Disagreement-Guided Weak-to-Strong Routing
We first characterize the optimal weak-to-strong routing policy, while allowing for multiple weak-model responses generated through local inference.
ProofPaper ↗
Key points
- Small language models are increasingly deployed on edge devices and robotic systems to meet cost and latency constraints, yet their limited capabilities can make reliable deployment challenging.
- We revisit the foundations of weak-to-strong routing and ask whether routing decisions can instead be guided by signals that do not require ground-truth supervision.
- In particular, we introduce a notion of model disagreement that can be measured directly from model-generated responses, without ground-truth supervision.
- Moreover, even when ground-truth correctness signals are available for training, disagreement-based routing can outperform methods trained directly to predict correctness
Sources (1)
- [1]Disagreement-Guided Weak-to-Strong RoutingarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 05:19 PM
We first characterize the optimal weak-to-strong routing policy, while allowing for multiple weak-model responses generated through local inference.
Small language models are increasingly deployed on edge devices and robotic systems to meet cost and latency constraints, yet their limited capabilities can make reliable deployment challenging.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026Prompt tips from Anthropic
- Oct 15, 2026Each Reward to Its Relevant Tokens: Reward-Specific Token-Level Credit Assignment for Multi-Reward RL
- Oct 15, 2026Learning to Think and Explore in 3D World
- Oct 15, 2026CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness
- Oct 15, 2026MJev-Spatial: System One Spatial Intelligence via Native Attention Matching
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models