Disagreement-Guided Routing Makes Weak Models Defer
University of Pennsylvania researchers introduce a ground-truth-free routing method that uses disagreement between weak and strong language models to decide when escalation is worthwhile. Tests across agentic environments and reasoning benchmarks show improved cost–accuracy trade-offs over uncertainty and consistency baselines.
This paper reframes model disagreement from a failure signal into a practical inference control surface for hybrid AI systems.
- –Corrected disagreement combines weak–strong disagreement with the strong model’s self-consistency.
- –The method routes agent actions without step-level correctness labels or trajectory rewards.
- –It outperforms label-free baselines across WebShop, BFCL, MATH-500, Spider, and AceCode.
- –The approach is especially valuable for edge and robotic deployments where strong-model calls are expensive.
- –The official repository is public, but implementation code is not yet released.
DISCOVERED
1h ago
2026-10-16
PUBLISHED
2h ago
2026-10-16
RELEVANCE
AUTHOR
arnaudmercier