We study how language models aggregate conflicting evidence in controlled multi-juror settings. As a null control, symbolic majority voting, where jurors provide only YES/NO votes, yields near-zero minority override across models, confirming this is not an artifact of the prompt family. Under semantic aggregation, where jurors instead provide short evidence statements paired with local verdicts, override becomes strength-dependent and position-sensitive: rates rise when a minority statement is operationalized as stronger than a weak majority, while a strong-majority/weak-minority control remains low-override. Authority cues applied to a minority statement can modulate this effect in some contexts, but do not universally dominate evidence strength. We validate these patterns with seed replications, cross-domain templates, phase-transition mapping across majority strength, minority strength, and position, and rationality-vs-salience controls. We further confirm the pattern on a larger open-weight model (Llama-3.3-70B-Instruct) and the GPT-4.1 flagship model: Llama-3.3-70B stays close to the smaller-model baseline, while GPT-4.1 shows a substantially larger override magnitude, indicating the effect’s size is model-dependent even as its qualitative shape persists. We emphasize that evidence “strength” here is a synthetic, operationalized manipulation rather than real-world ground truth, and that minority-following can be rational when minority evidence is genuinely decisive.

Large Language Models (LLMs) are increasingly used for scientific writing and research assistance, yet their ability to maintain consistent citations across multi-turn conversations remains unexplored. This paper introduces the concept of citation drift—the phenomenon where references mutate, disappear, or get fabricated during extended LLM interactions. We analyze 240 conversations across four LLaMA models using 36 authentic scientific papers from six domains and find significant citation instability. LLaMA-4-Maverick-17B achieves the highest stability (0.481) and lowest fabrication entropy, while LLaMA-4-Scout-17B fabricates up to 85.6% of citations. We introduce five new metrics—stability, fabrication rate, drift rate, drift entropy, and willingness-to-cite—providing a standardized framework for evaluating factual reliability in scientific dialogue systems. Our benchmark offers reproducible, model-agnostic evaluation tools for assessing citation reliability in AI-assisted research workflows.