Timezone: »
Reward hacking---where RL agents exploit gaps in misspecified proxy rewards---has been widely observed, but not yet systematically studied. To understand reward hacking, we construct four RL environments with different misspecified rewards. We investigate reward hacking as a function of agent capabilities: model capacity, action space resolution, and observation space noise. Typically, more capable agents are able to better exploit reward misspecifications, causing them to attain higher proxy reward and lower true reward. Moreover, we find instances of \emph{phase transitions}: capability thresholds at which the agent's behavior qualitatively shifts, leading to a sharp decrease in the true reward. Such phase transitions pose challenges to monitoring the safety of ML systems. To encourage further research on reward misspecification, we propose an anomaly detection task for aberrant policies and offer several baseline detectors.
Author Information
Alexander Pan (California Institute of Technology)
Kush Bhatia (UC Berkeley)
Jacob Steinhardt (UC Berkeley)
More from the Same Authors
-
2021 Spotlight: Learning Equilibria in Matching Markets from Bandit Feedback »
Meena Jagadeesan · Alexander Wei · Yixin Wang · Michael Jordan · Jacob Steinhardt -
2021 : Measuring Coding Challenge Competence With APPS »
Dan Hendrycks · Steven Basart · Saurav Kadavath · Mantas Mazeika · Akul Arora · Ethan Guo · Collin Burns · Samir Puranik · Horace He · Dawn Song · Jacob Steinhardt -
2021 : PixMix: Dreamlike Pictures Comprehensively Improve Safety Measures »
Dan Hendrycks · Andy Zou · Mantas Mazeika · Leonard Tang · Dawn Song · Jacob Steinhardt -
2021 : Effect of Model Size on Worst-group Generalization »
Alan Pham · Eunice Chan · Vikranth Srivatsa · Dhruba Ghosh · Yaoqing Yang · Yaodong Yu · Ruiqi Zhong · Joseph Gonzalez · Jacob Steinhardt -
2021 : What Would Jiminy Cricket Do? Towards Agents That Behave Morally »
Dan Hendrycks · Mantas Mazeika · Andy Zou · Sahil Patel · Christine Zhu · Jesus Navarro · Dawn Song · Bo Li · Jacob Steinhardt -
2021 : Measuring Mathematical Problem Solving With the MATH Dataset »
Dan Hendrycks · Collin Burns · Saurav Kadavath · Akul Arora · Steven Basart · Eric Tang · Dawn Song · Jacob Steinhardt -
2022 : Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations »
Yongyi Yang · Jacob Steinhardt · Wei Hu -
2022 : Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small »
Kevin Wang · Alexandre Variengien · Arthur Conmy · Buck Shlegeris · Jacob Steinhardt -
2023 Poster: Jailbroken: How Does LLM Safety Training Fail? »
Alexander Wei · Nika Haghtalab · Jacob Steinhardt -
2023 Poster: Supply-Side Equilibria in Recommender Systems »
Meena Jagadeesan · Nikhil Garg · Jacob Steinhardt -
2023 Poster: A case for reframing automated medical image classification as segmentation »
Sarah Hooper · Mayee Chen · Khaled Saab · Kush Bhatia · Curtis Langlotz · Christopher Ré -
2023 Poster: TART: A plug-and-play Transformer module for task-agnostic reasoning »
Kush Bhatia · Avanika Narayan · Christopher De Sa · Christopher Ré -
2023 Poster: Mass-Producing Failures of Multimodal Models »
Shengbang Tong · Erik Jones · Jacob Steinhardt -
2023 Poster: Goal Driven Discovery of Distributional Differences via Language Descriptions »
Ruiqi Zhong · Peter Zhang · Steve Li · Jinwoo Ahn · Dan Klein · Jacob Steinhardt -
