跳转至

Reproduction and implementation overview

简体中文

This is the underlying unified inventory of reproduction adapters, implementation status, and experiment conclusions. Updates to papers and reruns go here and in the individual paper README rather than new milestone audit or batch summary pages.

Browse industrial search, ads, and recommendation by company, by topic, or by month. Architecture, pretraining, multimodality, and inference efficiency are in foundation models; post-training algorithms are in LLM post-training. Paper detail and physical paths stay unchanged.

Each new batch builds a candidate set first and records an explicit terminal disposition per candidate. See the discovery and coverage audit for rules, the machine ledger, and mandatory checks.

Key original figures

Each write-up shows a key original figure between background/main changes and core formulas. Prefer architecture, model structure, or training/inference flow. Store the image in the paper directory as assets/paper-figure-01.png, link the original, and retain its copyright note. Without an architecture figure, choose the original flow, algorithm, or experimental screenshot that best explains the method. Explicitly state when public full text is inaccessible; local Mermaid drawings are not original screenshots.

Add or refresh paper figures:

python -m pip install -e '.[paper-figures]'
python scripts/sync_paper_figures.py --only <arXiv ID or paper directory keyword>

Check the collection:

python scripts/sync_paper_figures.py --workers 3
pytest tests/test_research_module_docs.py

Extraction uses arXiv HTML caption relations first, falling back to PDF caption localization and cropping. Some non-arXiv papers use manually verified public PDF mappings. Stable comment markers let repeated refreshes replace figures without duplicate sections.

Admission and reporting rules

  • Industrial threshold: Full text must report quantitative production online A/B, or the user must explicitly accept statistically significant full rollout, business gains, and guardrails. Deployment alone, offline SOTA, and simulator results do not count. Label unquantified full traffic without inventing percentages.
  • Foundation models: No online A/B required; need a public benchmark, matched-budget controls, and a core method trainable on public data such as WikiText. Formulas or fixed scores alone do not enter the table.
  • Named exceptions: SASRec, TIGER, Wide & Deep, DCN-V2, DIEN, BST, DeepFM, YouTube DNN, ESMM, MMoE, and PLE fill classic lineages at user request. Missing quantitative A/B is disclosed per item and does not lower the bar later.
  • Local results: Every README states baselines, experimental groups, data, main metrics, and relative changes. Separate paper online results, local cross-model comparisons, component ablations, and efficiency controls.
  • Fidelity: Public substitutes and scaling down are allowed. Replacing a defining architecture, training objective, or inference path with a heuristic is a concept demo. Default batch runs exclude concept demos.
  • Paper information: At the top of each README list the paper, organization, exact first release and source, upstream code status, adapter, and local code path. Explicitly say no when original author code is not found. Use official proceedings or institution pages if a conference paper lacks arXiv; do not invent IDs.

The common DIN experiment uses MovieLens-100K, chronological leave-two-out, full-item ranking, and seeds 42/43/44. SERAL, LEADRE, COBRA, ARGUS, GR4AD, and MM-LLM use DIN NDCG@10 0.02167. Cross-domain KD uses DIN 0.05518 on its independent target split. These are public small-data results, not the papers' private industrial evidence.

Current progress

  • oob-regularization · OOB Regularization: Full-dataset Gaussian quantiles and spherical anchors, CPU caches, and periodic refresh. A four-arm short run on public CIFAR used three A100 seeds. Negative results are retained; the full multi-dataset study is not reproduced.
  • forte · FORTE: Pinterest's paired feature-retirement funnel. All three real A100 MovieLens runs retained the original features; no pruning gains are claimed.
  • evo2 · Evo²: Fixed executor alternates harness edits with Proposer/Reflector training, retaining actual execution feedback, SFT, and role-specific GRPO. Local concept diagnostic.
  • rsi-anything · RSI-Anything: Recursive improvement and policy training driven by isolated program execution feedback. No gold candidates are injected; not evidence of general capability.
  • branch-opd · BranchOPD: Selects branch points on real student rollouts, samples continuations independently, scores them with a teacher, and applies branch distillation and answer formatting constraints.
  • cast-opd · CAST: Aligns credit across tokenizers to pass trainable token-level supervision between heterogeneous teacher and student. Its key is distinct from the older agent CAST.
  • road-mopd · Road-MOPD: Adds cross-teacher safeguards to domain routing, recording conflicts and protected student updates.
  • boundary-evo · BoundaryEvo: Collects teacher trajectories near the student's capability frontier and maintains replay, iterating data and model together.
  • msar · MSAR: Expands a single residual stream into multi-stream depth-attention residuals and compares real language-model training and resource overhead.
  • orthros · Orthros: Separates a sliding-window backbone from an independent full-context output stream and checks prefill/decode KV-cache equivalence.
  • dualquant · DualQuant: Quantizes real weights and activations with calibration-free dual scaling. Software INT4 does not imply native A100 low-bit acceleration.
  • janus · JANUS: Predicts relative program gains and uncertainty, then tests actual candidates, retaining errors and zero gains.
  • t2mem · T2Mem: Manages long conversations using staged summaries, time information, and retrieval. Only the independent scorer sees answers.
  • rated · RATED: Trains a real query/tool dual encoder and updates tool selection from feedback. Semantic retrieval diagnosis is not backend task success.
  • jev-workflow · Jev workflow evaluation: Compares workflow efficiency, quality, and cost with paired traces. An evaluation facility, not a training algorithm.
  • woven · WOVEN: Uses visual state transitions as a shared training primitive. Official validation and test data were unavailable; local experiments use a labeled split of the training set.

  • umrm · UMRM: Trains a reward head from frozen VLM intermediate layers. Object verification with public captions does not reproduce official CHAIR.

  • rptune · RPTune: Google's joint query–product scoring, pruning, and suffix scheduling. No quantitative online A/B; academic/evolve diagnostic only, not industrial evidence.

  • agent-web-rec · AgentWebRec: Combines platform semantics, private user sequence memory, and low-confidence neighbor patterns without exposing private records. No online A/B; academic/evolve diagnostic.
  • gear · GEAR: Orthogonal-basis BasisVQ/BasisRQ and contextual reranking of collided items. The paper reports seven days of Douyin Ads A/B at 5% per arm; local work covers the L1 mechanism only.
  • effective-training-time · Effective Training Time: Separates training on new data from initialization, compilation, checkpointing, publishing, and recovery. User-approved production infrastructure exception, not recommender A/B evidence.
  • gris · GrIS: Combines collaborative graphs and content semantics in hierarchical Semantic IDs. User-approved academic/evolve exception with no online A/B.
  • repair-state · REPAIR: Selects corrective evidence from frozen encoder caches to repair a lossy preference state via residuals. User-approved academic/evolve exception with no online A/B.
  • cohortmix-ts · CohortMix-TS: Builds fixed-strength mixed Beta priors from historical cohorts, selects slates by Thompson sampling, and updates from feedback. The 25-day randomized deployment's significant result is from a restricted full-window subgroup; local work diagnoses state updates.
  • recap-ctr · RECAP CTR: Multiple estimator routes in a weight-shared recurrent backbone, route averaging, and EMA over training. Executable evolve operator: rankmixer_recap. No online A/B; not industrial evidence.
  • evoskillrec · EvoSkillRec: Runs typed skill genomes, validation, promotion, and reuse via rankmixer_evoskill in unified evolve. L1 mechanism diagnostic, not a claim about large search gains.
  • promptshift · PromptShift: Computes Drift, SliceShift, difficulty-weighted hits, and adaptive inverse-popularity reranking on fixed relevance and identity-slice fixtures. No real identity profiles are generated or inferred.
  • grp · Snap GRP v0.1: Separate SID blocks, isolated generation/ranking gradients, and mGRPO with one-sided recall safeguards phase a shared model into retrieval and ranking. L1 public tensors, not Snap's online results.
  • helix · TikTok HELIX: Interleaves feature interactions and sequence modeling while preserving user-state reuse with one-way information flow. L1 public-tensor mechanism, not full-rollout A/B replication.
  • flvm · Google/YouTube FLVM: Three latent factors, a restricted confound baseline, and sparse feedback routing with three positive/negative-feedback seeds on public KuaiRand. Sparse negatives and unstable semantics make this a concept diagnostic.
  • recursive-opsd · Meta DCE+SRCL: Qwen3-4B on A100 runs a dynamic privileged teacher, token forward KL, and answer-validated concise self-rewrites. Two rounds on public GSM8K gave no accuracy gain evidence; concept diagnostic.
  • kuafu · Tencent KuaFu: A100 ran itemwise compression, dual-axis projection, and small public MRQA supervision. Dev F1 did not beat same-token head truncation; hallucination RL had zero within-group advantage. Concept diagnostic.
  • onetrans-v2 · ByteDance OneTrans-V2: Shared causal user backbone, three-level SID retrieval, and joint pre/fine ranking plus distillation ran with three MovieLens-1M proxy-label seeds. No real conversions, ads, or impression funnel; concept diagnostic.
  • xrec · TikTok X-Rec: Anchor-conditioned spherical flow matching, final-layer interactions, and multi-trigger full-catalog retrieval on MovieLens-1M. Three small-budget seeds underperformed same-budget U2I.
  • music-rationales · Google YouTube Music: A real 7B model creates discovery profiles offline; a public artist catalog filters cached reasons, with CF fallback. Concept diagnostic only.
  • light-heads · Google Light Heads: Dynamically adds gradient-free shallow heads above shared ranking towers, comparing window reset and removing stop-gradient.
  • unique · UNIQUE: Flat single-level quantization and early fusion for joint retrieval and ranking. Retains low recall and no ranking improvement on KuaiRand-Pure.
  • muser · MuSeR: Hierarchical temporal compression and multi-query interests. Public labels replace private multimodal semantics; concept diagnostic.
  • evopilot · EvoPilot: Constrains long-running online research with artifact attestation, a fail-closed pair verifier, and human review.
  • angle · ANGLE: Hierarchical text identifiers, joint generation/classification/ranking, and dynamic constrained retrieval.
  • sas-attention · SAS: Uses continuous log gates in softmax to train sparse context ranking end to end with language-model loss.
  • mima · MIMA: Uses same-request multiple positives, exclusive matching, and calibrated interest activation to reduce interest collapse.
  • chronicle-rec · ChronicleRec: Recent-dense and older-sparse compression, causal time anchors, multiple horizons, and alignment pretraining.
  • pindco · PinDCO: Incrementally fuses component towers with a page-pixel penalty, prescreening, and feedback from exploration.
  • gese · GESE: Set-level title exploration, fidelity constraints, and personalized contextual selection.
  • lazformer · LazFormer: Transferable generative pretraining, ranking residual adaptation, and recent-dense/older-sparse attention.
  • sirf · SIRF: Internalizes policy rules as trainable interaction features and validates decisions with separate high-quantile error constraints.
  • unirec · UniRec: The paper fuses pre-ranking and fine ranking. The old NumPy scorer is a concept demo; v3 CPPA/CAPP and AGRR KL are separate A100 mechanism components, not full reproduction.
  • sequenceo1 · SequenceO1: Fixed-budget Sketch Attention, a recent-history branch, and a cache for extremely long sequences.
  • baff · BAFF: Two-axis ranking/bidding interference filtering on shared RTB logs.
  • allecompanion · AlleCompanion: Category-constrained dual towers, ComCat complementary graph, and controlled Category Adapter.
  • autolr · AutoLR: Multi-expert review, evidence-weighted experiment budgets, and deterministic release gates.
  • beaconkv · BeaconKV: Query-cluster beacons predict distant KV revisits in long reasoning.
  • kvmem · KVMEM: Paged KV virtualization and query-dependent agent workspace materialization.
  • coral · CORAL: Meta's constrained configuration, short-term experiment memory, and production feedback loop.
  • recevolve · RecEvolve: Google's knowledge-driven proposals, isolated experiments, rollback, and anti-reward-hacking controller.
  • random-attention · Random Attention: Prompt protection, random per-head KV eviction, and real Qwen/A100 verification.
  • lngram-v2 · Lngram v2: Discrete latent n-gram memory, GQA zero sink, and real VLM/A100 verification.
  • rest · ReST: Dual-gated temporal encoding and reusable user prefixes for industrial ranking.
  • tgr · TGR: Unified generative ranking, semantic IDs, and offline reason tokens.
  • camie · CAMIE: Aligns multimodal product representations using shared engagement.

