Goal Misgeneralization in Deep Reinforcement Learning
Langosco, Lauro, Koch, Jack, Sharkey, Lee, Pfau, Jacob, Orseau, Laurent, Krueger, DavidarXiv.orgPublished May 28, 2021Original︎ ↗Archived copy
Examinations
None recorded.
Text66 passages · 1 marked
Abstract: We study goal misgeneralization, a type of out-of-distribution generalization failure in reinforcement learning (RL). Goal misgeneralization failures occur when an RL agent retains its capabilities out-of-distribution yet pursues the wrong goal. For instance, an agent might continue to competently avoid obstacles, but navigate to the wrong place. In contrast, previous works have typically focused on capability generalization failures, where an agent fails to do anything sensible at test time. We formalize this distinction between capability and goal generalization, provide the first empirical demonstrations of goal misgeneralization, and present a partial characterization of its causes.
Not yet read by this claim’s Steward.
Lineage across 1 claim
Record
- Kind
- unknown
- Authors
- Langosco, Lauro, Koch, Jack, Sharkey, Lee, Pfau, Jacob, Orseau, Laurent, Krueger, David
- Publisher
- arXiv.org
- Published
- May 28, 2021
- DOI
- None
- Link
- https://arxiv.org/abs/2105.14111
- Archived copy
- https://web.archive.org/web/20261011231757/https://arxiv.org/abs/2105.14111
- Stored text
- 4,679 characters, retrieved Nov 1, 2026 · 49 sections, 66 passages
- Claims extracted
- For
Demonstrations and definition of goal misgeneralization in reinforcement learning agents, and how it differs from capability failure.
· 2 claims (1 new) · Nov 1, 2026 for Goal misgeneralization: evidence in current language models and likelihood in advanced AI systems
- For
- Versions
- One known
- Copies
- None recorded
- Corrections
- None recorded
- Facts checked
- Nov 1, 2026
History 3
- Nov 1, 2026DividedSegmenter · Divided into 115 parts by its own structure
- Nov 1, 2026FactsFacts fetcher · Facts looked up
- Nov 1, 2026Entered the graphEntered the graph; copy stored