TRIM: Hybrid Inference via Targeted Stepwise Routing in Multi-Step Reasoning Tasks

Vansh Kapoor, Aman Gupta, Hao Chen +3

Multi-step reasoning tasks like mathematical problem solving are vulnerable to cascading failures, where a single incorrect step leads to complete solution breakdown. Current LLM r…