INTERESTING TTS WORK ACCEPTED BY ICML 2026

MODEL ROUTING AND TEST-TIME SCALING HAVE BEEN TREATED AS TWO SEPARATE DIMENSIONS, AND THAT DECOUPLING IS EXACTLY THE BOTTLENECK. UNISCALE UNIFIES THEM IN A SINGLE OPTIMIZATION SPACE. FORMULATES ADAPTIVE UNIFIED INFERENCE SCALING AS A CONTEXTUAL MULTI-ARMED BANDIT, LEARNS INFERENCE POLICIES VIA LINUCB, ADDS EFFICIENCY-AWARE LEARNING AND COST MODELING TO STAY STABLE OVER HIGH-DIMENSIONAL ACTION SPACES. THE RESULT: A SMOOTHER ACCURACY-COST PARETO WHER

High-priority alert based on source significance and keyword relevance.