{ "paper": { "arxiv_id": "2604.12965", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "url": "https://arxiv.org/abs/2604.12965", "organization": "Meta" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "hill-index", "same_split_and_candidates": true }, "baseline": { "name": "one-layer learned index", "hit_at_10": 0.10909090909090909, "ndcg_at_10": 0.05518934137305569, "fresh_hit_at_10": 0.07936507936507936, "head_share_at_10": 0.10363636363636364 }, "method": { "name": "HILL residual two-layer index", "hit_at_10": 0.05454545454545454, "ndcg_at_10": 0.035973142239490156, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.10681818181818181 }, "relative": { "hit_at_10_percent": -50.0, "ndcg_at_10_percent": -34.818678127851676, "fresh_hit_at_10_percent": -60.0, "head_share_at_10_percent": 3.0701754385964843 }, "stages": { "coarse_nodes": 6, "residual_children_per_node": 6, "flat_mean_scored_items": 61.328125, "hill_mean_scored_items": 83.875, "hill_candidate_fraction": 0.23298611111111112, "test_time_training_nodes": 28 }, "paper_results": { "two_layer_monn_medium_small_online_metric_percent": 2.57, "two_layer_monn_small_small_online_metric_percent": 1.22 }, "scope": "实际学习 coarse attention assignment、跨层 residual quantization、层级 beam 检索，并与同公开切分的单层索引比较候选量和质量；未复刻 Meta MoNN、十亿级 Ads 索引和分布式 FAISS EM。", "schema_version": 2, "manifest": { "adapter_key": "hill-index", "arxiv_id": "2604.12965", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "paper_url": "https://arxiv.org/abs/2604.12965", "track": "recommendation", "organization": "Meta / Facebook and Instagram Ads", "published": "2026-04-14", "code_url": null, "topics": [ "retrieval", "hierarchical-index", "residual-quantization", "test-time-training", "serving" ], "local_code_dir": "src/auto_research/reproductions/industrial_2026", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K (220 users / 360 items)" ], "baseline": "one-layer learned index with the same public item features and retrieval protocol", "metrics": [ "hit_at_10", "ndcg_at_10", "head_share_at_10", "mean_scored_items" ], "default_seeds": [ 42, 43, 44 ], "budget": "6 coarse nodes; up to 6 residual child nodes per parent", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "Meta Ads Retrieval", "metric": "online ads metric", "lift_percent": 2.57, "traffic": "production online A/B: 2-layer MoNN Large/Small versus MoNN Small", "source_url": "https://arxiv.org/html/2604.12965v1#S4.SS5", "source_location": "Section 4.5, Table 8", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [] }, "provenance": { "created_at": "2026-09-01T01:50:54.010622+00:00", "code_commit": "df64a338410bf9bdc3e6e2feb9d2d1f5000bf221", "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/2604.12965-hill-index/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": "6 coarse nodes; up to 6 residual child nodes per parent", "formal_comparison": true, "claim_policy": "formal multi-seed comparison" }, "seed_results": [ { "paper": { "arxiv_id": "2604.12965", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "url": "https://arxiv.org/abs/2604.12965", "organization": "Meta" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "hill-index", "same_split_and_candidates": true }, "baseline": { "name": "one-layer learned index", "hit_at_10": 0.10909090909090909, "ndcg_at_10": 0.05518934137305569, "fresh_hit_at_10": 0.07936507936507936, "head_share_at_10": 0.10363636363636364 }, "method": { "name": "HILL residual two-layer index", "hit_at_10": 0.05454545454545454, "ndcg_at_10": 0.035973142239490156, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.10681818181818181 }, "relative": { "hit_at_10_percent": -50.0, "ndcg_at_10_percent": -34.818678127851676, "fresh_hit_at_10_percent": -60.0, "head_share_at_10_percent": 3.0701754385964843 }, "stages": { "coarse_nodes": 6, "residual_children_per_node": 6, "flat_mean_scored_items": 61.328125, "hill_mean_scored_items": 83.875, "hill_candidate_fraction": 0.23298611111111112, "test_time_training_nodes": 28 }, "paper_results": { "two_layer_monn_medium_small_online_metric_percent": 2.57, "two_layer_monn_small_small_online_metric_percent": 1.22 }, "scope": "实际学习 coarse attention assignment、跨层 residual quantization、层级 beam 检索，并与同公开切分的单层索引比较候选量和质量；未复刻 Meta MoNN、十亿级 Ads 索引和分布式 FAISS EM。", "seed": 42 }, { "paper": { "arxiv_id": "2604.12965", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "url": "https://arxiv.org/abs/2604.12965", "organization": "Meta" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "hill-index", "same_split_and_candidates": true }, "baseline": { "name": "one-layer learned index", "hit_at_10": 0.09545454545454546, "ndcg_at_10": 0.05635913257742555, "fresh_hit_at_10": 0.06349206349206349, "head_share_at_10": 0.11227272727272727 }, "method": { "name": "HILL residual two-layer index", "hit_at_10": 0.07727272727272727, "ndcg_at_10": 0.045418828808666205, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.09636363636363636 }, "relative": { "hit_at_10_percent": -19.047619047619055, "ndcg_at_10_percent": -19.411767478375715, "fresh_hit_at_10_percent": -50.0, "head_share_at_10_percent": -14.170040485829965 }, "stages": { "coarse_nodes": 6, "residual_children_per_node": 6, "flat_mean_scored_items": 69.875, "hill_mean_scored_items": 52.1875, "hill_candidate_fraction": 0.1449652777777778, "test_time_training_nodes": 34 }, "paper_results": { "two_layer_monn_medium_small_online_metric_percent": 2.57, "two_layer_monn_small_small_online_metric_percent": 1.22 }, "scope": "实际学习 coarse attention assignment、跨层 residual quantization、层级 beam 检索，并与同公开切分的单层索引比较候选量和质量；未复刻 Meta MoNN、十亿级 Ads 索引和分布式 FAISS EM。", "seed": 43 }, { "paper": { "arxiv_id": "2604.12965", "title": "Efficient Retrieval Scaling with Hierarchical Indexing for Large Scale Recommendation", "url": "https://arxiv.org/abs/2604.12965", "organization": "Meta" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "hill-index", "same_split_and_candidates": true }, "baseline": { "name": "one-layer learned index", "hit_at_10": 0.09545454545454546, "ndcg_at_10": 0.04988833268900992, "fresh_hit_at_10": 0.047619047619047616, "head_share_at_10": 0.10454545454545454 }, "method": { "name": "HILL residual two-layer index", "hit_at_10": 0.05, "ndcg_at_10": 0.03448482384798187, "fresh_hit_at_10": 0.015873015873015872, "head_share_at_10": 0.10954545454545454 }, "relative": { "hit_at_10_percent": -47.61904761904762, "ndcg_at_10_percent": -30.875974422815986, "fresh_hit_at_10_percent": -66.66666666666667, "head_share_at_10_percent": 4.782608695652178 }, "stages": { "coarse_nodes": 6, "residual_children_per_node": 6, "flat_mean_scored_items": 60.046875, "hill_mean_scored_items": 78.15625, "hill_candidate_fraction": 0.21710069444444444, "test_time_training_nodes": 27 }, "paper_results": { "two_layer_monn_medium_small_online_metric_percent": 2.57, "two_layer_monn_small_small_online_metric_percent": 1.22 }, "scope": "实际学习 coarse attention assignment、跨层 residual quantization、层级 beam 检索，并与同公开切分的单层索引比较候选量和质量；未复刻 Meta MoNN、十亿级 Ads 索引和分布式 FAISS EM。", "seed": 44 } ], "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": 0.09999999999999999, "std": 0.007872958216222165, "ci95": 0.008909090909090905, "n": 3 }, "baseline.ndcg_at_10": { "mean": 0.05381226887983039, "std": 0.003448196526835189, "ci95": 0.003902001698254577, "n": 3 }, "baseline.fresh_hit_at_10": { "mean": 0.06349206349206349, "std": 0.015873015873015872, "ci95": 0.017962008374788358, "n": 3 }, "baseline.head_share_at_10": { "mean": 0.10681818181818181, "std": 0.0047455938676866155, "ci95": 0.0053701449980680695, "n": 3 }, "method.hit_at_10": { "mean": 0.06060606060606061, "std": 0.014611592062110537, "ci95": 0.016534572956283095, "n": 3 }, "method.ndcg_at_10": { "mean": 0.03862559829871274, "std": 0.0059299880674118875, "ci95": 0.006710413205742584, "n": 3 }, "method.fresh_hit_at_10": { "mean": 0.026455026455026454, "std": 0.009164289987136916, "ci95": 0.01037037037037037, "n": 3 }, "method.head_share_at_10": { "mean": 0.10424242424242425, "std": 0.006958159092851536, "ci95": 0.00787389824963127, "n": 3 }, "relative.hit_at_10_percent": { "mean": -38.88888888888889, "std": 17.224233662071985, "ci95": 19.49105524510926, "n": 3 }, "relative.ndcg_at_10_percent": { "mean": -28.36880667634779, "std": 8.003601784317592, "ci95": 9.056928023538168, "n": 3 }, "relative.fresh_hit_at_10_percent": { "mean": -58.88888888888889, "std": 8.388704928078614, "ci95": 9.492713254821915, "n": 3 }, "relative.head_share_at_10_percent": { "mean": -2.1057521171937674, "std": 10.483005165322375, "ci95": 11.862637074065617, "n": 3 }, "stages.coarse_nodes": { "mean": 6.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.residual_children_per_node": { "mean": 6.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.flat_mean_scored_items": { "mean": 63.75, "std": 5.342950415325319, "ci95": 6.046117566687596, "n": 3 }, "stages.hill_mean_scored_items": { "mean": 71.40625, "std": 16.887755595771157, "ci95": 19.11029446888509, "n": 3 }, "stages.hill_candidate_fraction": { "mean": 0.19835069444444445, "std": 0.046910432210475424, "ci95": 0.05308415130245857, "n": 3 }, "stages.test_time_training_nodes": { "mean": 29.666666666666668, "std": 3.7859388972001824, "ci95": 4.28419316921064, "n": 3 }, "paper_results.two_layer_monn_medium_small_online_metric_percent": { "mean": 2.57, "std": 0.0, "ci95": 0.0, "n": 3 }, "paper_results.two_layer_monn_small_small_online_metric_percent": { "mean": 1.22, "std": 0.0, "ci95": 0.0, "n": 3 } }, "manifest_ref": "reproduction:hill-index" }