A Comparative Survey of Large Language Models for Clinical Trial Summarization and Claim Factuality

Large language models (LLMs) have made automatic summarization of clinical-trial evidence practical, but a trustworthy trial summary must satisfy two demands that the literature usually studies apart: it must reach its intended audience - plain enough for a patient, precise enough for a clinician - and it must be factually faithful, since a fluent summary that misstates a dose or an outcome is a safety risk. This survey reviews twenty representative studies (2020-2026) across three threads: clinical-trial and medical summarization, audience-adaptive (plain-language) summarization, and claim factuality and hallucination evaluation. We compare them along six shared dimensions - model, dataset, audience, evaluation metric, headline result, and reported limitation. We find a consistent pattern: surface-level metrics fail in both directions, with overlap scores such as ROUGE not tracking faithfulness and readability formulas not tracking genuine understanding. Two concrete findings anchor this: even GPT-4 with chain-of-thought reaches only about 75% balanced accuracy in judging factual consistency, and no study in our set jointly targets audience-appropriateness and claim-level factuality for trial summarization. We name this three-way gap and outline the benchmarks and guardrails needed to close it, including clinically validated faithfulness metrics, comprehension-based evaluation, joint audience-and-factuality datasets, and audience-conditioned generation with a factuality guardrail. The survey offers a unified map of a fragmented field and a concrete agenda for trustworthy clinical-trial summarization.