  • setmir · SetMIR: Predicts sets of interests, gates presence, and dispatches dynamically to ANN.

  • Audited 94 main entries and 138 arXiv links in two industrial-deployment blog sections.

  • See this page and the domain indexes for adapters. Recommendation requires online A/B or full rollout; foundation models need public benchmarks and real training. Generated manifest counts are authoritative.
  • Deferred: AIGQ lacks equivalent query/CTR rewards; RaG needs video generation and quality feedback; RoleGen needs conversion trajectories and online feedback; LCU data require an NDA.
  • Skipped EGA-V1. Offline-only or unverified quantitative A/B papers do not enter the implementation queue.
  • Remaining qualifying 2026 papers entered core-mechanism reproduction. The 2026-07-27 P1 batch added eight industrial papers and executable Engram, Looped Latent Attention, and GaugeQuant operators to LLM evolve. GRACE, DLMRec, LO-FAR, and PRL lacked quantitative online evidence and were not included.
  • The 2026-07-28 batch added Meta Mosaic, Kuaishou UniR², Meituan CORE, and DataOrchestra. The first three passed quantitative online A/B; DataOrchestra has official code and public pretraining benchmarks.
  • The 2026-07-30 batch added Tencent CCFormer, Teads Open Web UFM, Meta ROCS, and WIDE. The first three disclose A/B, full traffic, or quantitative deployment, respectively. WIDE provides code and public pruning and throughput experiments.
  • The 2025 industrial backfill added MIM, FilterLLM, FuXi-α, RecGPT-V2, HiGR, DRL-PUT, AdaF²M², MGOE, and Click A Buy B. All nine report production A/B and ran independent core mechanisms on MovieLens-1M.
  • 2025 LLM evolve P0 added DeepSeek NSA, Qwen Gated Attention, and Moonshot Muon. Structures and optimizers compose; ran matched WikiText-2 controls and four evolve rounds.

Google / Meta full-text evidence backfill (10 P0 adapters)

  • memory-layer · Meta Memory Layer: Training-time eta=1 writeback makes item cache model state. Always-on attributes cover cold start and missing IDs.
  • scalr · Meta SCALR: Learns cross-domain event conditionals and samples probabilistically instead of copying top-k to cover more training events.
  • hill-index · Meta HILL: Coarse-to-fine residual quantization learns hierarchical indexes and reduces online candidate scoring.

  • semantic-native-longseq · Google Semantic-Native Long Sequence Modeling: Hierarchical SIDs, bigrams, temporal folding, and global-local pooling extend video histories.

  • ha-moe · Google Discover HA-MoE: Heterogeneity-gated experts; 7-day A/B on 1% traffic.

  • dual-sid · YouTube Dual-purpose Semantic IDs: Semantic IDs encode collaborative identity and reconstruct content embeddings.
  • agentic-rec-tune · Google Discover AgenticRecTune: Actor–Critic–SkillHub multi-round recommendation optimization.
  • mfli · Meta MFLI: Learnable real-time retrieval indexes with multiple slices.
  • kunlun · Meta Kunlun: Unified deep architecture with GDPA, hierarchical seed pooling, and CompSkip.
  • ultra-hstu · Meta ULTRA-HSTU: Semi-local attention, LBSL, and Mixture of Transducers.

Candidates were retrieved by affiliation and topic, then checked in full text and tables without requiring A/B in the abstract. All six had public-data runs, stable metrics, an original key figure, and executable evolve mappings (AgenticRecTune maps to a controller regression, not an invented architecture).

2026 historical P0 · H02 (8 adapters)

  • drem · DrEM: Stabilizes multi-channel pxtr ranking with preference-preserving noise filtering, pair-label flip-risk inversion, and consistency fusion.
  • incrementality · From Prediction to Incrementality: Uses potential-outcome differences and uncertainty for matched-budget causal targeting instead of ranking treated outcomes.
  • tm20k · TM20K: The teacher retains full history; contiguous token merging and attention distillation compress the student's long sequence.
  • transx · TransX: Separates cacheable behavioral and per-request serving streams with cross-attention to reduce repeated work.
  • snaplgr · SnapLGR: Co-engagement propagation, hierarchical residual SIDs, token grounding, and SID sequence SFT.
  • psg · PSG: Generates ordered pairs and expands them into a slate, reducing six-item decoding to three steps.
  • director · DIRECTOR: Uses Sinkhorn transport and global matching for parallel duplicate-free dynamic-index slates.
  • reward-guided-decoding · Reward Guided Decoding: A KL-regularized closed-form policy adds business rewards without retraining the generator.

All eight retain public-data results for seeds 42/43/44, original key figures, online evidence, and reproduction limits. The structures are executable genrec evolve mutations.

2026 historical P0 · H03 (9 adapters)

  • specformer · SpecFormer: Spectrally softened attention and spectral residual position encoding.
  • egr · EGR: A shared encoder jointly learns item index vectors and user query vectors.
  • zorro · ZoRRO: Training-free ranking over recency, semantics, and category relations.
  • elise · Apple Music ELISE: Dense/lexical quantile calibration and hybrid retrieval.
  • poem · POEM: Partially ordered behavior sequences from real-time multi-objective ranking signals.
  • uniformer · UniFormer: Unified token interactions in feature and task spaces.
  • rag-generation · Recommendation as Generation: Disentangled SIDs connect recommendation and content generation.
  • onerank · OneRank: Transformer-native private task channels and gradient isolation.
  • piano · PIANO: Query-driven interest refinement and a listwise information node.

All nine execute independent core computation, retain three-seed public results and original key figures, and are executable genrec evolve operators. Local results establish mechanisms, not production lift.

2026 historical P0 · H04 (9 adapters)

  • atomic-intent · Atomic Intent Reasoning: Decomposes queries with hierarchical atomic intent trees, then reranks by candidate alignment with intent paths.
  • toolrec · ToolRec: Joint relevance, confidence calibration, and business cost filter over-triggered tool recommendations.
  • ssrlive · SSRLive: Static and dynamic SIDs plus user–live-room crosses unify content identity and real-time state.
  • taiji · Taiji: Adaptively selects Pareto operating points between semantic and collaborative matching.
  • primal-dual-decoding · Stochastic Primal-Dual Decoding: Updates decoding-time constraint multipliers to balance relevance and business constraints.
  • muchator · MuChator: Discovers latent intents in multi-turn conversations and ranks candidates using explicit and implicit feedback.
  • causal-representation · Causal Representation Learning for Generalisable Recommendation: Residualizes environment confounders for more stable causal preferences across environments.
  • policy-facet · Policy-Grounded Dynamic Facet Suggestions: Offline taxonomy, online retrieval, and policy-constrained ranking generate dynamic facets.
  • pa-bridge · PA-Bridge: Aligns representation distributions between active search and passive recommendation.

