Context
van Dort and Heuss use activation patching in Llama-3.1-8B-Instruct to identify distributed components involved in deciding whether to emit inline citations, then test interventions on PopQA-derived prompts and HotpotQA.
What changed
A primary experimental study adds internal-mechanism and intervention evidence to citation evaluation: the decision to output a citation marker can be influenced by distributed model components, including early document-question entity matching. Targeted interventions change citation behavior, although multi-document strict citation accuracy remains low.
Why it matters
Citation presence or correctness is not equivalent to demonstrating that retrieved evidence causally informed the answer. This distinction matters for measuring AI visibility, source selection and grounded answer quality.
Visibility OS interpretation
Visibility OS interpretation: treat retrieval, answer grounding, citation correctness and citation faithfulness as separate measurement layers. This model-specific experiment is not evidence of citation behavior in ChatGPT Search, Google AI Mode, Perplexity or other production systems.
What remains uncertain
Will the same citation mechanism replicate across larger models, architectures and live search-augmented systems? Can mechanistic faithfulness be measured without internal model access? How robust are the results to different prompts and document types?