The authors construct DIB-Bench, a protocol that applies six classes of deployment tampering to LLM-agent stacks across three model pairs and three random seeds, yielding 7,560 scored examples.
Stronger LLM judges accept more tampered outputs, not fewer
In 7,560 scored examples, tampered-output acceptance rises from 88.6% for rule-based scorers to 94.1% for GPT-5-nano, and trigger-conditional tampering yields up to 100% harmful outputs.
Academic
Jianchang Su · Dawei Xiang · Yifan Zhang · Wei Zhang
University of Connecticut
Research Digest··2 min read
The authors introduce DIB-Bench, a benchmark for measuring whether LLM evaluators detect deployment tampering.
Why this paper
From University of Connecticut
In one line
Output-only evaluation cannot detect deployment tampering; stronger judges accept more tampered outputs, so identity evidence is required.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§