All nine retain production evidence and public-data limits, execute independent mechanisms with three MovieLens seeds, and register runnable genrec evolve mutations. Negative local results are neither hidden nor rewritten as improvements.

2026 historical P0 · H05 (10 adapters)

  • marc · MARC: Extracts task-relevant intermediate representations, decoupling recommendation adaptation from compression.
  • rankup · RankUp: Randomly permuted multiple embeddings, global tokens, and task decoupling improve effective rank.
  • sid-coord · SID-Coord: Hierarchical SIDs, target-aware HID–SID gating, and interest alignment support long-tail generalization.
  • rclrec · RCLRec: Builds short backward curricula from conversion-related behaviors for intermediate supervision.
  • tagllm · TagLLM: A user-interest handbook constrains fine multimodal labels, then distills tagging into a small model.
  • genfacet · GenFacet: Jointly generates facets and rewrites queries, aligning preferences with retrieval satisfaction.
  • cgr · CGR: Constraint-aware reward pruning directly within autoregressive reranking.
  • hpgr · HPGR: Session-level pretraining and preference-guided sparse attention restructure flat behavior sequences.
  • climber-pilot · Climber-Pilot: Distills multi-step consumption look-ahead into one-step retrieval with business instructions in attention.
  • rolegen · RoleGen: Finds bridge items for dormant users via functional roles and counterfactual conversion paths.

All ten have independent core computation, public seed 42/43/44 artifacts, original key figures, and executable genrec evolve operators. Paper online numbers and local mechanism results remain separate.

2026 historical P0 · H06/H07 (15 adapters)

  • unimvt · UniMVT: Jointly learns debiased CTR and causal uplift with explicit treatment and counterfactual estimates.
  • rq-gmm · RQ-GMM: Replaces hard clustering with residual Gaussian-mixture quantization for probabilistic SIDs in multimodal CTR.
  • capts · CAPTS: Value-aware matching and adaptive cross-channel routing select I2I retrieval triggers.
  • mlcc · MLCC: Hierarchical compression and low-dimensional feature crossing grow capacity within a controlled compute budget.
  • ug-sep · UG-Sep: Separates user-side and general computation to reuse representations computed once.
  • smes · SMES: Progressive sparse routing and deduplicated execution allocate expert capacity dynamically by task.
  • pit · PIT: Combines item tokens dynamically by user so SIDs encode personal preferences.
  • zenith · Zenith: Prime Token Fusion and Boost scale live-stream ranking.
  • easq · EASQ: Separate LoRA and multitask paths incorporate sparse survey satisfaction into online learning.
  • s2gr · S2GR: Injects semantic guidance in stages into latent reasoning for intermediate decisions.
  • sparsectr · SparseCTR: Combines global, transition, and local sparse attention over thousand-event histories.
  • hcub · HCUB: Hierarchical contextual uplift bandits learn catalog personalization incrementality.
  • airbnb-ebr · Airbnb EBR: Journey-level retrieval and multi-stage ranking unify experience recommendations.
  • promise · PROMISE: Process rewards supervise intermediate reasoning in generative recommendation.
  • harmonrank · HarmonRank: Relation-aware attention and consistency constraints coordinate multi-stage ranking.

All fifteen have independent mechanisms, public results at seeds 42/43/44, original key figures, and executable evolve operators. H06/H07 close the fixed 55-paper 2026 historical P0 queue.

2026-08-24 cross-domain additions

  • oneshot-index · OneShot: Meta/Instagram jointly trains a hierarchical index and ranking, replacing pure dot-product retrieval with neural scoring. Deployed to all global traffic.
  • clockrope · ClockRoPE: YouTube/Google DeepMind samples rotary frequencies from a periodic kernel's Fourier spectrum to model daily and weekly routines.
  • next-vlm · NEXT: Meta uses an 8B VLM to build a directed item→intent→item NKG; reasoning is offline and only verified edges are injected online.
  • onemodel · OneModel: Xiaohongshu uses a shared long-sequence backbone, scenario gates, and hierarchical user representations for recommendation, ads, and merchant ranking. Reports quantitative A/B in three production settings.
  • rare · RARE: Projects editing directions into the MoE router null space and corrects drift at downstream guard layers to preserve expert paths.

2026-08-26 additions

  • tagr · TAGR: Kuaishou livestream ads use stable two-level semantic/collaborative IDs, multiscale interests, and behavior-value gates. Production A/B: LRE +8.5%, SCC +7.4%, revenue +16.1%.
  • wemm-embedding · WeMM-Embedding: WeChat combines multimodal alignment, cross-scale teacher refinement, and Matryoshka embeddings. Reports full rollout of 14 online A/B gains and releases code and weights.

2026-08-27 additions

  • dceo · DCEO: Taobao Search uses a context-conditioned actor to optimize relative causal long-term user value. A 41-day production A/B reported GMV +0.36%.
  • transretrieval · TransRetrieval: Cross-domain ad retrieval with norm-aware aggregation, target-token compression, and positional domain embeddings. One-month 5%-traffic A/B revenue +2.53%.
  • vbvr-pro · VBVR-Pro: Task-specific executable scoring replaces generic VLM judging for state-level verifiable reward in native visual reasoning.
  • mllmclip · MLLMCLIP: Selects teacher tokens by attention across MLLM layers and aligns small CLIP visual/text features via CKA.

2026-08-29 additions

  • friend-gnn · VK Friend-GNN: Multi-hash sharing compresses massive user embeddings, while time-sorted adjacency and binary cutoffs prevent future leakage. VK production A/B: friend adds +16.0%, unique friend-adding users +11.5%.
  • pace-vlm · PACE: A shallow visual preview adapts token budgets with fused LLM/ViT attention. Verified real Qwen2.5-VL and RealWorldQA on A30.
  • twinkv · TwinKV: Pairs orphan and donor entries to repair arbitrary KV eviction policies at a fixed cache budget. Real Qwen3 KV and WikiText-2 long context verified on A30.

2026 historical scan B01–B03 (18 adapters)

  • dynamic-codebook · Dynamic Single-Level Large Semantic Codebook: Exposure-weighted dynamic large codebook, independent collision codes, and short SID decoding.
  • netflix-mediafm · Netflix Multimedia Asset Personalization: Frozen multimodal features, a unified asset dual encoder, and query-aware scoring.
  • ogr · Once Generated, Ranked: Unified semantic–collaborative IDs, slate planning, and conservative policy alignment.
  • inthq · IntHQ: Hierarchical task-interactive queries over short and long streams.
  • pushdualgen · PushDualGen: Generate SIDs first and interpretable copy on demand.
  • recharness · RecHarness: Bandit routing over multi-round architecture experiments under a limited budget.
  • gala · GALA: Multimodal triplet pretraining, GRPO behavioral alignment, and gated ID fusion.
  • feedback-policy · Feedback-Grounded Policy Discovery: Discovers generative policies from feedback and distills to a lightweight online ranker.
  • real-estate-rerank · LLM-Based Re-Ranking for Real Estate Search: Reranks using conversational intent, property text, and candidate-set statistics.
  • adaptive-ad-load · Adaptive Ad Load Design: Chooses ad count dynamically under revenue and conversion constraints.
  • guess-where-you-go · Guess Where You Go: Spatiotemporal SIDs, curriculum learning, and long-term feedback policy optimization.
  • genpage · GenPage: One model generates whole pages and is post-trained with long-term rewards.
  • journeyformer · JourneyFormer: Jointly encodes Airbnb guest journeys at short and long timescales.
  • l2rec · L2Rec: Personalized dual-view LoRA-MoE and cross-view fusion.
  • qgs · QGS: Joint query–item sequences, Linear HSTU, and generative search.
  • tubifm · TubiFM: A shared user-story foundation model unifies three ranking tasks.
  • pearl-percentile · PEARL: Multi-sample contrastive learning estimates behavior percentiles with low bias and variance.
  • dadf · DADF: Freezes the base model and learns distribution-aware multiplicative residual corrections.

All-domain P0 gap closeout (19 adapters)

Each write-up has upstream code status, core formulas, original figures, paper results, fair local baselines, stable metrics, and executable commands. Negative local results are retained.

All-domain P1 gap closeout (8 adapters)

2026-08-09 P0/P1 additions (9 adapters)

The first four industrial papers met quantitative online A/B or rollout criteria; the other five belong to foundation models. Core mechanisms ran and preserve seed 42 metrics, including negative and nonsignificant results.

2026-08-13 MR7 additions (3 adapters)

  • sona · Sona: Compressed history, autoregressive Semantic IDs, and an item ranker in one model replace a music recommendation cascade.
  • metastrategy · MetaStrategy: An LLM writes typed strategies compiled deterministically for multi-objective ranking.
  • gas · GAS: A removable MoT generation branch performs continuous next-embedding prediction for multimodal understanding without deployment overhead.

