Cross-domain paper lineages and gaps¶
This long-term audit covers search, ads and recommendation; foundation models; LLM post-training; and agents. Individual paper READMEs remain authoritative for implementations and metrics. This page asks whether the major lineages are covered and what remains missing.
A second-level topic review on 2026-08-08 confirmed real gaps. The earlier “P0/P1 complete” status referred only to a 15-paper batch, not the whole library. See the all-domain gap audit for evidence, priorities, and rejected candidates.
Common admission criteria¶
| Domain | Evidence needed for implementation | Default public evaluation |
|---|---|---|
| Industrial search, ads, and recommendation | Quantitative online A/B or full-rollout evidence explicitly accepted by the user; classic baselines are named exceptions only | MovieLens / Amazon / KuaiRand, full-catalog ranking |
| Foundation models | Public benchmarks, same-budget controls, trainable core operators | WikiText-2, Tiny Shakespeare, multimodal and efficiency suites |
| LLM post-training | Executable preference/RL objective and public evaluation | GSM8K candidate, preference pairs, process traces |
| Agents | Auditable planning, tools, memory, or reflection trajectories | Deterministic tools and cross-episode mini-suites |
Search, ads, and recommendation: From classic ranking to generative recommendation¶
flowchart LR
W["Wide & Deep"] --> F["Feature interactions<br/>DeepFM / DCN-V2"]
W --> I["Interest modeling<br/>DIN / DIEN"]
F --> M["Multitask learning and modern ranking<br/>ESMM / MMoE / PLE"]
I --> S["Sequence modeling<br/>BST / SASRec"]
M --> H["Long sequences and large-scale ranking<br/>HSTU / RankMixer / HyFormer"]
S --> H
H --> G["TIGER / OneRec / CQ-SID"]
G --> R["RL / value-aware generation"]Coverage includes classic feature interactions, sequential interests, long-sequence ranking, dual-encoder retrieval, Semantic ID generation, LLM content and knowledge augmentation, sampling and distillation, RL, long-term value, and serving. The topic index separates retrieval, coarse/fine ranking, reranking and blending, content understanding, moderation and risk, and commercial ad decisions.
| Classic exception | Value | Current status |
|---|---|---|
| DeepFM | Common CTR baseline with shared embeddings across FM and deep branches | Implemented as a user-approved classic exception |
| YouTube DNN | Representative two-stage industrial retrieval | Retrieval and sampled softmax implemented as a user-approved classic exception |
| ESMM / MMoE / PLE | Main multitask CVR and shared-bottom lineage | Public multitask protocol and separate implementations |
No quantitative online A/B evidence meeting the industrial bar has been verified for RecoChain or DIG (2026). No placeholder adapters were created. GloRank, Dual-Rerank, OneRanker, RADAR, DualGR, MPFormer, HAP, OnePiece, IntSR, CDM, and CWM from the 2026-08-08 review are implemented. Netflix GenRec was added on 2026-08-13 alongside the existing JD.com paper with the same name.
Foundation models: Architecture, pretraining, multimodality, and efficiency¶
flowchart LR
T["Dense Transformer"] --> M["Switch Transformer MoE"]
T --> S["Sparse / Switch Attention"]
T --> A["Mamba selective SSM"]
T --> E["Engram / Memory Grafting"]
T --> P["RoPE / long context / MTP"]Coverage includes dense baselines, sparse MoE, Mamba, dynamic and sparse attention, conditional memory, long-context position encoding, MTP, quantization, RoPE/ALiBi, GQA, Hymba, MoBA, DoReMi and data mixtures, BLT, and classic multimodal methods. Trainable operators join the LLM/VLM evolve search space.
See foundation model research for browsing, classification, and evaluation.
- RoPE / ALiBi: Standalone same-budget long-context adapters and evolve operators exist.
- Chinchilla scaling: Requires curves over multiple compute and data budgets; one small-model run is not a substitute.
- New MoE routing and linear sequence models: Admit only when real operators, gradients, and public benchmarks are verifiable.
LLM post-training¶
The lineage covers PPO-RLHF, Constitutional AI, RRHF, RAFT, DPO, KTO, ORPO, IPO, SimPO, GRPO, RLOO, ReMax, DAPO, GSPO, LUSPO, CoBA-RL, Lightning OPD, Relay-OPD, GPRL, TCR, and CoRT. Four sequence objectives have real-tokenizer, free-generation, multi-seed paths. The P1 additions SLiC-HF, SteerLM, and SPIN cover sequence calibration, multi-attribute conditional SFT, and self-play. See post-training lineages for differences.
Paper retrieval, auditable objective mapping, and compositional genomes run through the unified multi-round controller. Further work is to scale public rollouts and unify reporting for reward, KL, length bias, and group normalization.
Agents¶
Classic coverage includes ReAct, MRKL, HuggingGPT, Toolformer, Tree of Thoughts, Reflexion, Self-Refine, LATS, ReWOO, AutoGen, WebGPT, SayCan, PAL, ART, Generative Agents, MemGPT, MetaGPT, CRITIC, SWE-agent, OpenHands, Agent Lightning, SEED, CAST, TurnOPD, Voyager, HiSkill, UniMem, and recent memory and planning methods. Software agents have real file editing and regression test paths. See agent lineages.
Memory, planner, tool, critic, policy, recovery, and capacity are composable genome axes. The same task suite measures success, token/tool costs, cross-episode reuse, and recovery from errors. Real code executors have a fair comparison matrix. Official SWE-bench Lite, the full ToolHop suite, and real browser environments remain deferred at the user's request; local fixtures are not official benchmark evidence.
Completed batches and new priorities¶
- Earlier P0 batch complete: Wide & Deep, DCN-V2, DIEN, BST, CS3, CQ-SID, Switch Transformer, Mamba, and Switch Attention. This does not establish completion of later audit findings.
- P1 infrastructure complete: Paper-constrained genomes for new LLM architectures, post-training, and agents join the unified multi-round controller.
- P1 classic papers complete: DeepFM, YouTube DNN, ESMM, MMoE, and PLE were implemented as user-approved classic exceptions.
- 2026-08-08 P0/P1 queue closed: See the all-domain gap audit for results and scope. New candidates do not reuse that day's queue.
Status as of 2026-08-24¶
The unified roadmap and TODOs maintain priorities, IDs, acceptance criteria, and status. Registered static P0/P1 items are closed. This page does not duplicate the queue. Further work comes from new discovery or deferred items reactivated by the user.