AI/ML Research Digest: self-distillation, MoE pruning and long context

A research digest dated Oct 10, 2026 covers on-policy self-distillation for spatial and multimodal reasoning, serving-time MoE expert pruning with up to 1.81x faster decoding, TokenRouter boosting decoding throughput 2-64x, and V-CoLA dropping half the vision tokens while retaining 99.5% of performance.

• Self-distillation of 3D cues boosts VSI-Bench without 3D sensors
• Serving-time MoE expert pruning speeds decoding up to 1.81x
• TokenRouter raises decoding throughput by 2-64x
• V-CoLA drops half the vision tokens, keeps 99.5% performance