{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreigzsz6vvmu3kc7khwpoglxqvh2omrkkodxzw6pzqth5edvyh7w6kq",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mptkvf2wehu2"
  },
  "path": "/t/presenting-tis-token-importance-scoring-a-new-way-to-compress-kv-cache/177429#post_10",
  "publishedAt": "2026-07-04T15:53:08.000Z",
  "site": "https://discuss.huggingface.co",
  "textContent": "I see the combined benchmark as interesting, but not central to my current work.\n\nMy main focus is upstream admission control: deciding which evidence/state should enter the model context before prefill. From that perspective, KV-cache compression is a downstream optimization for cases where long prefill has already happened or cannot be avoided.\n\nSo I would not currently prioritize implementing TIS integration myself. It is a valid experiment, but it addresses a different intervention point.\n\nIf Phase 4 of TIS moves in that direction, I would be interested to see whether it adds measurable value on top of upstream evidence/state selection. But for DESi, the primary question remains: can we avoid unnecessary context admission in the first place?",
  "title": "Presenting TIS (Token Importance Scoring) - A new way to compress KV cache"
}