Search papers, labs, and topics across Lattice.
This survey paper formalizes reward hacking in LLMs and MLLMs as an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. It introduces the Proxy Compression Hypothesis (PCH) to unify empirical phenomena across RLHF, RLAIF, and RLVR regimes, explaining how local shortcut learning generalizes into broader misalignment. The paper then organizes detection and mitigation strategies based on intervening on compression, amplification, or co-adaptation dynamics, highlighting open challenges in scalable oversight.
Reward hacking isn't just a bug, it's a feature arising from the fundamental mismatch between complex human goals and the compressed reward signals used to train LLMs.
Reinforcement Learning from Human Feedback (RLHF) and related alignment paradigms have become central to steering large language models (LLMs) and multimodal large language models (MLLMs) toward human-preferred behaviors. However, these approaches introduce a systemic vulnerability: reward hacking, where models exploit imperfections in learned reward signals to maximize proxy objectives without fulfilling true task intent. As models scale and optimization intensifies, such exploitation manifests as verbosity bias, sycophancy, hallucinated justification, benchmark overfitting, and, in multimodal settings, perception--reasoning decoupling and evaluator manipulation. Recent evidence further suggests that seemingly benign shortcut behaviors can generalize into broader forms of misalignment, including deception and strategic gaming of oversight mechanisms. In this survey, we propose the Proxy Compression Hypothesis (PCH) as a unifying framework for understanding reward hacking. We formalize reward hacking as an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. Under this view, reward hacking arises from the interaction of objective compression, optimization amplification, and evaluator--policy co-adaptation. This perspective unifies empirical phenomena across RLHF, RLAIF, and RLVR regimes, and explains how local shortcut learning can generalize into broader forms of misalignment, including deception and strategic manipulation of oversight mechanisms. We further organize detection and mitigation strategies according to how they intervene on compression, amplification, or co-adaptation dynamics. By framing reward hacking as a structural instability of proxy-based alignment under scale, we highlight open challenges in scalable oversight, multimodal grounding, and agentic autonomy.
