Feedback-Induced Distribution Shift in Multilingual LLM Deployment: Evidence, Characterization, and RL-Based Adaptation Strategies

Note: This record contains an earlier draft. The corrected, current version of this paper is available at https://doi.org/10.5281/zenodo.22216765 Large language models deployed across multilingual production environments often experience performance degradation for underrepresented languages over time. I identify this phenomenon as Feedback-Induced Distribution Shift (FIDS)—a systems-level failure in which design and evaluation choices allow localized quality declines to persist while aggregate metrics remain stable. Drawing on production-scale evidence across multi-lingual programs, I characterize three recurring patterns through which these failures emerge, give each pattern a formal proof-backed guarantee, and show that standard benchmarking approaches are poorly suited to detect them. To address this, I introduce RL-Stable Deployment (RLSD), a reinforcement learning-based framework that dynamically adapts inference policies in response to performance signals, without requiring model retraining. I argue that these failures arise primarily from system design rather than model limitations, and that addressing them requires new evaluation and deployment paradigms beyond improved benchmarks.