{ "paper": { "arxiv_id": "2607.28940", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "url": "https://arxiv.org/abs/2607.28940", "organization": "LinkedIn" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "transx", "same_split_and_candidates": true }, "baseline": { "name": "monolithic recent-sequence scorer", "hit_at_10": 0.10909090909090909, "ndcg_at_10": 0.05400546778896325, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.23772727272727273 }, "method": { "name": "cached behavior encoder + serving-stream cross attention", "hit_at_10": 0.07727272727272727, "ndcg_at_10": 0.04475488754459977, "fresh_hit_at_10": 0.1111111111111111, "head_share_at_10": 0.065 }, "relative": { "hit_at_10_percent": -29.166666666666668, "ndcg_at_10_percent": -17.128969756380787, "fresh_hit_at_10_percent": 250.0, "head_share_at_10_percent": -72.65774378585087 }, "stages": { "behavior_events": 73, "serving_events": 4, "cached_behavior_tokens": 9, "monolithic_attention_pairs": 5929, "cross_stream_attention_pairs": 36, "attention_pair_reduction": 0.9939281497723056 }, "paper_results": { "ctr_lift_percent": 6.0, "conversion_lift_percent": 4.4, "online_compute_reduction_percent": 80 }, "scope": "实际拆分 nearline 行为流和实时 serving 流，构造 global-local cache 并执行 candidate-to-cache cross attention；未复刻 LinkedIn Kafka、KV 服务、TTL/version 管理和生产候选特征。", "schema_version": 2, "manifest": { "adapter_key": "transx", "arxiv_id": "2607.28940", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "paper_url": "https://arxiv.org/abs/2607.28940", "track": "recommendation", "organization": "LinkedIn", "published": "2026-07-31", "code_url": null, "topics": [ "ranking", "long-sequence", "cross-attention", "serving" ], "local_code_dir": "src/auto_research/reproductions/industrial_2026", "fidelity": "core_mechanism", "evaluation_tier": "l2_public_dataset", "datasets": [ "MovieLens 100K" ], "baseline": "monolithic recent-sequence scorer", "metrics": [ "hit_at_10", "ndcg_at_10", "attention_pair_reduction" ], "default_seeds": [ 42, 43, 44 ], "budget": "220 users / 360 items; global token plus 8 cached local tokens", "device_capabilities": [ "cpu", "mps", "cuda" ], "requires_gpu_validation": false, "gpu_validation_artifact": null, "online_evidence": [ { "product": "LinkedIn recommender systems", "metric": "CTR", "lift_percent": 6.0, "traffic": "large-scale online A/B test", "source_url": "https://arxiv.org/html/2607.28940v1", "source_location": "Abstract and Section 6", "significance": "statistically significant for users with eligible history", "retrieved_at": "2026-09-01" } ], "selection_exception": null, "evolve_operators": [] }, "provenance": { "created_at": "2026-09-01T10:55:35.958213+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/2607.28940-transx/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; global token plus 8 cached local tokens", "formal_comparison": true, "claim_policy": "formal multi-seed comparison" }, "seed_results": [ { "paper": { "arxiv_id": "2607.28940", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "url": "https://arxiv.org/abs/2607.28940", "organization": "LinkedIn" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "transx", "same_split_and_candidates": true }, "baseline": { "name": "monolithic recent-sequence scorer", "hit_at_10": 0.10909090909090909, "ndcg_at_10": 0.05400546778896325, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.23772727272727273 }, "method": { "name": "cached behavior encoder + serving-stream cross attention", "hit_at_10": 0.07727272727272727, "ndcg_at_10": 0.04475488754459977, "fresh_hit_at_10": 0.1111111111111111, "head_share_at_10": 0.065 }, "relative": { "hit_at_10_percent": -29.166666666666668, "ndcg_at_10_percent": -17.128969756380787, "fresh_hit_at_10_percent": 250.0, "head_share_at_10_percent": -72.65774378585087 }, "stages": { "behavior_events": 73, "serving_events": 4, "cached_behavior_tokens": 9, "monolithic_attention_pairs": 5929, "cross_stream_attention_pairs": 36, "attention_pair_reduction": 0.9939281497723056 }, "paper_results": { "ctr_lift_percent": 6.0, "conversion_lift_percent": 4.4, "online_compute_reduction_percent": 80 }, "scope": "实际拆分 nearline 行为流和实时 serving 流，构造 global-local cache 并执行 candidate-to-cache cross attention；未复刻 LinkedIn Kafka、KV 服务、TTL/version 