{
  "paper": {
    "arxiv_id": "2606.25147",
    "title": "TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems",
    "url": "https://arxiv.org/abs/2606.25147",
    "organization": "Google DeepMind / YouTube"
  },
  "dataset": {
    "name": "MovieLens-1M",
    "users": 320,
    "items": 480
  },
  "setup": {
    "seed": 42,
    "steps_per_model": 120,
    "maximum_history": 32,
    "sid_levels": 2,
    "sid_cardinality": 16,
    "same_split_optimizer_and_full_catalog": true
  },
  "baseline": {
    "name": "dense-only sequential user encoder",
    "hit_at_10": 0.034375,
    "ndcg_at_10": 0.020270109562740432,
    "head_share_at_10": 0.105
  },
  "method": {
    "name": "TokenMinds dense embedding + generated SID user tokens",
    "hit_at_10": 0.034375,
    "ndcg_at_10": 0.02019862876969453,
    "head_share_at_10": 0.1034375
  },
  "relative": {
    "hit_at_10_percent": 0.0,
    "ndcg_at_10_percent": -0.35264137485124736,
    "head_share_at_10_percent": -1.4880952380952328
  },
  "training": {
    "baseline": {
      "initial_ranking_loss": 7.678776359558105,
      "final_ranking_loss": 6.986705493927002,
      "final_sid_loss": null,
      "parameters": 22272,
      "device": "cpu"
    },
    "tokenminds": {
      "initial_ranking_loss": 7.683404588699341,
      "final_ranking_loss": 6.9884271144866945,
      "final_sid_loss": 2.6777411699295044,
      "parameters": 24353,
      "device": "cpu"
    }
  },
  "paper_results": {
    "sfv_engaged_users_percent": 0.11,
    "sfv_satisfied_engagement_percent": 0.62,
    "training_compute_reduction_percent": 50,
    "serving_compute_reduction_percent": 31,
    "full_user_traffic": true
  },
  "reproduction_fidelity": {
    "level": "core_mechanism",
    "label": "核心机制复现",
    "description": "论文中心算法被实际执行，但生产模型、私有特征或基础设施未复刻。",
    "omitted_core_components": [
      "Gemini 370M MoE encoder and 370M decoder with PLUM CPT",
      "YouTube LFV/SFV/search private training corpus",
      "multi-context beam decoding and asynchronous billion-user serving"
    ]
  },
  "schema_version": 2,
  "manifest_ref": "reproduction:tokenminds",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2606.25147-tokenminds/metrics/movielens-1m-seed42.json",
    "code_commit": "working tree before commit",
    "dataset": "local public MovieLens-1M",
    "created_at": "2026-08-09T00:00:00+08:00",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}