Both industrial papers report quantitative real-traffic A/B in the full text. MovieLens-1M experiments retain negative results rather than importing online lifts. GAS is foundation/multimodal and does not require online A/B; public-pixel runs and paper numbers are separate.

2026-08-20 industrial additions (2 adapters)

  • connectionmind · ConnectionMind: Meta and Michigan State University combine temporal heterogeneous graphs, shortest-path SFT, rule-reward GRPO, and path distillation for social recommendation. Multi-week A/B with tens of millions of users reports quantified impression, watch-time, and session gains.
  • dream · DREAM: Taobao adds three-level intent, strategy memory, typed parameter compilation, and offline/online Reward Dual Loop to existing retrieval and ranking. Fine ranking plus reranking: IPV +2.71%, Core IPV +3.06%, GMV +1.31%.

Both run their core training or control loops. ConnectionMind uses Delicious as in the paper and retains NDCG@10 −7.44%. DREAM's positive local correlation comes with pronounced head bias.

2026 historical scan B04–B06 (19 adapters)

Full text verifies production A/B for all 19. Independent mechanisms run on identical MovieLens-100K splits, candidates, and seed 42. Online and local results are reported separately.

B04

B05

  • aigq · AIGQ: Hybrid query generation and IL-GRPO.
  • safro · SaFRO: Satisfaction-aware dual relative policy optimization.
  • sort-ranking · SORT: System optimizations for an industrial Ranking Transformer.
  • quasid · QuaSID: Qualification-aware SID collision constraints.
  • gpl-prerank · Generative Pseudo-Labeling: LLM pseudo-labels improve online pre-ranking.
  • ltv-video-ranking · Long-term Value Prediction: Position debiasing and long-term value prediction.

B06

  • rgalign-rec · RGAlign-Rec: Ranker-guided latent query alignment.
  • linkedin-feed-sr · LinkedIn Feed SR: Industrial long-sequence feed ranking.
  • cadet · CADET: Ad CTR conditioned on post-candidate context.
  • diffureason · DiffuReason: Diffusion latent reasoning and GRPO.
  • sarm · SARM: MLLM semantic anchors for livestream ranking.
  • ml-dcn · ML-DCN: Masked low-rank DCN feature crosses.
  • rag-qac · RAG-QAC: Retrieval augmentation and multi-objective alignment for query completion.

2026 historical scan B07 (9 foundation and infrastructure adapters)

