{
  "id": "2304.08467",
  "title": "Learning to Compress Prompts with Gist Tokens",
  "authors": "Jesse Mu, Xiang Lisa Li, Noah Goodman",
  "first_public": "2023-04-17",
  "first_public_source": "https://arxiv.org/abs/2304.08467",
  "arxiv_first_submission": "2023-04-17",
  "date_label": "arXiv v1",
  "arxiv": "2304.08467",
  "paper_url": "https://arxiv.org/abs/2304.08467",
  "venue": "NeurIPS 2023",
  "lens": [
    "context-compression"
  ],
  "related_terms": [
    "gist tokens",
    "prompt compression",
    "activation caching",
    "amortized inference"
  ],
  "summary": "Compresses an instruction into a few reusable token activations.",
  "contribution": "Trains with an attention mask that forces later tokens to use gist activations in place of the original prompt.",
  "limitations": "Compression can reduce output quality. The cached activations are tied to the trained model.",
  "program_form": "Cached activations at a few gist-token positions.",
  "conditioning": "The instruction or prompt.",
  "execution": "The language model generates outputs while attending to the cached gist.",
  "relationship": "PAW cites gisting as another way to reuse instruction computation. Gisting stores activations; PAW stores a program with text and adaptation parameters.",
  "relationship_source": "https://arxiv.org/html/2607.02512v1#S10",
  "provenance": "arXiv record and paper.",
  "author_names": [
    "Jesse Mu",
    "Xiang Lisa Li",
    "Noah Goodman"
  ],
  "record_url": "https://programasweights.com/reading-list/papers/2304.08467/",
  "legacy_record_url": "https://programasweights.com/reading-list/papers/#paper-2304-08467",
  "citation_date": "2023-04-17",
  "json_url": "https://programasweights.com/reading-list/papers/2304.08467/record.json",
  "markdown_url": "https://programasweights.com/reading-list/papers/2304.08467/index.md",
  "bibtex_url": "https://programasweights.com/reading-list/papers/2304.08467/citation.bib",
  "csl_url": "https://programasweights.com/reading-list/papers/2304.08467/citation.csl.json",
  "publication": {
    "title": "Learning to Compress Prompts with Gist Tokens",
    "author_names": [
      "Jesse Mu",
      "Xiang Li",
      "Noah Goodman"
    ],
    "year": 2023,
    "type": "paper-conference",
    "container_title": "Advances in Neural Information Processing Systems",
    "url": "https://proceedings.neurips.cc/paper_files/paper/2023/hash/3d77c6dcc7f143aa2154e7f4d5e22d68-Abstract-Conference.html",
    "source_url": "https://proceedings.neurips.cc/paper_files/paper/2023/hash/3d77c6dcc7f143aa2154e7f4d5e22d68-Abstract-Conference.html",
    "verified_at": "2026-10-19",
    "volume": "36",
    "doi": "10.52202/075280-0848",
    "pages": "19327-19352"
  },
  "publication_bibtex_url": "https://programasweights.com/reading-list/papers/2304.08467/publication.bib",
  "publication_csl_url": "https://programasweights.com/reading-list/papers/2304.08467/publication.csl.json",
  "method_comparisons": [
    {
      "title": "What gets reused?",
      "text": "Gisting caches a few instruction activations. Generative Adapter stores context in parameter updates. PAW saves a function as a program containing text and a neural module.",
      "source_ids": [
        "2304.08467",
        "2411.05877",
        "2607.02512"
      ],
      "source_urls": [
        "https://arxiv.org/abs/2304.08467",
        "https://arxiv.org/abs/2411.05877",
        "https://arxiv.org/html/2607.02512v1#S2"
      ],
      "source_labels": [
        "Gist paper",
        "Generative Adapter paper",
        "PAW program definition"
      ],
      "topic_id": "context-compression",
      "topic_label": "Context compression",
      "topic_url": "https://programasweights.com/reading-list/guides/compilation/#context-compression"
    }
  ]
}