MALT: A Dataset of Natural and Prompted Behaviors That Threaten Eval Integrity - METR
MALT (Manually-reviewed Agentic Labeled Transcripts) is a dataset of natural and prompted examples of behaviors that threaten evaluation integrity (like generalized reward hacking or sandbagging).
MALT: A Dataset of Natural and Prompted Behaviors That Threaten Eval Integrity - METR Our Work Research Notes Updates Risk Assessment About Donate Careers Search --> Our Work Research Notes Updates Risk Assessment About Donate Careers Menu × MALT: A Dataset of Natural and Prompted Behaviors That Threaten Eval Integrity CONTRIBUTORS Neev Parikh and Hjalmar Wijk DATE October 14, 2025 SHARE Copy Link Citation BibTeX Citation × @misc { metr-2025-malt-dataset-of-natural-and-prompted-behaviors , title = {MALT: A Dataset of Natural and Prompted Behaviors That Threaten Eval Integrity} , au
saved by
related reading
- Early work on monitorability evaluations - METRmetr.org
- Agentic misalignment: How LLMs could be insider threats \ Anthropicanthropic.com
- Surfacing Pathological Behaviors in Language Models | Transluce AItransluce.org
- An alignment assessment of recent cybersecurity incidentsanthropic.com
- Reward Hacking in Reinforcement Learning | Lil'Loglilianweng.github.io
- Training a Misaligned Reward Seekeralignment.anthropic.com
- Natural-emergent-misalignment-from-reward-hacking-paper.pdfassets.anthropic.com
- Recent Frontier Models Are Reward Hacking - METRmetr.org
- [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervisionarxiv.org
- Metagaming matters for training, evaluation, and oversightalignment.openai.com
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscationarxiv.org
- [2604.16812] Introspection Adapters: Training LLMs to Report Their Learned Behaviorsarxiv.org