2023 Poster: Skill-it! A data-driven skills framework for understanding and training language models »
Mayee Chen · Nicholas Roberts · Kush Bhatia · Jue WANG · Ce Zhang · Frederic Sala · Christopher Ré -
2023 Poster: Improved Bayes Risk Can Yield Reduced Social Welfare Under Competition »
Meena Jagadeesan · Michael Jordan · Jacob Steinhardt · Nika Haghtalab -
2023 Poster: Embroid: Unsupervised Prediction Smoothing Can Improve Few-Shot Classification »
Neel Guha · Mayee Chen · Kush Bhatia · Azalia Mirhoseini · Frederic Sala · Christopher Ré -
2023 Oral: Jailbroken: How Does LLM Safety Training Fail? »
Alexander Wei · Nika Haghtalab · Jacob Steinhardt -
2022 Workshop: Workshop on Machine Learning Safety »
Dan Hendrycks · Victoria Krakovna · Dawn Song · Jacob Steinhardt · Nicholas Carlini -
2022 Poster: How Would The Viewer Feel? Estimating Wellbeing From Video Scenarios »
Mantas Mazeika · Eric Tang · Andy Zou · Steven Basart · Jun Shern Chan · Dawn Song · David Forsyth · Jacob Steinhardt · Dan Hendrycks -
2022 Poster: Capturing Failures of Large Language Models via Human Cognitive Biases »
Erik Jones · Jacob Steinhardt -
2022 Poster: Forecasting Future World Events With Neural Networks »
Andy Zou · Tristan Xiao · Ryan Jia · Joe Kwon · Mantas Mazeika · Richard Li · Dawn Song · Jacob Steinhardt · Owain Evans · Dan Hendrycks -
2021 Poster: Grounding Representation Similarity Through Statistical Testing »
Frances Ding · Jean-Stanislas Denain · Jacob Steinhardt -
2021 Poster: Learning Equilibria in Matching Markets from Bandit Feedback »
Meena Jagadeesan · Alexander Wei · Yixin Wang · Michael Jordan · Jacob Steinhardt -
2020 Poster: Online learning with dynamics: A minimax perspective »
Kush Bhatia · Karthik Sridharan -
2020 Poster: Preference learning along multiple criteria: A game-theoretic perspective »
Kush Bhatia · Ashwin Pananjady · Peter Bartlett · Anca Dragan · Martin Wainwright -
2018 Workshop: Workshop on Security in Machine Learning »
Nicolas Papernot · Jacob Steinhardt · Matt Fredrikson · Kamalika Chaudhuri · Florian Tramer -
2018 Poster: Gen-Oja: Simple & Efficient Algorithm for Streaming Generalized Eigenvector Computation »
Kush Bhatia · Aldo Pacchiano · Nicolas Flammarion · Peter Bartlett · Michael Jordan -
2018 Poster: FastGRNN: A Fast, Accurate, Stable and Tiny Kilobyte Sized Gated Recurrent Neural Network »
Aditya Kusupati · Manish Singh · Kush Bhatia · Ashish Kumar · Prateek Jain · Manik Varma -
2018 Poster: Semidefinite relaxations for certifying robustness to adversarial examples »
Aditi Raghunathan · Jacob Steinhardt · Percy Liang -
2017 Workshop: Aligned Artificial Intelligence »
Dylan Hadfield-Menell · Jacob Steinhardt · David Duvenaud · David Krueger · Anca Dragan -
2017 Workshop: Machine Learning and Computer Security »
Jacob Steinhardt · Nicolas Papernot · Bo Li · Chang Liu · Percy Liang · Dawn Song -
2017 Poster: Consistent Robust Regression »
Kush Bhatia · Prateek Jain · Parameswaran Kamalaruban · Purushottam Kar -
2017 Poster: Certified Defenses for Data Poisoning Attacks »
Jacob Steinhardt · Pang Wei Koh · Percy Liang -
2016 : Opening Remarks »
Jacob Steinhardt -
2016 Workshop: Reliable Machine Learning in the Wild »
Dylan Hadfield-Menell · Adrian Weller · David Duvenaud · Jacob Steinhardt · Percy Liang -
2015 Poster: Learning with Relaxed Supervision »
Jacob Steinhardt · Percy Liang