← All threads

Token Pruning Efficiency

Methods for identifying and removing low-value tokens in language and vision-language models to reduce computational cost while maintaining reasoning performance.

10 papers

Where this stands

The written synthesis of this thread is for subscribers. Subscribe.

How this thread developed

  1. Oct 2026 · Central South University, SenseTime Research

    Simple clustering preserves video understanding with far fewer visual tokens

    Proposes SimpleCluster, a clustering-based token reduction method for video understanding.

    released code

  2. Oct 2026 · University of Toronto, Vector Institute

    Self-distillation helps vision-language models reason with far fewer visual tokens

    Demonstrates that self-distillation allows VLMs to maintain reasoning performance after aggressive visual token pruning.

  3. Oct 2026 · Purdue University, Rochester Institute of Technology

    Language models can cut reasoning tokens by targeting low-value text

    Shows that normalized log probabilities can identify low-value tokens in reasoning chains, enabling token reduction.

  4. 3 further papers

    Oct 2026 · Tencent, Code Models

    Pruning experts by replacement damage better preserves MoE reasoning performance

    Prunes MoE experts rather than tokens but achieves analogous inference speedup with minimal reasoning loss, expanding token-pruning ideas to expert removal.

  5. 2 further papers

    Oct 2026 · Amazon AGI

    Token graph communities speed long-context decoding without model retraining

    Proposes CommunityKV, a training-free sparse attention method that prunes low-value tokens based on token graph communities for long-context decoding.

5 of 10 papers shown