{
  "paper": {
    "arxiv_id": "2608.25528",
    "title": "TransRetrieval"
  },
  "dataset": {
    "name": "MovieLens-1M",
    "users": 320,
    "items": 480
  },
  "setup": {
    "seed": 42,
    "full_catalog": true,
    "history_length": 20
  },
  "variants": {
    "Transformer two-tower": {
      "hit_at_10": 0.10625,
      "ndcg_at_10": 0.05428956179752216,
      "head_share_at_10": 0.0596875
    },
    "TransRetrieval core": {
      "hit_at_10": 0.11875,
      "ndcg_at_10": 0.05472929672758298,
      "head_share_at_10": 0.0571875
    }
  },
  "relative": {
    "hit_at_10_percent": 11.76470588235294,
    "ndcg_at_10_percent": 0.8099806215066692,
    "head_share_at_10_percent": -4.188481675392662
  },
  "diagnostics": {
    "token_norm_std_before": 8.258790355892755e-09,
    "token_norm_std_after": 6.929635869225387e-17,
    "target_tokens_before": 8,
    "target_tokens_after": 1,
    "target_flops_reduction_percent": 85.0
  },
  "paper_results": {
    "revenue_lift_percent": 2.53,
    "rpm_lift_percent": 1.28
  },
  "scope": "在 MovieLens-1M 上执行 norm-aware weighted aggregation、8→1 target token compression 与 position-style domain embedding；未复刻 400 亿交互、52M 广告库和生产 ANN serving。",
  "runtime": {
    "requested_device": "auto",
    "cpu_threads": null,
    "platform": "Darwin arm64"
  },
  "seed": 42,
  "schema_version": 2,
  "manifest": {
    "adapter_key": "transretrieval",
    "arxiv_id": "2608.25528",
    "title": "TransRetrieval: Scaling Up Transformer-Based Retrieval for Industrial Recommendation",
    "paper_url": "https://arxiv.org/abs/2608.25528",
    "track": "recommendation",
    "organization": "Renmin University of China / Taobao & Tmall Group, Alibaba",
    "published": "2026-08-26",
    "code_url": null,
    "topics": [
      "retrieval",
      "transformer-scaling",
      "target-token-compression",
      "multi-domain"
    ],
    "local_code_dir": "src/auto_research/reproductions/transretrieval",
    "fidelity": "core_mechanism",
    "evaluation_tier": "l2_public_dataset",
    "datasets": [
      "MovieLens-1M"
    ],
    "baseline": "same-feature Transformer-style two-tower mean aggregation",
    "metrics": [
      "Hit@10",
      "NDCG@10",
      "head share@10",
      "token norm dispersion"
    ],
    "default_seeds": [
      42
    ],
    "budget": "adapter-defined fixed budget",
    "device_capabilities": [
      "cpu"
    ],
    "online_evidence": [
      {
        "product": "Alibaba display advertising",
        "metric": "platform revenue",
        "lift_percent": 2.53,
        "traffic": "5% traffic; month-long A/B; p<0.0001",
        "source_url": "https://arxiv.org/html/2608.25528v1#S4.SS5",
        "source_location": "Section 4.5 / Table 7",
        "experiment_duration": "one month",
        "significance": "p<0.0001",
        "retrieved_at": "2026-08-28"
      },
      {
        "product": "Alibaba display advertising",
        "metric": "RPM",
        "lift_percent": 1.28,
        "traffic": "5% traffic; month-long A/B; p<0.0001",
        "source_url": "https://arxiv.org/html/2608.25528v1#S4.SS5",
        "source_location": "Section 4.5 / Table 7",
        "experiment_duration": "one month",
        "significance": "p<0.0001",
        "retrieved_at": "2026-08-28"
      }
    ],
    "selection_exception": null,
    "evolve_operators": []
  },
  "provenance": {
    "created_at": "2026-08-27T16:29:37.484173+00:00",
    "code_commit": "cfb01264c6575f3d1634226940df384a3632a036",
    "python": "3.14.5",
    "platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
    "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
    "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
    "packages": {
      "numpy": "2.4.3"
    },
    "artifact_path": "docs/reproductions/2608.25528-transretrieval/metrics/movielens-1m-seed42.json"
  },
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "tier_label": "L2 公开数据集训练",
    "seeds": [
      42
    ],
    "budget": "paper-specific",
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "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": [
      "40B industrial interactions",
      "52M-item advertisement corpus",
      "production ANN serving"
    ]
  },
  "manifest_ref": "reproduction:transretrieval"
}