Full reproduction and implementation status

Fidelity Adapter / paper Reported paper evidence Local conclusion
Core mechanism state-opd · State-OPD No industrial A/B; offline interactive-task results Three arms × three seeds made actual A100 updates with receipts. Accuracy was zero on all eight questions, not a capability gain.
Deferred research material MagicBeanStalk (historical key magicbeanstalk, not registered) No industrial A/B; synthetic math and finance training Too costly; not reproduced. Full-method acceptance is pending. Code and real failures are retained, with no executable entry point.
Core mechanism gppo-framework · GPPO framework No industrial A/B; offline proximal optimization A100 three-seed mechanism diagnosis on eight questions with a separate GRPO control. No stable-improvement claim.
Core mechanism tapo-trajectory · TAPO trajectory No industrial A/B; offline long-trajectory results Three A100 seeds did not beat GRPO at this budget. Negative results retained.
Core mechanism cv-opd · CV-OPD No industrial A/B; offline visual-dependence results Actual updates for method and ablation at three A100 seeds each; test means 4.17%/12.5%. Negative results retained.
Core mechanism oal · OAL No industrial A/B; offline attention replacement A100 three-seed Q/K grouping, calibration, and LoRA conversion. Seed 44 was not better; not an original-scale capability or efficiency comparison.
Core mechanism replayssm · ReplaySSM No industrial A/B; state precision and system experiments Actual low-bit kernels and rollback ran on A100 for three seeds. JIT compatibility limits retained; no quality or HBM gains claimed.
Core mechanism lewam · LeWAM No industrial A/B; robotics and planning experiments A100 native joint training and a three-arm new-candidate loop at three seeds; no planning advantage. Single-task scaling diagnostic only.
Core mechanism smolwam · SmolWAM No industrial A/B; LIBERO experiments A100 three-seed teacher KD, curriculum, and new-student loop. Student and no-op both had zero success, not capability improvement.
Core mechanism near · NEAR No industrial A/B; offline memory retrieval A30 shared-bank three-seed run and strict bank-disjoint single-seed run completed. Three seeds on the same data are not independent replications.
Core mechanism beyond-final-embedding · Beyond the Final Embedding No industrial A/B; offline sparse-memory retrieval A30 token matching, ID fusion, and query adapter completed. Strict protocol has only two test queries, not the official benchmark.
Core mechanism skilldreamer · SkillDreamer No industrial A/B; skill and tool tasks Complete A100 mechanism receipt for three arms × three seeds. CSI, loading, and tools executed but produced zero submissions. Negative result retained.
Core mechanism synthself · SynthSelf No industrial A/B; multidomain post-training A100 three-seed aligner and SFT with acquisition and retention evaluation. Did not beat gold-SFT.
Core mechanism upturn · UPTURN No industrial A/B; inference interventions All three A100 seeds ran. 42/43 had rescue; 44 did not and the failure is retained. Only one seed has a valid receipt; no all-seed pass claim.
Core mechanism doublecat-semicat · DoubleCat/SemiCat No industrial A/B; diffusion LM results Corrected core without an extra forward pass passed on A30 at three seeds. One-hot and DoubleCat generation collapse retained.
Core mechanism soar · SOAR No industrial A/B; provenance-consistent reasoning Actual updates in six A100 method/ablation runs across three seeds. Ablation seed 43 had zero accuracy; four validation items are diagnostic only.
Core mechanism grpo-difficulty-curriculum · GRPO difficulty curriculum No industrial A/B; difficulty curricula Fifteen A30 runs, five arms × three seeds; 151 real updates and 38 exactly skipped equal-value groups. Negative results and previous numerical errors retained.
Core mechanism progressive-contraction · Progressive Contraction No industrial A/B; offline teacher-pool contraction A100 three-seed real Monte Carlo trees, student routing, and updates. Matches full-SFT per seed; no total-cost savings claim.
Core mechanism uxsid · UxSID Version 4 full text reports industrial A/B; see detail page Four MovieLens-100K structures with three A100 seeds passed cache consistency. No stable improvement or substitute for industrial reproduction.
Core mechanism wegr · WeGR WeChat Search production A/B: orders +2.539%, CTR +1.508% A100 three-seed MovieLens-100K diagnostic of CoCo-VAE, DHPO, and MSFT. Low performance and no gains over control retained.
Core mechanism casi · CASI Xiaohongshu seven-day randomized production A/B: Advv +3.51% CPU component with frozen snapshots, exact and permutation credit, and bounded zero-sum advantages. No LLM policy training; high-cost full experiment not reproduced.
Core mechanism pn-mix · PN-Mix No industrial A/B; offline cache quantization L1 reference with actual bit packing and attention consumption, not pretrained-LM quality or fused-kernel throughput.
Core mechanism smrc-pmrc · SMRC / PMRC No industrial A/B; optimizer training experiments Real parameter and optimizer-state updates, paper quadratic counterexample, and small MLP control. Full pretraining not reproduced.
Core mechanism oob-regularization · OOB Regularization No industrial A/B; self-supervised representation experiments Real Gaussian/spherical anchors and periodic refresh; four arms, three A100 seeds on official CIFAR training subset. No paper-scale or stable capability gain.
Core mechanism pace-vlm · PACE No industrial A/B; 10% visual tokens retained 93.8% average performance and up to 3.1× TTFT speedup Auditable APC/DDAE formulas, matched-checkpoint RealWorldQA quality/tokens/latency/memory, and evolve recipes.
Core mechanism twinkv · TwinKV No industrial A/B; offline LongBench/LooGLE/RULER/MMLU-Pro including failures Auditable fixed-budget repair, real Qwen3 KV reconstruction/latency/memory, and evolve attention operator.
Core mechanism prl-puts · PRL-PUTS P2P impressions +0.30% Two-head Q and Pareto utility sweep; NDCG@10 +39.64%.
Core mechanism ektm · Effective Knowledge Transfer for Multi-Task Recommendation Models eCPM +3.93% Task-similarity-driven transfer; NDCG@10 +10.98%.
Core mechanism adasid · AdaSID GPM +1.16% Dynamic SID from collision load and semantic compatibility; NDCG@10 -6.30%.
Core mechanism unirec-coa · UniRec GMV +5.60% Attribute chains and business preference alignment; NDCG@10 -4.93%.
Core mechanism uniscale · UniScale GMV +2.04% Joint Entire-Space scaling; NDCG@10 -2.20%.
Core mechanism gatesid · GateSID GMV +2.60% Cold-start-aware semantic/collaborative gating; NDCG@10 +34.95%.
Core mechanism aigq · AIGQ GMV +10.68% Hybrid query generation and IL-GRPO; NDCG@10 +8.09%.
Core mechanism safro · SaFRO Watch time +0.61% Satisfaction-aware dual relative policy optimization; NDCG@10 +81.86%.
Core mechanism sort-ranking · SORT Orders +6.35% Industrial Ranking Transformer system optimization; NDCG@10 -13.11%.
Core mechanism quasid · QuaSID GMV-S2 +2.38% Qualification-aware SID collision constraint; NDCG@10 -0.55%.
Core mechanism gpl-prerank · Generative Pseudo-Labeling CTR +3.07% LLM pseudo-labels for online pre-ranking; NDCG@10 +12.84%.
Core mechanism ltv-video-ranking · Long-term Value Prediction VV +2.49% Position debiasing and long-term value; NDCG@10 +9.73%.
Core mechanism rgalign-rec · RGAlign-Rec CTR@3 incremental over QE-Rec +0.13% Ranker-guided latent query alignment; NDCG@10 -14.03%.
Core mechanism linkedin-feed-sr · LinkedIn Feed SR Time spent +2.10% Industrial long-sequence feed ranking; NDCG@10 -0.12%.
Core mechanism cadet · CADET CTR +11.04% Post-candidate-context-conditioned ad CTR; NDCG@10 -6.20%.
Core mechanism diffureason · DiffuReason Ad consumption +1.15% Diffusion latent reasoning and GRPO; NDCG@10 -6.50%.
Core mechanism sarm · SARM Watch count +1.19% MLLM semantic anchors for livestream ranking; NDCG@10 +44.74%.
Core mechanism ml-dcn · ML-DCN Platform-wide CTR +1.89% Masked low-rank DCN feature crossing; NDCG@10 -2.08%.
Core mechanism rag-qac · RAG-QAC Suggestions taken +3.46% RAG and multi-objective query completion; NDCG@10 -4.48%.
Core mechanism dynamic-codebook · Dynamic Codebook Consumption +0.792%; decoding FLOPs about -48% Dynamic codebook and collision code; NDCG@10 -18.73%, negative result retained.
Core mechanism netflix-mediafm · Netflix MediaFM Search playthrough +0.36% Frozen multimodal features and query scoring; NDCG@10 -9.52%.
Core mechanism ogr · OGR Effective Views +1.120% Unified SID, slate planning, and conservative alignment; NDCG@10 +13.25%.
Core mechanism inthq · IntHQ UVCTR +1.60% Hierarchical task queries; NDCG@10 -7.38%.
Core mechanism pushdualgen · PushDualGen Effective play rate +8.50%, dissatisfaction -37.70% SID then optional copy; NDCG@10 +1.19%.
Core mechanism recharness · RecHarness ADVV +2.084%, revenue +0.534% UCB budget routing; NDCG@10 -9.02%.
Core mechanism gala · GALA Orders +0.55% Gated multimodal behavior alignment; NDCG@10 -5.60%.
Core mechanism feedback-policy · Feedback Policy Revenue +4.506%, ADVV +4.621% Feedback policy and conservative increments; NDCG@10 +2.76%.
Core mechanism real-estate-rerank · Real Estate Re-ranker CTR +5.3%, property tour bookings +4.8% Text/attribute set reranking; NDCG@10 -8.21%.
Core mechanism adaptive-ad-load · Adaptive Ad Load Revenue +36.8%, conversion -1.1% Constrained ad load; unchanged NDCG@10.
Core mechanism guess-where-you-go · Gwhere P-CTR +5.83%, U-CTR +6.20% Spatiotemporal SID next POI; NDCG@10 +9.38%.
