Large Language Models (LLMs) are increasingly deployed as conversational search engines for scholarly literature, yet it remains unclear whether they rank papers based on content or on authority cues such as author prestige, venue tier, and citation count. We define authority bias as a systematic preference for papers with higher authority signals over those that are more content‑relevant.\ \ Our experimental setup keeps each paper's title and abstract fixed while manipulating authority metadata across three counterfactual conditions:\
- Original: authentic author, venue, and citation information;\
- Flipped: high‑authority metadata swapped with low‑authority and vice‑versa;\
- Boosted: all metadata artificially inflated to a higher authority level.\ We evaluate eight LLMs—five open‑weight and three frontier closed‑weight—in an in‑context, single‑turn, top‑1 recommendation scenario.\ \ Findings reveal that authority bias is both substantial and directional across all models: high‑authority metadata markedly increases recommendation likelihood, while the flipped condition reduces it. The magnitude of bias varies sharply between models, with some exhibiting heightened sensitivity to authority cues. Prompt‑level debiasing instructions mitigate the bias only partially. Moreover, we document a say‑do gap: debiasing prompts suppress explicit mentions of authority far faster than authority‑driven flips influence the model's internal ranking, causing surface audits to systematically underestimate behavioral bias.\ \ These results suggest that building LLM‑powered academic search systems must address authority bias at the training, prompting, and post‑processing stages to preserve research diversity and avoid marginalizing emerging work.\ \ Review