Shaoqiu Zhang

Quantized reasoning models overthink more, negating efficiency gains, but a new method precisely targets and penalizes these "overthinking" tokens without retraining.

Video segmenters suffer from persistent identity drift because they blindly write noisy observations to memory; gating state transitions through propose-verify graph reasoning preserves object history through extended occlusions without altering the underlying base model.

ET-Prune achieves a remarkable balance between efficiency and accuracy, outperforming traditional pruning methods by retaining critical evidence while cutting down on unnecessary tokens.