Core mechanism genpage · GenPage Core engagement +0.24%, latency -20% Joint page utility; NDCG@10 -3.98%.
Core mechanism journeyformer · JourneyFormer Bookers +0.55%, booked nights +0.82% Long and short journey encoding; NDCG@10 -2.77%.
Core mechanism l2rec · L2Rec CTR +9.24%, reply rate +3.15% Personalized dual-view adaptation; NDCG@10 -5.04%.
Core mechanism qgs · QGS CTR +0.62%, PV duration +3.55% Query-conditioned SID scoring; NDCG@10 +0.73%.
Core mechanism tubifm · TubiFM Search TVT +3.9%, carousel TVT +0.30% Unified three-task model; NDCG@10 +0.45%.
Core mechanism pearl-percentile · PEARL Watch duration +2.10%, report rate -6.91% Multi-sample percentile; NDCG@10 -10.65%.
Core mechanism dadf · DADF Time spent +0.649%, production MAE -12.57% Frozen-base multiplicative residual; unchanged NDCG@10.
Core mechanism connectionmind · ConnectionMind Meta multi-week A/B with tens of millions of users: impressions +0.33%, watch time +0.43%, video sessions +0.22% Delicious-2K path SFT, GRPO, and distilled mixed inference. NDCG@10 -7.44%, negative result retained.
Core mechanism dream · DREAM Taobao fine ranking and reranking: IPV +2.71%, Core IPV +3.06%, GMV +1.31% Intent, strategy memory, safe compilation, and dual-loop gates. NDCG@10 +39.36% but head share +94.37%.
Core mechanism gas · GAS Multimodal: from-scratch overall 47.25→48.25; inference overhead 0% MoT, NEP, and EMA; Fashion-MNIST QA 61.1%→63.5%, deployment parameter overhead 0%.
Core mechanism sona · Sona Yandex Music active users +4.53%, duration +6.30%, likes +11.42% History compression, SID decoder, and ranker. NDCG@10 -92.05%, negative transfer retained.
Core mechanism metastrategy · MetaStrategy Taobao click PV +2.11%, IPV +3.12%, transaction value +2.83% Typed bundles and deterministic compiler. NDCG@10 -8.51%, head share -42.94%.
Core mechanism crsd · CRSD Meituan AdCTR +0.91%, AdCVR +1.06%, GTV +0.40% Two-stage reasoning distillation; MovieLens-100K NDCG@10 -2.14%, negative transfer retained.
Core mechanism twin-v2 · TWIN-V2 Kuaishou watch time in three settings +0.672%/+0.800%/+0.728%, main traffic Hierarchical compression with GSU/ESU; NDCG@10 +22.45%.
Core mechanism medusa · Medusa LLM: Medusa-1 2.2×, Medusa-2 2.3–3.6× Three future heads on WikiText-2; backbone calls -50%, identical outputs.
Core mechanism awq · AWQ LLM: TinyChat more than 3× over FP16 Activation-aware W4; output MSE -4.22%.
Core mechanism smolvlm · SmolVLM Multimodal: competitive at lower visual token cost Pixel-shuffle reduces visual tokens 16→4; Fashion-MNIST QA +19.0 points.
Core mechanism siglip2 · SigLIP 2 Multimodal: improved semantics and localization on several image–text benchmarks Sigmoid contrastive loss and masked-view self-distillation; Fashion-MNIST QA +56.9 points.
Core mechanism llava · LLaVA Multimodal: ScienceQA 92.53% MLP projector versus linear connector -3.4 points; negative small-data result retained.
Core mechanism blip2 · BLIP-2 Multimodal: zero-shot VQAv2 +8.7 points over Flamingo-80B Four-query Q-Former reduces tokens 16→4; accuracy -5.3 points, negative result retained.
Core mechanism speculative-decoding · Speculative Decoding LLM: T5-XXL 2–3× with unchanged distribution Low-rank draft and exact verification; target calls -75%, identical outputs.
Core mechanism clip · CLIP Multimodal: zero-shot ImageNet matches supervised ResNet-50 Bidirectional contrastive learning; Fashion-MNIST image–text accuracy 72.0%, shuffled images 10.3%.
Core mechanism sim · SIM Alibaba CTR +7.1%, RPM +4.4%, main traffic GSU/ESU long-history retrieval; NDCG@10 +32.97%.
Core mechanism dme · DME Douyin Search Lifetime +0.1%, internal offline +2.92% Typed latent and cross-reconstruction; MovieLens-100K NDCG@10 -8.84%, negative result retained.
Core mechanism steps · STEPS Active days +0.2843%, permission disablement -1.9089%; fully rolled out Planning/execution/filtering loop; MovieLens-100K NDCG@10 +66.05%
Core mechanism spear · SPEAR Query-view CTR +0.259 and reading depth +0.733; fully rolled out Dual embeddings and multiplicative gate; MovieLens-100K NDCG@10 +26.95%
Core mechanism open-language-model · OpenLanguageModel Standalone LLM: 90.6% weak scaling on four GPUs; no online A/B olm_composable genome; essentially equivalent WikiText-2 PPL at matched budget, verifying compositional semantics
Core mechanism gryphon-v2 · Gryphon-v2 Yandex Music active users +1.41% Dual rollout/impression teacher distillation; NDCG@10 -26.84%, retaining negative transfer
Core mechanism degr · DEGR JD.com UCTR +1.22%, PV +0.20% Diversity and adaptive-reward ORPO; NDCG unchanged, head share -1.21%
Core mechanism rd-attnres · RD-AttnRes Standalone LLM: 120M/343M PPL -2.97%/-2.43% PPL +0.61% versus Block AttnRes (worse); QK/V routing JS=0.00489
Core mechanism retoken · ReToken Visual Haystacks: Qwen3-VL-8B +13.4 points; no online A/B One retrieval target and value-cache Top-K; WikiText-2 PPL +3.70% (worse)
Core mechanism ccformer · CCFormer Tencent Video CTR +3.57%, ad revenue +1.71%; full rollout after the experiment Field-specific ID/content gates and hierarchical history compression; MovieLens-1M NDCG@10 +22.44%, tokens 24→12
Core mechanism open-web-ufm · Open Web UFM Teads 50/50 A/B: CTR +2.13%, eCPC -1.13%, visit rate +2.37% Two-crop contrastive pretraining and next-item proxy; MovieLens-1M NDCG@10 +0.00%, preserving the null result
Core mechanism rocs · ROCS Meta retrieval up to 3× QPS; short-video ranking LogLoss -0.5%, QPS +50% Request-once/candidate-late interaction; MovieLens-1M NDCG@10 +8.19%, in-process candidate scoring 129.58×
Core mechanism wide · WIDE Standalone LLM: at 50% sparsity, kernel prefill 1.98× and decode 4.95× Per-token head/FFN group Top-K; WikiText-2 PPL +0.81% (worse), no kernel speedup in dense PyTorch
Core mechanism asarl · ASARL QQ Channels CTR +2.69%, group GSB +16.66%, 12 million DAU Reason/Critic/Gen plus SCT/PGO/SD; NDCG@10 -72.72%, preserving negative transfer on the proxy
Core mechanism oxygenrec-v2 · OxygenREC-v2 JD.com UCTCVR +1.61%–+4.44%, homepage GMV +21.21% Behavioral instructions and privileged distillation; NDCG@10 -54.09%, preserving negative transfer on the proxy
Core mechanism reco-reward · RecoReward Kuaishou effective-user penetration +0.265%, outbound exposure +0.791% RAS target/non-target reward; Hit@10 -16.00%, retaining the negative result
Core mechanism twice · TWICE Expected revenue +2.486%, conversion +2.061%; full rollout Dual clocks and delay CDF; NDCG@10 +14.31%
Core mechanism swag-bid · SWAG GMV +3.42%, ROAS +5.65% Sliding-window masked planner; NDCG@10 +0.00%
Core mechanism youtube-freshness · YouTube Freshness 1-day new-release engagement +4.33% IPS, bias tower, and uncertainty; head share -28.72%
Core mechanism melo · Melo System-level A/B retention >2pp Grounding and retries; fresh Hit@10 +50.00%, with system attribution reported separately
Core mechanism penelope · Penelope Standalone LLM: improved structured-reasoning accuracy and compute Two-step localized recurrence; composite loss -0.63%
Core mechanism mim · MIM Taobao CTR +14.14%, RPM +4.12% Masked multimodal alignment and CiUBM; NDCG@10 +0.94%, Hit@10 declined
Core mechanism filterllm · FilterLLM Cold-PV +5.13%, GMV +10.86% Text-to-user distribution; NDCG@10 -9.82%, retaining the negative result
Core mechanism fuxi-alpha · FuXi-α Songs played +4.67%, play duration +5.10% Three-channel interaction; Hit@10 slightly higher, NDCG@10 -4.25%
Core mechanism recgpt-v2 · RecGPT-V2 Taobao CTR +2.98%, IPV +3.71%, NER +11.46% Hierarchical agents and constrained routing; NDCG@10 +18.52%
Core mechanism higr · HiGR Tencent plays +1.73%, watch time +1.22% Hierarchical SID slate and ORPO; NDCG@10 -12.32%
Core mechanism drl-put · DRL-PUT Pinterest revenue +0.27%, CTR +1.62% Logged-bandit policy weight adjustment; NDCG@10 +19.13%
Core mechanism adaf2m2 · AdaF²M² Douyin active days +1.37%, duration +1.89% Feature-mask multi-forward with adapter; NDCG@10 +2.42%
Core mechanism mgoe · MGOE Alibaba GMV +16.46%, CVR +5.88% Macro task graph experts; NDCG@10 +7.03%
Core mechanism click-a-buy-b · Click A Buy B Pinterest primary business metric +0.25% CABA/CABB with taxonomy; NDCG@10 +33.70%
Core mechanism mosaic · Mosaic Meta three surfaces: +0.10%/+0.15%/+0.28% Four specialists with MRM and CRL; NDCG@10 +3.49%, Hit@10 -7.69%
Core mechanism unir2 · UniR² Kuaishou plays +1.177%, like rate +2.560%; Lite gift spending +2.569% DQ-PCA and ranking LoRA; SID code accuracy +34.04%, NDCG@10 -13.19%
Core mechanism core-relevance · CORE Meituan NDCG@5 +0.20%, Badcase@5 -15.9% Cascaded heads, step-GRPO, and PostCoT; NDCG@5 +0.98%, Badcase@5 -50.00%, accuracy -0.52 points
Core mechanism data-orchestra · DataOrchestra Standalone LLM: consistent average benchmark gains at 0.5B/1.5B/7B Per-example routing; PPL -1.03% versus fixed cleaning, +8.60% versus raw (worse)
Core mechanism native-sparse-attention · NSA Standalone LLM: quality at least as good as full attention, with significant speedups in three stages at 64K Three-path sparse attention; attention edges -43.65%, PPL -3.17%
Core mechanism gated-attention · Gated Attention Standalone LLM: improved scaling, stability, and long-context extrapolation Per-head post-SDPA gate; PPL -0.72%
Core mechanism muon · Muon Standalone LLM: about 2× compute efficiency Orthogonal matrix updates ran; PPL +5.61% (worse)
Core mechanism wide-deep · Wide & Deep Google Play acquisition +3.9% Wide feature crosses and deep tower; NDCG@10 +5.45%
