JevAdvBench: A Benchmark and Black-Box Attacks for Reinforcement Learning for Calibrated Decisions Models
JevAdvBench shows appended opinions flip 12.1% of Jev decisions and sink many below review confidence.
JevAdvBench is an adversarial benchmark for reinforcement-learning-for-calibrated-decisions models, with 812 typed questions over 66 scenarios and 9,744 single-edit black-box variants. On jev-1.13.0, rewording stays within 1.2 percentage points of a re-run baseline, and fields outside the schema never reach the model. One unverified opinion appended to the state flips 12.1% of decisions, statistically tied with the strongest injected command at 10.1%, and pushes 38% of confident answers below the 0.8 human-review threshold. The authors argue that applications should treat model state as untrusted input.
- Benchmark has 812 questions and 9,744 single-edit black-box variants.
- Rewording stays within 1.2 points of the re-run baseline.
- Appended opinions flip 12.1% of jev-1.13.0 decisions.
- 38% of confident answers fall below the 0.8 review threshold.
From the original article
1 min read at arxiv.orgModels trained with reinforcement learning for calibrated decisions (RLCD), such as Jev, answer a typed question about an input, the state, with a probability, a choice, or a score, and software acts on the answer without a person reading it.
By Jianyi Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.