{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiaqnn3tsjv4gxwbmo23m6gf5oi6u2e4znxvsjit6ak6lodd4tanlm",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mput5gouji72"
  },
  "path": "/t/presenting-tis-token-importance-scoring-a-new-way-to-compress-kv-cache/177429#post_14",
  "publishedAt": "2026-07-05T04:03:56.000Z",
  "site": "https://discuss.huggingface.co",
  "textContent": "Interesting connection. I agree there is a shared concern around token/evidence importance. But I would place DESi one architectural layer earlier than GLM-5/DSA. DSA optimizes attention after tokens have entered the model; DESi tries to decide which evidence/state should enter the context before prefill. So I see them as complementary, but not equivalent: GLM-5 reduces attention cost inside the model, while DESi reduces epistemic and token pressure before the model is invoked.",
  "title": "Presenting TIS (Token Importance Scoring) - A new way to compress KV cache"
}