Core mechanism deepfm · DeepFM User-approved classic exception Shared embeddings across FM and deep paths; NDCG@10 +23.58%
Core mechanism youtube-dnn · YouTube DNN User-approved classic exception; no quantified lift disclosed Nonlinear user tower; NDCG@10 -6.61%, retaining the negative result
Core mechanism esmm · ESMM User-approved classic exception Entire-space CTR×CVR; mean AUC +1.69%
Core mechanism mmoe · MMoE User-approved classic exception Shared experts and task-specific gates; mean AUC +1.30%
Core mechanism ple · PLE User-approved classic exception Shared and task-specific experts; mean AUC +1.34%
Core mechanism dcn-v2 · DCN-V2 Classic exception: significant online result without disclosed lift Low-rank cross experts; NDCG@10 +22.87%
Core mechanism dien · DIEN CTR +20.7%, eCPM +17.1% GRU, auxiliary objective, and interest evolution; NDCG@10 -1.98%
Core mechanism bst · BST CTR +7.57% Target-token Transformer; NDCG@10 +20.29%
Core mechanism cs3 · CS3 Revenue +8.356%/+1.366%/+2.177% Cycle/sync/cascade; NDCG@10 -16.06%
Core mechanism cq-sid · CQ-SID GMV +1.15%, UCTCVR +0.40% Category SID and EG-GRPO; NDCG@10 +1.66%
Core mechanism switch-transformer · Switch Transformer Standalone LLM: about 7× faster pretraining Top-1 MoE; 30-step PPL -3.29%
Core mechanism mamba · Mamba Standalone LLM: up to about 5× faster inference Selective scan; 30-step PPL +48.82% (worse)
Core mechanism switch-attention · Switch Attention Standalone LLM: >4× decoding at 32K Dynamic full/local router; 30-step PPL +0.19% (slightly worse)
Core mechanism onemall · OneMall Product-card GMV +13.01%, short-video orders +15.32%, livestream orders +2.78% Surface prompts, SID, and cross-behavior fusion; NDCG@10 +4.33%
Core mechanism dos · DOS Meituan revenue +1.15% Dual-stream ORQ; NDCG@10 +11.26%
Core mechanism mdl · MDL Douyin LT30 +0.0626%, rewrite -0.3267% Three token types and domain attention; NDCG@10 +13.34%, head share +73.43%
Core mechanism hisac · HiSAC Taobao CTR +1.65% Hierarchical interest agents; NDCG@10 +1.31%
Core mechanism pinclip · PinCLIP Fresh Repin +15%, new ads click +8.7% Neighbor alignment NDCG@10 -1.41%; reported gains did not transfer
Core mechanism pin-scale · Pin-SCALE Repin +3.67%, DAU +0.05% Engagement-aware SID; NDCG@10 +13.61%, fresh Hit +50.00%
Core mechanism causal-retrieval · Causal Retrieval Triggering -85%, sessions +0.26%, Save +1.10% DR uplift triggering; synthetic-treatment NDCG@10 +80.77%
Core mechanism podcast-mtl · Podcast MTL eCPS -22%, streams +18%–24% Shared MTL NDCG@10 -20.63%, showing negative transfer
Core mechanism engram · Engram Standalone LLM: MMLU +3.4, BBH +5.0, HumanEval +3.0 O(1) memory integrated into evolve; 30-step PPL +50.12% (worse)
Core mechanism looped-latent-attention · Looped Latent Attention Standalone LLM: up to 32× KV compression Integrated into evolve; parameters -42.28%, PPL +5.56%
Core mechanism gaugequant · GaugeQuant Standalone LLM: LLaMA-2 7B W4A4 PPL 8.22→6.73 Integrated into evolve; local W4A4 STE PPL -1.56%
Core mechanism nova · NOVA Tencent GMV +1.25%/+1.70%/+2.02% for three pCVR objectives Four-level verification cascade, failure direction, and architecture gradient actually integrated into evolve
Core mechanism evorec · EvoRec Revenue +1.85%, CTR +1.02% Three generations of dual-track model and method evolution with persistent skill memory
Complete core pipeline tokenmixer-large · TokenMixer-Large Orders +1.66%, payment GMV +2.98% Mixing/reverting, two SwiGLUs, interval residuals, and auxiliary losses actually trained
Core mechanism msn · MSN Watch time +0.2958%, finish rate +0.2071% Product-Key Memory, sparse top-k reads, and gate
Core mechanism idproxy · IDProxy Content interactions +0.50%, ad ADVV +1.93% Contrastive loss 6.073→5.358; NDCG@10 +5.32%
Core mechanism glide · GLIDE Non-habitual streaming +5.4%, new-show discovery +14.3% Residual Semantic ID generation and separate long/short-term prompts
Core mechanism genrec · GenRec Clicks +9.5%, transactions +8.7% Page-wise NTP, Token Merger, and GRPO-SR/NLL
Core mechanism genrec-netflix · Netflix GenRec 10% of traffic for 4 weeks; short-term homepage engagement +0.115%, long-term primary metric +0.006% SmolLM2 LoRA, catalog head, and reward weighting; NDCG@10 +35.07% (one seed)
Core mechanism rankgraph2 · RankGraph-2 CTR +0.96%, CVR +2.75% Debiased edges, multi-hop PPR, and two-level residual index; NDCG@10 +109.65%
Core mechanism solaris · SOLARIS Full-traffic top-line revenue +0.67% Future-pair predictor, asynchronous latent cache, and fallback
Core mechanism minimax-sparse-attention · MiniMax Sparse Attention Standalone LLM: 1M-context attention compute -28.4× (reduction, as reported) Attention pairs -79.95%, PPL +0.41% (worse), unfused MPS runtime +67.58%
Core mechanism pinequalizer · PinEqualizer Related Pins ranking architecture: all-fresh +8.63%, underexplored +6.57% MovieLens-1M fresh NDCG +448.62%, underexplored exposure +471.02%, but overall NDCG -16.44%
Core mechanism gzip-sparse-attention · Gzip-guided Sparse Attention Standalone LLM: PG-19 BPB 2.34→1.71 versus BigBird At matched parameter count and 256-token context, BPB +1.02% versus BigBird (worse), attention edges -70.82%
Core mechanism windowed-mtp · Windowed-MTP Standalone LLM serving: 1M step latency metric +28.3%–+44.3% as reported 16K KV reads -99.56%, MPS draft latency -50.25%; exactly identical outputs, acceptance 41.67%→25.00%
Core mechanism adadsf · AdaDSF Standalone LLM: at 80% retention, PPL 21.6→18.9 versus MoD Matched teacher and budget: PPL +0.30% versus Uniform MoD (worse)
Core mechanism barge · BARGE Tencent CTR +0.60%, unique clicking users +1.34%, reading time +1.70% OSQ+ICA+HPR+DPD; NDCG@10 +85.77%, but head share +165.02%
Core mechanism mobius-rope · Möbius RoPE Standalone LLM: needle score 63.3%→90.3%, PPL flat PPL -0.03%, single-seed needle -2.08 points; paper gains did not transfer
Core mechanism naju · Naju Standalone LLM: WikiText-103 PPL 28.31→26.20 Preserve-first gates correct; WikiText-2 PPL +25.67% (worse)
Core mechanism dynamic-rubric · DynamicRubric WeChat Search: statistically significant at full AI traffic; exact lift undisclosed Alpaca preference accuracy 83.89%→87.78%, relative +4.64%
Core mechanism tsgr · TSGR Taobao IPV +0.43%, transactions +1.12%, GMV +1.64% Parallel value SID and joint VRM; NDCG@10 +115.73%
Core mechanism off-context-grpo · Off-Context GRPO Pure LLM: Qwen2.5-7B +13.8% over GRPO Official GSM8K Pass@1 2.67%→3.33%, +25.00% relative
Core mechanism ramp · RAMP Industrial CVR Total Advertiser Value >+3% Traffic with restricted personalized fields: NDCG@10 +417.23%
Core mechanism whale · WHALE Meta primary metric +0.113%, Metric 1 +0.824% Progressive Wukong–HSTU exchange; NDCG@10 -83.20%, gains did not transfer
Core mechanism tmallgs · TMallGS Tmall UCTCVR +1.38%, GMV +1.52% Field-adaptive gated Transformer; NDCG@10 +310.42%
Core mechanism long-history-transformer · Long-History User Transformers Yandex Search primary metric +2.77%, Revenue +2.26% Cached full history plus recent online encoder; NDCG@10 +57.08%
Core mechanism downstream-rewards · Downstream Rewards Across surfaces +0.11%–+0.36% Selected long-term reward head; NDCG@10 -5.10%, gains did not transfer
Core mechanism recgpt-mobile · RecGPT-Mobile Taobao CLICK +1.8%, PAY +2.7%, GMV +2.5% Real 135M LoRA semantic intent accuracy +100.00%; INT8 relative -6.25%, size -53.68%, trigger skipped 96.21% of inference calls
Core mechanism sort-gen · SORT-Gen Versus deployed baseline: CLICK +4.13%, GMV +8.10% Ordered regression plus single-pass mask-driven queue generation; Click +5.10%, Pay +8.46%, GMV proxy +9.00%, ILAD +2.89%
Core mechanism recgpt-v3 · RecGPT-V3 Taobao Feed IPV +1.28%, CTR +1.00%, GMV +3.97%; resources -52.4% After two-stage teacher distillation: NDCG@10 +36.96%, memory tokens -65%, latent slots -90%, but head share +71.43%
Core mechanism slimper · SlimPer Statistically significant full-traffic lift on Instagram Reels/Feed; exact lift undisclosed Parameter-matched NDCG@10 +1.29%; attention-score elements -94.12%
Core mechanism tokenminds · TokenMinds YouTube SFV Engaged Users +0.11%, Satisfied Engagement +0.62%, full traffic across multiple surfaces Actual dense + SID user-token dual-output training; one-seed NDCG@10 -0.35%, no stable improvement claim
Core mechanism recap · RECAP Kuaishou usage time per user +0.139% GRPO reward 0.5245→0.7096, but NDCG@10 -6.77% and head share -20.27%
Core mechanism uame · UAME LongView up to +1.614%, Forward up to +1.598% Three public proxy inputs: NDCG@10 -62.28%; online gains did not transfer
Core mechanism conv-llm · Convolution for LLMs Pure LLM: Qwen3-1.7B PPL 13.42→12.79 Matched-budget WikiText-2 test PPL 305.664→304.787 (-0.29%)
Core mechanism ppl-factory · PPL-Factory Pure LLM with 10% of data: GSM8K +0.9 and MATH +4.8 points 20% middle selection had 1.79% worse PPL than random; easy selection was best
Core mechanism fluid · FLUID QWD +0.55%, cold-start room plays +2.05% Without candidate IDs: NDCG -20.63%, fresh Hit +100.00%, head share -58.32%
Core mechanism memory-grafting · Memory Grafting Pure LLM: benchmark average 53.86 PPL -3.59% versus Transformer, but +0.03% versus Engram; did not beat directly trainable memory
Core mechanism mhc · mHC Pure LLM: benchmark +2.1%–2.3% No PPL improvement; residual row/column errors reached zero, spectral norm 1.089→1.000
Core mechanism degre · DeGRe Taobao Flash CTR +2.85%, GMV +3.75% Evaluator→beam→dense distillation; NDCG@10 +3.31%
Core mechanism harness-lm · HARNESS-LM Bing Ads Revenue +1.0%, Clicks +0.4% Three stages converged but test NDCG -28.05%
