{
  "paper": {
    "arxiv_id": "2607.26427",
    "title": "PSG: Pair-Space Generation for Efficient Generative Reranking",
    "url": "https://arxiv.org/abs/2607.26427",
    "organization": "Kuaishou Technology"
  },
  "dataset": {
    "name": "MovieLens 100K",
    "users": 220,
    "items": 360
  },
  "setup": {
    "adapter": "psg",
    "same_split_and_candidates": true
  },
  "baseline": {
    "name": "item-space autoregressive next-item scoring",
    "hit_at_10": 0.10909090909090909,
    "ndcg_at_10": 0.05400546778896325,
    "fresh_hit_at_10": 0.031746031746031744,
    "head_share_at_10": 0.23772727272727273
  },
  "method": {
    "name": "ordered pair-token generation and unfolding",
    "hit_at_10": 0.09090909090909091,
    "ndcg_at_10": 0.04826018605777078,
    "fresh_hit_at_10": 0.07936507936507936,
    "head_share_at_10": 0.18954545454545454
  },
  "relative": {
    "hit_at_10_percent": -16.66666666666666,
    "ndcg_at_10_percent": -10.638333425132556,
    "fresh_hit_at_10_percent": 150.0,
    "head_share_at_10_percent": -20.26768642447419
  },
  "stages": {
    "item_space_steps": 6,
    "pair_space_steps": 3,
    "decode_step_reduction": 0.5,
    "request_pair_vocabulary": 129240,
    "generated_items": 6,
    "duplicate_items": 0
  },
  "paper_results": {
    "stay_time_lift_percent": 0.178,
    "industrial_speedup": 1.83,
    "traffic_percent_per_bucket": 10,
    "experiment_days": 7
  },
  "scope": "实际构造 request-specific n(n-1) 有序 pair space、pair token 打分、半长度解码与无重复展开；未复刻 Kuaishou GoalRank evaluator、曝光日志和线上 50-list beam serving。",
  "schema_version": 2,
  "manifest": {
    "adapter_key": "psg",
    "arxiv_id": "2607.26427",
    "title": "PSG: Pair-Space Generation for Efficient Generative Reranking",
    "paper_url": "https://arxiv.org/abs/2607.26427",
    "track": "recommendation",
    "organization": "Kuaishou Technology",
    "published": "2026-07-29",
    "code_url": null,
    "topics": [
      "reranking",
      "generative-recommendation",
      "pair-token",
      "serving"
    ],
    "local_code_dir": "src/auto_research/reproductions/industrial_2026",
    "fidelity": "core_mechanism",
    "evaluation_tier": "l2_public_dataset",
    "datasets": [
      "MovieLens 100K"
    ],
    "baseline": "item-space autoregressive scorer",
    "metrics": [
      "hit_at_10",
      "ndcg_at_10",
      "decode_step_reduction",
      "duplicate_items"
    ],
    "default_seeds": [
      42,
      43,
      44
    ],
    "budget": "220 users / 360 items; six-item slate",
    "device_capabilities": [
      "cpu",
      "mps",
      "cuda"
    ],
    "requires_gpu_validation": false,
    "gpu_validation_artifact": null,
    "online_evidence": [
      {
        "product": "Kuaishou single-column feed",
        "metric": "Stay Time per user",
        "lift_percent": 0.178,
        "traffic": "7-day test; two disjoint 10% traffic buckets",
        "source_url": "https://arxiv.org/html/2607.26427v1",
        "source_location": "Section 5 online A/B test, Table 8",
        "retrieved_at": "2026-09-01"
      }
    ],
    "selection_exception": null,
    "evolve_operators": []
  },
  "provenance": {
    "created_at": "2026-09-01T10:55:36.860473+00:00",
    "code_commit": "996ac89dd9bb16d0ac38be27d53639bc9d4ff859",
    "python": "3.12.9",
    "platform": "macOS-26.5.2-arm64-arm-64bit",
    "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
    "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
    "packages": {
      "auto-research": "0.1.0",
      "numpy": "2.5.2",
      "torch": "2.13.0",
      "transformers": "5.14.1"
    },
    "artifact_path": "docs/reproductions/2607.26427-psg/metrics/public-seeds42-44.json"
  },
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "tier_label": "L2 公开数据集训练",
    "seeds": [
      42,
      43,
      44
    ],
    "budget": "220 users / 360 items; six-item slate",
    "formal_comparison": true,
    "claim_policy": "formal multi-seed comparison"
  },
  "seed_results": [
    {
      "paper": {
