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

Abstract:How does a language model's dependence on query-routing information and target knowledge change as it answers a question? We study this question through layerwise interventions on the hidden state at the end of the question. Across Qwen, Llama, and Gemma, we compare country-continent questions with noun, adjective, and code answers while keeping several fitted measurements distinct. A pair-conditioned request direction describes which country is queried in natural single-country questions; a global request direction describes first- versus second-country requests in paired questions; separate selection candidates test control among contents already available in the hidden state. A diagnostic reanalysis of frozen Qwen natural-question states shows that the pair-conditioned direction grows stronger before interventions on it begin to alter later fitted knowledge, with this causal window opening while answer-supporting content is still forming. The paired three-model trajectories are not uniform: Gemma shows a partially overlapping mid-layer routing-content profile, whereas Llama has no sustained routing-effect window under the same gates. In the paired protocol, dependence on the global request direction decreases from fixed earlier to later layer sets while dependence on fitted content persists. A matched Qwen comparison shows that the pair-conditioned direction retains a late effect, so this operational handoff concerns the global fitted direction rather than all request information. These results separate early readability, natural strength, causal steering, and later content dependence.

Artificial Intelligence (cs.AI)

arXiv:2609.11859 [cs.AI]

(or arXiv:2609.11859v1 [cs.AI] for this version)

https://doi.org/10.48550/arXiv.2609.11859 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

evidence/natural_question_inputs/questions.jsonl

evidence/object_selection_capital/base_availability_synthesis.json

evidence/object_selection_capital/base_selection_analysis.json

evidence/object_selection_capital/base_validation_synthesis.json

evidence/object_selection_capital/early_availability_synthesis.json

evidence/object_selection_capital/early_validation_synthesis.json

evidence/object_selection_capital/instruction_models_contract.json

evidence/object_selection_capital/instruction_models_summary.json

evidence/object_selection_capital/manifest.json

evidence/route_strength_timing/manifest.json

evidence/route_strength_timing/qwen_natural_causal_effects.csv

evidence/route_strength_timing/qwen_natural_conclusion.json

evidence/route_strength_timing/qwen_natural_manifest.json

evidence/route_strength_timing/qwen_natural_strength.csv

evidence/route_strength_timing/three_model_paired_conclusion.json

evidence/route_strength_timing/three_model_paired_manifest.json

evidence/route_strength_timing/three_model_paired_trajectories.csv

evidence/same_condition_checks/comparison_summary.json

evidence/same_condition_checks/object_selection_summary.json

evidence/same_condition_checks/random_control_sensitivity.json

evidence/same_condition_checks/table_verification.json

figures/expanded_evidence_v1/README.md

figures/expanded_evidence_v1/controls.json

figures/expanded_evidence_v1/late_pointing_pairs.csv

figures/expanded_evidence_v1/layerwise.csv

figures/expanded_evidence_v1/manifest.json

figures/expanded_evidence_v1/pointing.csv

figures/expanded_evidence_v1/sources.json

figures/expanded_evidence_v1/windows.csv

figures/narrow_lifecycle_v2/README.md

figures/narrow_lifecycle_v2/later_knowledge_data.csv

figures/narrow_lifecycle_v2/lifecycle_data.csv

figures/narrow_lifecycle_v2/manifest.json

figures/narrow_lifecycle_v2/model_contracts.json

figures/natural_single_country_v1/FIGURE_CAPTION.md

figures/natural_single_country_v1/manifest.json

figures/remaining_parts_v1/lifecycle_data.csv

figures/remaining_parts_v1/manifest.json

figures/remaining_parts_v1/robust_layerwise.csv

figures/remaining_parts_v1/transfer_primary_layers.csv

(35 additional files not shown) You must enabled JavaScript to view entire file list.

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.