Core mechanism grc · GRC Revenue +1.79%, CTR +2.11%, GMV +2.04% Structured SFT→GRPO→EGRS; NDCG -11.12%
Core mechanism mbgr · MBGR Meituan CTCVR +3.98% BID/MBP/LDR; NDCG -5.92%
Core mechanism growthgr · GrowthGR New-item GMV +5.3%, site-wide GMV +0.3% ItemLTV→SID→MoPO; NDCG +2.05%
Core mechanism mesh · MESH Fresh repins +5.5%, retention +0.46% Three towers and RGBC; NDCG -3.54%
Core mechanism sam · SAM CTR +1.1%, GMV +0.9%, bad cases -74.5% ASGU/TTNP; NDCG -6.60%
Core mechanism danet · DANet pCVR +3.63%, GMV +2.23% TFTM/DCM; NDCG -1.46%, fresh Hit +50.00%
Core mechanism proximity-features · Proximity Features First-time bookers +2.0%, bookings +0.16% Adaptive ZIP buckets; Hit@10 +16.67%, NDCG +22.91%
Core mechanism nontp · NONTP Meituan DSP CTR +1.8%, GMV +2.1% EMA-teacher TCL, cross-domain TDL, and zero additional inference executed; Hit@10 -4.93%, NDCG -8.62%
Core mechanism akt-rec · AKT-Rec Tmall CTR +2.76%, GMV +3.47% Actual small LLM, RQ-VAE, and asymmetric transfer; AUC +3.44%, GAUC +5.53%, tail AUC +2.15%
Complete core pipeline s-grec · S-GRec WeChat GMV +1.19%, CTR +1.16%, dislike -2.02% Actual LLM PSJ, SID generator, and 5% A2PO. Validation promoted A2PO; test HR@10 +0%, NDCG -4.53%, no constraint violations
Complete core pipeline pinterest-ads-llm · Complementary LLM Predictor US Shopping RoAS +4.94%, opt-in +6.69% SFT selected; GRPO Recall@20 +0%, LLM ranking feature AUC +2.59%, retrieval quota=0
Complete core pipeline lwgr · LWGR Ads revenue +1.35%, CTR +1.17% Reference selected; LWGR Recall@10 +0%, NDCG -4.29%; dual update ran without improving the constraint
Complete core pipeline sigma · SIGMA AliExpress Order +2.80%, GMV +7.84% Top-1 prefix HR@20 1/128→9/128; APF -11.11% relative to top-1
Complete core pipeline univa · UniVA WeChat Channels GMV +1.50%, GMV(normal) +1.42% Public Office proxy; HR@100 +4.76%, ValueHR +6.56%, but wNDCG -8.43%; 50/50 trie paths valid
Core mechanism prompt-generation · Prompt Generation Taobao Search transactions +0.47%, GMV +0.51%; Shop Search +4.01% Same-source Amazon Office plus Qwen2.5-0.5B; selected Title HR@10 -11.11%, mean merger -90.38% versus raw Title scoring
Complete core pipeline precise · PRECISE WeChat ranking Clicks +1.961%, Shares +1.433% SmolLM tokens, MoE, and UT/TT; Recall@10 +40.0%, Cold Recall -50.0%
Complete core pipeline lum · LUM Taobao CTR +2.9%, RPM +1.2% Next-condition-item, grouped queries, and DLRM; AUC +14.60%, positive in 3/3 seeds
Complete core pipeline lsvcr · LSVCR Kuaishou comment watch time +4.1264% q/v-LoRA + SSC/VCC; comment NDCG +50.40%, item NDCG -56.42%
Complete core pipeline msd · MSD Meituan CTR +2.12%, CPM +2.59% Teacher→T5 distillation plus LoRA/cache fusion; mean AUC +1.55%, positive in 2/3 seeds
Core mechanism sessionrec · SessionRec Meituan Pay PV +0.603%, PVCTCVR +0.564% Real KuaiRand sessions and exposed negatives; NDCG@20 -22.05%, positive in only 1/3 seeds
Core mechanism saviorrec · SaviorRec Taobao Clicks +13.31%, Orders +13.44%, CTR +12.80% Behavior encoder + RQ-SID + MBA + BiTargetAttn; mean cold AUC +6.92%, positive in only 1/3 seeds
Complete core pipeline pinrec · PinRec Grid Clicks +4.01%, Time Spent +0.55% Actual OC + unordered MT training; Recall@10 -27.78%
Complete core pipeline genrank · GenRank Engagements +1.2474% Action-oriented latency -25.66%, AUC -0.46%
Complete core pipeline learn · LEARN Cold-item Revenue +8.77% NDCG +233.10%, but head share 69.50%
Complete core pipeline notellm · NoteLLM I2I CTR +16.20% GCL+CSFT; NDCG +7.15%, positive in 3/3 seeds
Complete core pipeline kar · KAR Huawei news Recall +7%; music plays +1.70% Actual LLM knowledge generation and hybrid experts; mean AUC +0.81%, positive in 2/3 seeds
Complete core pipeline bahe · BAHE Ads CTR +9.65%, CPM +2.41% Atomic behavior caching plus upper-layer aggregation; runtime -53.61%, AUC -2.94%
Complete core pipeline beque · BEQUE Taobao GMV +0.40%, transactions +0.34% SFT, leakage-free self-sampling, offline feedback, and PRO; feedback +30.03%, increment -66.02%
Complete core pipeline onerec-v2 · OneRec-V2 Kuaishou stay +0.467%, Lite +0.741% Actual KuaiRand duration feedback; Lazy latency -54.78%, mean GBPO +21.66% with only 2/3 positive seeds
Complete core pipeline plum · PLUM YouTube Panel CTR +0.76%/+4.96% CPT lowered loss; Recall@10 R1/CR1 0.5%, R2/CR2 0; no verified retrieval gain
Complete core pipeline onerec · OneRec Kuaishou watch time +1.68% Core paths executed; DPO reduced local NDCG@10 from 0.0157 to 0
Complete core pipeline g2rec · G2Rec Meta +0.06%–+0.19% Soft graph and generative dual loss on Beauty; NDCG@10 +11.92%
Complete core pipeline mixformer · MixFormer Douyin duration +0.2799% Matched-budget trainable blocks; NDCG@10 +17.41%
Complete core pipeline rankmixer · RankMixer Active Days +0.3%, duration +1.08% Dense per-token FFN performed best; sparse MoE did not match dense
Complete core pipeline hyformer · HyFormer Watch time +0.293%, finish +1.111% NDCG@10 +143.77%, with higher head share
Complete core pipeline onetrans · OneTrans Feeds GMV/U +5.6848% NDCG@10 +123.58%, but 92% of recommendations were head items
Complete core pipeline rec-distill · Rec-Distill Ads ADVV +1.00%, Rec Finish/U +1.2725% After α search, transferability -4.11%; no verified distillation gain
Complete core pipeline sasrec · SASRec None; user-designated classic baseline exception Original BCE and full-catalog inference; NDCG@10 0.02933, -1.24% versus popularity
Core mechanism hstu · HSTU Meta engagement +12.4%, consumption +4.4% Matched sampled-softmax SASRec control; NDCG@10 -17.73%
Core mechanism m6rec · M6-Rec Alipay mini-app CTR >+1.0% Real frozen pretrained LM; mean option-adapter AUC +0.12% ± 0.41%
Core mechanism din · DIN Alibaba CTR +10.0%, RPM +3.8% Local activation and Dice actually trained; NDCG@10 -6.97% versus mean pooling
Core mechanism tiger · TIGER None; user-designated classic paper exception Actual RQ-VAE and autoregressive retrieval training; NDCG@10 -39.16% versus capacity-matched random IDs
Core mechanism transact-v2 · TransAct V2 Pinterest Repin +6.35%, Hide -12.80%, Time Spent +1.41% NDCG@10 +92.65%, but head share rose to 98.99%
Core mechanism pinfm · PinFM Pinterest Homefeed Saves +1.20%–+5.70% Two rounds of pretraining/fine-tuning with validation selection; test -3.57%, head share fell to 20.16%
Core mechanism sis · SIS Not part of this batch's A/B set SIS formula executed; Qwen3/GRPO not trained
Core mechanism mdcns · MDCNS Public offline paper results Author's Beauty split; three-source sampling and dual-model updates actually ran
Core mechanism memento · Memento Meta CTR +1.0%, CVR +1.2% Query-conditioned MMR executed; production replay and serving omitted
Core mechanism llm-ad-retrieval · LLM Retrieval Meta top-line +0.45%, final recall +1.2% Domain SFT and LLM attribute graph; Recall@20 +11.90%, score drift -77.36%
Complete core pipeline seral · SERAL Taobao clicks +29.56%, transactions +27.6% NDCG +50.60% versus DIN; novelty did not improve
Complete core pipeline leadre · LEADRE WeChat GMV +1.57%/+1.17% +12.94% versus DIN; DPO ablation -4.53%
Complete core pipeline cobra · COBRA Conversion +3.60%, ARPU +4.15% +25.75% versus DIN; higher popularity concentration
Core mechanism argus · ARGUS Listening +2.26%, likes +6.37% -4.12% versus DIN; no verified gain
Core mechanism gr4ad · GR4AD Kuaishou ad revenue +4.2% +69.67% versus DIN, head share 0.505
Core mechanism cross-domain-kd · Cross-domain KD Music discovery +1.12% Target split -68.46% versus DIN
Core mechanism mm-llm · MM-LLM Meta engagement +0.02% -13.23% versus DIN; no verified gain
Core mechanism cluster-goobs · Cluster GOOBS Meta CTR +53% Online sampler executed; genre replaced private LLM clusters
Core mechanism llatte · LLaTTE Meta conversion +4.3% BERT semantic features, MLA, target-aware online attention, and DHEN all actually trained
Core mechanism self-evolving-rec · Self-Evolving RecSys Google +0.03%–+0.14% Local instruction LLM reads the journal, proposes changes each round, receives validation feedback, and uses isolated test
Core mechanism cmsl · CMSL Meta +0.092%–+0.171% Learned contextual lenses and HSTU-style backbone actually trained
Core mechanism longer · LONGER Douyin Ads/e-commerce A/B InnerTrans, global tokens, and hybrid attention trained end to end

Running reproductions

# One paper
auto-research reproduce --paper <adapter> --seed 42

# All papers
auto-research reproduce --paper all --seed 42

# Include explicitly downgraded concept demos
auto-research reproduce --paper all --include-concept-demos --seed 42

Raw run artifacts:

runs/reproductions/<arxiv-id>-<adapter>/<timestamp>/
├── report.md
└── result.json

Git excludes runs/. The single-run source is result.json. The paper README summarizes reviewed stable conclusions, protocol, and reproduction limits.

Adapter layout

src/auto_research/reproductions/<adapter>/
├── adapter.py
├── model.py or algorithm.py
├── experiment.py
└── report.py

Shared public splits and per-user/matrix full-catalog metrics are in reproductions/rec_utils.py, all-position sequence training in reproductions/sequence_training.py, and downloads in datasets.py. Paper-specific networks, sampling, tuning, and reporting remain in each implementation. Every README contains metadata, original background, main changes, Mermaid architecture, core equations, paper offline/online results, local protocol, and fidelity limits. scripts/sync_reproduction_metadata.py synchronizes metadata. See architecture for extension rules.