        "arxiv_id": "2607.26427",
        "title": "PSG: Pair-Space Generation for Efficient Generative Reranking",
        "url": "https://arxiv.org/abs/2607.26427",
        "organization": "Kuaishou Technology"
      },
      "dataset": {
        "name": "MovieLens 100K",
        "users": 220,
        "items": 360
      },
      "setup": {
        "adapter": "psg",
        "same_split_and_candidates": true
      },
      "baseline": {
        "name": "item-space autoregressive next-item scoring",
        "hit_at_10": 0.10909090909090909,
        "ndcg_at_10": 0.05400546778896325,
        "fresh_hit_at_10": 0.031746031746031744,
        "head_share_at_10": 0.23772727272727273
      },
      "method": {
        "name": "ordered pair-token generation and unfolding",
        "hit_at_10": 0.09090909090909091,
        "ndcg_at_10": 0.04826018605777078,
        "fresh_hit_at_10": 0.07936507936507936,
        "head_share_at_10": 0.18954545454545454
      },
      "relative": {
        "hit_at_10_percent": -16.66666666666666,
        "ndcg_at_10_percent": -10.638333425132556,
        "fresh_hit_at_10_percent": 150.0,
        "head_share_at_10_percent": -20.26768642447419
      },
      "stages": {
        "item_space_steps": 6,
        "pair_space_steps": 3,
        "decode_step_reduction": 0.5,
        "request_pair_vocabulary": 129240,
        "generated_items": 6,
        "duplicate_items": 0
      },
      "paper_results": {
        "stay_time_lift_percent": 0.178,
        "industrial_speedup": 1.83,
        "traffic_percent_per_bucket": 10,
        "experiment_days": 7
      },
      "scope": "实际构造 request-specific n(n-1) 有序 pair space、pair token 打分、半长度解码与无重复展开；未复刻 Kuaishou GoalRank evaluator、曝光日志和线上 50-list beam serving。"
    },
    {
      "paper": {
        "arxiv_id": "2607.26427",
        "title": "PSG: Pair-Space Generation for Efficient Generative Reranking",
        "url": "https://arxiv.org/abs/2607.26427",
        "organization": "Kuaishou Technology"
      },
      "dataset": {
        "name": "MovieLens 100K",
        "users": 220,
        "items": 360
      },
      "setup": {
        "adapter": "psg",
        "same_split_and_candidates": true
      },
      "baseline": {
        "name": "item-space autoregressive next-item scoring",
        "hit_at_10": 0.10909090909090909,
        "ndcg_at_10": 0.05400546778896325,
        "fresh_hit_at_10": 0.031746031746031744,
        "head_share_at_10": 0.23772727272727273
      },
      "method": {
        "name": "ordered pair-token generation and unfolding",
        "hit_at_10": 0.09090909090909091,
        "ndcg_at_10": 0.04826018605777078,
        "fresh_hit_at_10": 0.07936507936507936,
        "head_share_at_10": 0.18954545454545454
      },
      "relative": {
        "hit_at_10_percent": -16.66666666666666,
        "ndcg_at_10_percent": -10.638333425132556,
        "fresh_hit_at_10_percent": 150.0,
        "head_share_at_10_percent": -20.26768642447419
      },
      "stages": {
        "item_space_steps": 6,
        "pair_space_steps": 3,
        "decode_step_reduction": 0.5,
        "request_pair_vocabulary": 129240,
        "generated_items": 6,
        "duplicate_items": 0
      },
      "paper_results": {
        "stay_time_lift_percent": 0.178,
        "industrial_speedup": 1.83,
        "traffic_percent_per_bucket": 10,
        "experiment_days": 7
      },
      "scope": "实际构造 request-specific n(n-1) 有序 pair space、pair token 打分、半长度解码与无重复展开；未复刻 Kuaishou GoalRank evaluator、曝光日志和线上 50-list beam serving。"
    },
    {
      "paper": {
        "arxiv_id": "2607.26427",
        "title": "PSG: Pair-Space Generation for Efficient Generative Reranking",
        "url": "https://arxiv.org/abs/2607.26427",
        "organization": "Kuaishou Technology"
      },
      "dataset": {
        "name": "MovieLens 100K",
        "users": 220,
        "items": 360
      },
      "setup": {
        "adapter": "psg",
        "same_split_and_candidates": true
      },
      "baseline": {