管理和生产候选特征。" }, { "paper": { "arxiv_id": "2607.28940", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "url": "https://arxiv.org/abs/2607.28940", "organization": "LinkedIn" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "transx", "same_split_and_candidates": true }, "baseline": { "name": "monolithic recent-sequence scorer", "hit_at_10": 0.10909090909090909, "ndcg_at_10": 0.05400546778896325, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.23772727272727273 }, "method": { "name": "cached behavior encoder + serving-stream cross attention", "hit_at_10": 0.07727272727272727, "ndcg_at_10": 0.04475488754459977, "fresh_hit_at_10": 0.1111111111111111, "head_share_at_10": 0.065 }, "relative": { "hit_at_10_percent": -29.166666666666668, "ndcg_at_10_percent": -17.128969756380787, "fresh_hit_at_10_percent": 250.0, "head_share_at_10_percent": -72.65774378585087 }, "stages": { "behavior_events": 73, "serving_events": 4, "cached_behavior_tokens": 9, "monolithic_attention_pairs": 5929, "cross_stream_attention_pairs": 36, "attention_pair_reduction": 0.9939281497723056 }, "paper_results": { "ctr_lift_percent": 6.0, "conversion_lift_percent": 4.4, "online_compute_reduction_percent": 80 }, "scope": "实际拆分 nearline 行为流和实时 serving 流，构造 global-local cache 并执行 candidate-to-cache cross attention；未复刻 LinkedIn Kafka、KV 服务、TTL/version 管理和生产候选特征。" }, { "paper": { "arxiv_id": "2607.28940", "title": "TransX: Scaling Transformer-based Recommendation via Behavioral and Serving Stream Crossings", "url": "https://arxiv.org/abs/2607.28940", "organization": "LinkedIn" }, "dataset": { "name": "MovieLens 100K", "users": 220, "items": 360 }, "setup": { "adapter": "transx", "same_split_and_candidates": true }, "baseline": { "name": "monolithic recent-sequence scorer", "hit_at_10": 0.10909090909090909, "ndcg_at_10": 0.05400546778896325, "fresh_hit_at_10": 0.031746031746031744, "head_share_at_10": 0.23772727272727273 }, "method": { "name": "cached behavior encoder + serving-stream cross attention", "hit_at_10": 0.07727272727272727, "ndcg_at_10": 0.04475488754459977, "fresh_hit_at_10": 0.1111111111111111, "head_share_at_10": 0.065 }, "relative": { "hit_at_10_percent": -29.166666666666668, "ndcg_at_10_percent": -17.128969756380787, "fresh_hit_at_10_percent": 250.0, "head_share_at_10_percent": -72.65774378585087 }, "stages": { "behavior_events": 73, "serving_events": 4, "cached_behavior_tokens": 9, "monolithic_attention_pairs": 5929, "cross_stream_attention_pairs": 36, "attention_pair_reduction": 0.9939281497723056 }, "paper_results": { "ctr_lift_percent": 6.0, "conversion_lift_percent": 4.4, "online_compute_reduction_percent": 80 }, "scope": "实际拆分 nearline 行为流和实时 serving 流，构造 global-local cache 并执行 candidate-to-cache cross attention；未复刻 LinkedIn Kafka、KV 服务、TTL/version 管理和生产候选特征。" } ], "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": 0.07727272727272727, "std": 0.0, "ci95": 0.0, "n": 3 }, "method.ndcg_at_10": { "mean": 0.04475488754459977, "std": 0.0, "ci95": 0.0, "n": 3 }, "method.fresh_hit_at_10": { "mean": 0.1111111111111111, "std": 0.0, "ci95": 0.0, "n": 3 }, "method.head_share_at_10": { "mean": 0.065, "std": 0.0, "ci95": 0.0, "n": 3 }, "relative.hit_at_10_percent": { "mean": -29.166666666666668, "std": 0.0, "ci95": 0.0, "n": 3 }, "relative.ndcg_at_10_percent": { "mean": -17.128969756380787, "std": 0.0, "ci95": 0.0, "n": 3 }, "relative.fresh_hit_at_10_percent": { "mean": 250.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "relative.head_share_at_10_percent": { "mean": -72.65774378585087, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.behavior_events": { "mean": 73.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.serving_events": { "mean": 4.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.cached_behavior_tokens": { "mean": 9.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.monolithic_attention_pairs": { "mean": 5929.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.cross_stream_attention_pairs": { "mean": 36.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "stages.attention_pair_reduction": { "mean": 0.9939281497723056, "std": 0.0, "ci95": 0.0, "n": 3 }, "paper_results.ctr_lift_percent": { "mean": 6.0, "std": 0.0, "ci95": 0.0, "n": 3 }, "paper_results.conversion_lift_percent": { "mean": 4.4, "std": 0.0, "ci95": 0.0, "n": 3 }, "paper_results.online_compute_reduction_percent": { "mean": 80.0, "std": 0.0, "ci95": 0.0, "n": 3 } }, "manifest_ref": "reproduction:transx" }