Paper page - Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges

Source: https://huggingface.co/papers/2604.13602 Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

Reward hacking in aligned language models stems from optimizing expressive policies against compressed reward signals, leading to systematic misalignment behaviors that generalize beyond initial shortcuts.

Reinforcement Learning from Human Feedback(RLHF) and related alignment paradigms have become central to steering large language models (LLMs) andmultimodal large language models(MLLMs) toward human-preferred behaviors. However, these approaches introduce a systemic vulnerability:reward hacking, where models exploit imperfections in learnedreward signalsto maximizeproxy objectiveswithout 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, includingdeceptionandstrategic gamingof oversight mechanisms. In this survey, we propose the Proxy Compression Hypothesis (PCH) as a unifying framework for understandingreward hacking. We formalizereward hackingas an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. Under this view,reward hackingarises 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, includingdeceptionand 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 framingreward hackingas a structural instability of proxy-based alignment under scale, we highlight open challenges inscalable oversight,multimodal grounding, andagentic autonomy.

View arXiv pageView PDFProject pageGitHub9Add to collection

Get this paper in your agent:

hf papers read 2604\.13602

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2604.13602 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2604.13602 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.13602 in a Space README.md to link it from this page.

Collections including this paper1