Do Deployment Constraints Make LLMs Hallucinate Citations? An Empirical Study across Four Models and Five Prompting Regimes

TL;DR

This study investigates how deployment constraints influence hallucinated citations in large language models, revealing significant fabrication rates and emphasizing the need for post-hoc verification in academic and software engineering contexts.

Contribution

It provides an empirical analysis of citation hallucination across multiple models and prompting regimes, highlighting the impact of deployment constraints on citation verifiability.

Findings

No model exceeds 47.5% citation existence rate

Temporal and combination constraints cause the steepest declines in verifiable citations

A substantial fraction of unresolved cases are fabricated citations

Abstract

LLMs are increasingly used to draft academic text and to support software engineering (SE) evidence synthesis, but they often hallucinate bibliographic references that look legitimate. We study how deployment-motivated prompting constraints affect citation verifiability in a closed-book setting. Using 144 claims (24 in SE&CS) and a deterministic verification pipeline (Crossref + Semantic Scholar), we evaluate two proprietary models (Claude Sonnet, GPT-4o) and two open-weight models (LLaMA~3.1-8B, Qwen~2.5-14B) across five regimes: Baseline, Temporal (publication-year window), Survey-style breadth, Non-Disclosure policy, and their combination. Across 17,443 generated citations, no model exceeds a citation-level existence rate of 0.475; Temporal and Combo conditions produce the steepest drops while outputs remain format-compliant (well-formed bibliographic fields). Unresolved outcomes dominate (36-61%); a 100-citation audit indicates that a substantial fraction of Unresolved cases are fabricated. Results motivate post-hoc citation verification before LLM outputs enter SE literature reviews or tooling pipelines.