Lost in the Fallback: How Reliably Do Typed Decision Models Use Fallback Options?
We find that correct recognition does not reliably transfer to correct fallback options, and errors persist even when the correct category is explicitly supplied.Moreover, logically equivalent formulations of the same fallback rule can also produce substantial disagreement in model decisions.
ProofPaper ↗
Key points
- Typed decision models (TDMs) are increasingly used in applications such as request routing and model selection, making reliable decisions critical to system performance.
- While recent works have examined adversarial vulnerabilities in TDMs, their ability to follow fallback rules in non-adversarial settings remains underexplored.
- We introduce the Paired Fallback Robustness (PFR) Test to evaluate two aspects of fallback reliability: whether models correctly switch between regular and fallback options, and whether they make consistent decisions under logically equivalent definitions of the same rule.
- These findings reveal a crucial reliability gap in current TDMs: fallback mechanisms designed to improve reliability can instead introduce errors.
Sources (1)
- [1]Lost in the Fallback: How Reliably Do Typed Decision Models Use Fallback Options?arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 09:21 AM
We find that correct recognition does not reliably transfer to correct fallback options, and errors persist even when the correct category is explicitly supplied.Moreover, logically equivalent formulations of the same fallback rule can also produce substantial disagreement in model decisions.
Typed decision models (TDMs) are increasingly used in applications such as request routing and model selection, making reliable decisions critical to system performance.
Extractive summary: sentences quoted from the sources.