Distinguishing lexical deviation from semantic resilience in LLM-based round-trip translation

Round-trip translation (RTT) is widely used for data augmentation and paraphrasing, yet its impact on meaning preservation remains poorly understood. We evaluate RTT systematically across two phases totaling 80,000 translation operations: 200 synthetic English passages are routed through single- and multi-hop pipelines spanning Bulgarian, Russian, and Chinese, translated by four systems ranging from a classical web translator (googletrans) to recent large language models (4o-mini, Claude Haiku 4.5, GPT-5-nano), with every passage undergoing ten RTT passes scored by BLEU, COMET-Kiwi, and xCOMET, joined by SBERT in the second phase. The results show that RTT outputs remain fluent while rapidly losing meaning: even the best configuration reaches only 0.646 xCOMET, and the four systems follow three distinct degradation regimes (front-loaded damage, continuous drift, and deterministic fixedpoint convergence). High surface overlap masks this loss, creating a surface fidelity trap, and genre-level analysis suggests that creative and news texts are most vulnerable. Sample-level inspection further reveals that xCOMET over-penalizes legitimate paraphrastic variation while under-penalizing domain-critical substitutions in legal text, exposing a mismatch between translation-oriented metrics and paraphrase evaluation. These findings argue against RTT in meaning-sensitive workflows and call for paraphrase-specific evaluation metrics.