query-routing information (what is being asked) versus target knowledge (the actual answer) evolve across its layers?

For a question like "What is the capital of France?", the model must: 1. Route the query — parse the question and extract key elements ("France", "capital"). 2. Retrieve target knowledge — locate "Paris" within its parameters.

The paper examines whether these happen in the same layers or in a division of labor, and whether this differs across models.

Methodology: "Brain Surgery" on LLMs

The authors use layerwise interventions: they modify hidden states at specific layers and observe the effect on the final output—akin to blacking out a mystery film at different points to find when the audience learns the killer's identity.

Three probes are designed:

Pair-Conditioned Request Direction: a classifier direction distinguishing which of a pair of entities (France vs. Germany) the model is processing.
Global Request Direction: a direction distinguishing the question type (asking about capitals vs. other concepts).
Selection Candidates: injecting candidate answers (Paris, Berlin, London) into hidden states to see which best explains model behavior, and at which layer the answer is "formed".

Key Findings

Qwen: a clear relay race

Early layers (~1–15): request directions strengthen as the model parses the question, but interventions do not yet affect answers—the librarian has read the slip but not yet walked to the shelves.
Middle layers (~15–25): a causal window opens; intervening on request directions now changes the final answer (France → Germany flips Paris → Berlin).
Late layers (~25–32): the answer is consolidated; further interventions have diminishing effect.

Model-specific personalities

Qwen: a clean "understand → route → answer" pipeline.
Gemma: partial overlap between routing and content dependence—question understanding and answer retrieval run partly in parallel.
Llama: strikingly, no sustained routing-effect window under the same conditions—possibly relying on more end-to-end pattern matching rather than explicit routing-then-retrieval.

Operational handoff

global request direction

answer content

What This Means

Knowledge retrieval in LLMs is staged and separable, not a single black-box operation: early layers parse, middle layers route and activate, late layers select and generate.
The findings blur the understanding-vs-memorization dichotomy. A simple lookup table would not show staged processing, yet LLM "routing" is not human-like understanding either. The authors characterize it as "structured retrieval".
For AI interpretability, staged retrieval enables fine-grained control: correcting errors by inspecting mid-layer routing, injecting knowledge at middle layers, and detecting hallucinations via the selection-candidates probe (an answer with no strong content-candidate support in late layers is likely fabricated).

Open Questions

Side effects of interventions on other layers (mitigated with "frozen" techniques, but not eliminated).
Whether other task types (reasoning, code, creative writing) use different pipelines.
How scaling to hundreds of billions or trillions of parameters changes the routing-retrieval mechanism.

Future directions include knowledge editing (correcting stored facts without retraining), capability decoupling (mixing layers from different models), and probing for forms of metacognition.

Reference

Wei, W., Fang, Y., Jiang, R., Cheng, H., & Yu, X. (2026). From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge. *arXiv preprint* arXiv:2609.11859. https://arxiv.org/abs/2609.11859