        "name": "item-space autoregressive next-item scoring",
        "hit_at_10": 0.10909090909090909,
        "ndcg_at_10": 0.05400546778896325,
        "fresh_hit_at_10": 0.031746031746031744,
        "head_share_at_10": 0.23772727272727273
      },
      "method": {
        "name": "ordered pair-token generation and unfolding",
        "hit_at_10": 0.09090909090909091,
        "ndcg_at_10": 0.04826018605777078,
        "fresh_hit_at_10": 0.07936507936507936,
        "head_share_at_10": 0.18954545454545454
      },
      "relative": {
        "hit_at_10_percent": -16.66666666666666,
        "ndcg_at_10_percent": -10.638333425132556,
        "fresh_hit_at_10_percent": 150.0,
        "head_share_at_10_percent": -20.26768642447419
      },
      "stages": {
        "item_space_steps": 6,
        "pair_space_steps": 3,
        "decode_step_reduction": 0.5,
        "request_pair_vocabulary": 129240,
        "generated_items": 6,
        "duplicate_items": 0
      },
      "paper_results": {
        "stay_time_lift_percent": 0.178,
        "industrial_speedup": 1.83,
        "traffic_percent_per_bucket": 10,
        "experiment_days": 7
      },
      "scope": "实际构造 request-specific n(n-1) 有序 pair space、pair token 打分、半长度解码与无重复展开；未复刻 Kuaishou GoalRank evaluator、曝光日志和线上 50-list beam serving。"
    }
  ],
  "aggregate_metrics": {
    "dataset.users": {
      "mean": 220.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "dataset.items": {
      "mean": 360.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "baseline.hit_at_10": {
      "mean": 0.10909090909090909,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "baseline.ndcg_at_10": {
      "mean": 0.05400546778896325,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "baseline.fresh_hit_at_10": {
      "mean": 0.031746031746031744,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "baseline.head_share_at_10": {
      "mean": 0.2377272727272727,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "method.hit_at_10": {
      "mean": 0.0909090909090909,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "method.ndcg_at_10": {
      "mean": 0.04826018605777078,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "method.fresh_hit_at_10": {
      "mean": 0.07936507936507936,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "method.head_share_at_10": {
      "mean": 0.18954545454545454,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "relative.hit_at_10_percent": {
      "mean": -16.66666666666666,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "relative.ndcg_at_10_percent": {
      "mean": -10.638333425132556,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "relative.fresh_hit_at_10_percent": {
      "mean": 150.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "relative.head_share_at_10_percent": {
      "mean": -20.26768642447419,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "stages.item_space_steps": {
      "mean": 6.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "stages.pair_space_steps": {
      "mean": 3.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "stages.decode_step_reduction": {
      "mean": 0.5,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "stages.request_pair_vocabulary": {
      "mean": 129240.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "stages.generated_items": {
      "mean": 6.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "stages.duplicate_items": {
      "mean": 0.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "paper_results.stay_time_lift_percent": {
      "mean": 0.17800000000000002,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "paper_results.industrial_speedup": {
      "mean": 1.83,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "paper_results.traffic_percent_per_bucket": {
      "mean": 10.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    },
    "paper_results.experiment_days": {
      "mean": 7.0,
      "std": 0.0,
      "ci95": 0.0,
      "n": 3
    }
  },
  "manifest_ref": "reproduction:psg"
}