ZLKong/Awesome-Collection-Token-Reduction

What it solves

Token reduction (pruning, merging, clustering, compressing, adaptive thinking) is a set of techniques that reduce the number of tokens processed by machine learning and generative AI models, making them faster and more efficient without sacrificing quality. This project curates the latest academic work in this fast-moving field, so researchers and engineers can easily find relevant papers and track new developments.

How it works

The repository is a manually maintained collection of papers, organized by publication year and application area (Vision, Language, Multimodal, Agentic Systems, Hardware Co-design, State Space Models). Each entry includes a link to the paper and, when available, a link to open-source code. A companion Google Sheet provides structured summaries of each paper's task, token reduction type, contribution, and methodology. The project also hosts a position paper (arXiv:2505.18227) that argues token reduction should go beyond efficiency and outlines future research directions. Contributions are encouraged via issues and pull requests.

Who it's for

This resource is designed for AI researchers, graduate students, and practitioners who work on or follow token efficiency in vision, language, and multimodal models. It is also useful for anyone looking to apply token reduction techniques to their own projects and needing a quick survey of the state of the art.

Highlights

• Covers token pruning, merging, clustering, compression, and adaptive thinking across multiple modalities and application scenarios.
• Organized by year (2024–2026) and area, making it easy to scan recent progress.
• Includes a position paper that frames token reduction as a broader tool than just speed-up, with future directions.
• Links to many open-source implementations, enabling hands-on experimentation.
• Actively maintained, with frequent updates (e.g., a new Agentic Systems section added in 2026) and an open invitation for community contributions.