Search, ads, recommendation, and LLM applications: lineage and gaps

简体中文

This page maps major research lineages, how their approaches evolved, and what remains missing. See the reproduction overview and individual papers for formulas, online results, and local experiments. Expanding the catalog does not relax admission rules.

Lineage coverage

Lineage	Representative methods	Repository coverage	Status
Classic CTR and feature interaction	Wide & Deep, DeepFM, DIN, DCN-V2	Wide/deep models, FM, candidate-aware interest, explicit feature crosses	Implemented
Sequential interest and long histories	DIEN, BST, SASRec, HSTU, LONGER, RankMixer, HyFormer	GRU interest evolution, causal attention, token mixing, long-sequence compression	Implemented
Multitask learning and conversion	ESMM, MMoE, PLE, TWICE	Entire-space modeling, task gates, specialized and shared experts, dual clocks for delayed conversion	Implemented
Dual-encoder and graph retrieval	YouTube DNN, CS3, RankGraph-2, HARNESS-LM	Sampled retrieval, dual-encoder cyclic correction, multi-hop PPR, asymmetric lightweight retrieval	Implemented
Generative recommendation	TIGER, OneRec, OneMall, CQ-SID, RecGPT-V3	Semantic IDs, sequence generation, multi-scenario prompts, generative correction, and latent intent	Implemented
LLM content and knowledge augmentation	KAR, NoteLLM, SERAL, RecoReward, Melo	Knowledge generation, content compression, cognitive profiles, recommender rewards, agent grounding	Implemented
Reranking and blending	SORT-Gen, Memento, DeGRe, DRL-PUT	Listwise value, MMR, look-ahead distillation, utility-policy reweighting	Implemented
Content understanding and relevance screening	MM-LLM, MIM, PinCLIP, CORE, ASARL	Multimodal representations, content/collaborative alignment, relevance cascades, and data screening	Implemented
Privacy and policy risk	RAMP, Proximity Features, Causal Retrieval, UAME	Missing fields, privacy compliance, incremental triggers, and uncertainty constraints	Partial
Ad generation and long-term value	GR4AD, UniVA, GrowthGR, SWAG, Causal Retrieval	Generative ads, value alignment, long-term objectives, sliding-window bidding, and causal incrementality	Implemented
Freshness, cold start, and exploration	PinCLIP, Pin-SCALE, YouTube Freshness, Pinequalizer	Content/collaborative alignment, engagement SIDs, IPS, uncertainty, and exposure debiasing	Implemented
Training and serving efficiency	OneTrans, SOLARIS, Rec-Distill, TokenMixer-Large	KV caching, asynchronous latents, distillation, and sparse capacity scaling	Implemented
Self-evolving recommenders	Self-Evolving RecSys, NOVA, EvoRec	Experiment memory, validation cascades, and dual-track model/method evolution	Integrated with evolve

Technical evolution

flowchart LR
    W["Wide & Deep"] --> F["Feature interactions<br/>DeepFM / DCN-V2"]
    W --> I["Interest modeling<br/>DIN / DIEN"]
    F --> M["Multitask learning<br/>ESMM / MMoE / PLE"]
    I --> S["Sequence modeling<br/>BST / SASRec"]
    M --> L["Long sequences and large-scale ranking<br/>HSTU / RankMixer / LONGER"]
    S --> L
    L --> G["Generative retrieval<br/>TIGER / OneRec / CQ-SID"]
    G --> R["Feedback and value alignment<br/>GRPO / long-term value / causal incrementality"]
    R --> A["LLM agents and automated evolution<br/>Melo / NOVA / EvoRec"]

These are composable capability layers, not a strict publication timeline. Traditional ranking remains a serving baseline for most generative and LLM systems. Generative models, agents, and RL are marked implemented only after their core mechanisms run on public data.

Industrial evidence threshold

• New industrial papers must disclose quantitative production A/B results or a statistically significant full rollout explicitly accepted by the user.
• Report online lifts, local public-data results, and component ablations separately. None substitutes for another.
• Named classic exceptions such as DeepFM, YouTube DNN, ESMM, MMoE, and PLE fill gaps in the technical lineage. They are not precedents for waiving online evidence on later papers.
• If system-level A/B cannot isolate model, product, and UI effects, state the attribution limits explicitly, as for Melo.

See the full selection and reporting rules.

Current gaps

Gap	Reason	Prerequisite for implementation
Content safety, fraud, and ad compliance	Current implementations mainly cover relevance and privacy constraints and cannot replace safety moderation	Labeled public data, precision/recall/false-positive metrics, and production guardrails
Private large-scale ad auctions and conversion logs	Public data lacks bids, budgets, GMV, and mature labels	Public logs or an equivalent benchmark that preserves the decision constraints
End-to-end recreation of LLM recommendation products	Retrieval indexes, model checkpoints, and traffic orchestration are usually private	Official code or model, or an auditable public end-to-end environment
AIGQ, RaG, RoleGen, LCU	Missing rewards, video feedback, conversion trajectories, or public data access, respectively	The corresponding public data and executable feedback loop
RecoChain / DIG	Quantitative production A/B evidence has not been verified	Verifiable production experiment numbers in the paper body

These remain gaps. Do not create adapters that provide only a name, a formula, or fixed scores.