📝 SummarXiv

Abstract

Large language models are increasingly aligned to human preferences through reinforcement learning from human feedback (RLHF) and related methods such as Direct Preference Optimization (DPO), Constitutional AI, and RLAIF. While effective, these methods exhibit recurring failure patterns i.e., reward hacking, sycophancy, annotator drift, and misgeneralization. We introduce the concept of the Alignment Gap, a unifying lens for understanding recurring failures in feedback-based alignment. Using a KL-tilting formalism, we illustrate why optimization pressure tends to amplify divergence between proxy rewards and true human intent. We organize these failures into a catalogue of Murphys Laws of AI Alignment, and propose the Alignment Trilemma as a way to frame trade-offs among optimization strength, value capture, and generalization. Small-scale empirical studies serve as illustrative support. Finally, we propose the MAPS framework (Misspecification, Annotation, Pressure, Shift) as practical design levers. Our contribution is not a definitive impossibility theorem but a perspective that reframes alignment debates around structural limits and trade-offs, offering clearer guidance for future design.