SCRouter: Selecting models and budgets based on safety and utility conditions
Start here
On October 13, 2026, researchers published a paper on arXiv proposing the SCRouter router: for each query, it predicts utility and safety separately, and combines cost to select the model and token budget. It estimates the probability that a response will be completed within the budget for each model-budget pair, and uses lightweight probes to determine the weights of the two types of predictions. The accompanying benchmark SCRouterBench contains 38,869 utility and safety queries, each answered under 144 combinations of 12 models × 12 budgets; SCRouter improves baselines by 13-29 points in-domain and 19-23 points out-of-domain.
Generated by AI from reports · Updated 29 days ago
Attention and developments
Attention, hourly
Now 23, 0% in 24 hours·Peak 23 (Nov 6 03:00)
Related research
This event's paper
- Rethinking Quality in Large Language Model Routing: Security-Conditioned Model-and-Budget Selection
arXiv · 2026-10-13 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesRethinking Quality in LLM Routing: Safety-Conditioned Model and Budget Selection
The researchers propose the SCRouter router, which predicts utility and safety for each query separately, combining cost to select the model and token budget. It estimates the probability that each model–budget pair completes the response within the budget, and uses lightweight probes to determine the weights of the two types of predictions. The accompanying benchmark SCRouterBench contains 38,869 utility and safety queries, each answered under 144 combinations of 12 models × 12 budgets; SCRouter improves baselines by 13–29 points in-domain and 19–23 points out-of-domain.