policy

reward-hacking-misalignment

UKGovernmentBEIS · PyTorch

or hover any field below to flag it

Overview

Name
reward-hacking-misalignment
Author
UKGovernmentBEIS
Framework
PyTorch
License
MIT
Skill type
other
Evidence level
untested
Task description
Reproducing "Natural Emergent Misalignment from Reward Hacking" (MacDiarmid et al., Anthropic 2025) with open-source models. Includes reward-hackable RL environments, misalignment evaluations, training configs, and evaluation scripts. Models trained on OLMo (7B, 32B) and GPT-OSS (20B, 120B).

Spaces

Action space
other · 0-dim · 0Hz
Observation space
  • type: other

Links

HuggingFace repo
null
Paper (arXiv)
null

Compatible robots

3+17 mentioned but not in catalog yet

Compatible environments

0

No environments list reward-hacking-misalignment yet.

Datasets that reference this policy

0

No datasets reference reward-hacking-misalignment yet.