Systematic all-domain gap audit (2026-08-08)

简体中文

This fixed two-dimensional review covers search, ads, recommendation, and LLM applications; foundation models; LLM post-training; and agents, including each domain's second-level subtopics. The machine ledger records every candidate as implemented, deferred, or rejected so discoveries are not lost between batches.

Audit method and limits

1. Build the covered set from adapters, detail pages, and method indexes. A name mention is not an implementation.
2. Search classic lineages, 2025, and 2026 through August 8 separately, then verify via institutions, citation chains, and official repositories.
3. Industrial papers require quantitative online A/B, explicit full-rollout evidence, or a named classic exception approved by the user. Offline metrics and unquantified deployment do not qualify automatically.
4. Foundation models, post-training, and agents need trainable or executable core mechanisms and public benchmarks. Pure analysis can be indexed without placeholder adapters.
5. Deferred means verified but not implemented or awaiting evidence/environment. Rejected means below the current bar, not missed.

Earlier 2026-08-08-p0-p1-closed-audit closed only its 15 selected papers and did not claim subtopic coverage. It did not establish no omissions. Tests now require audit records for 24 second-level subtopics in all-domain batches.

Findings and completion

All 38 P0 and 15 P1 items identified here were implemented. P0 consists of 19 reproduction adapters, 11 post-training objectives, and eight agent methods (fixed-seed metrics). P1 consists of eight reproduction adapters, three post-training objectives, and four agent methods (fixed-seed metrics).

Domain	Main gaps	P0	P1 and limits
Search, ads, recommendation, and LLM applications	Recent generative recommendation, reranking, long histories, heterogeneous coarse ranking, and content relevance	11	Classic long sequences and candidates awaiting evidence; no qualifying new moderation or risk candidate
Foundation models	Position encoding, GQA and hybrid models, data mixtures, tokenizer-free models, multimodality, and inference decoding	8	Five local or single-GPU P1 methods implemented
LLM post-training	RLAIF, process rewards, test-time RL, self-play curricula, and off-policy stability	11	Three P1 methods; no duplicate adapters for theory equivalent to existing objectives
Agents	Tool RL, deep research, skill consolidation, self-evolving RL, and public benchmarks	8	Four P1 methods or evaluators implemented

Counts refer to the explicit priority queue, not arXiv hits. The ledger remains authoritative for candidates, rejection rationale, and priority.

Search, ads, recommendation, and LLM applications

Implemented P0 items

Subtopic	Paper	Gap addressed	Reported online evidence
Generative recommendation	GloRank	Global Semantic ID action space, SFT and GRPO	Kuaishou, 7.8% of traffic for 14 days; watch time +0.095% and improvements in multiple engagement metrics
Reranking and blending	Dual-Rerank	Sequential distillation and latency-aware dual reranking	5% traffic for one month; Long-View +1.107%, significantly lower P99 and mean latency
Fine ranking	OneRanker	Unifies value-aware generation and ad ranking	WeChat Ads full rollout; GMV +1.34%
Retrieval	RADAR	Asynchronously uses a full ranker for next-request retrieval	About 2× Recall@200; online engagement +0.8%
Retrieval	DualGR	Long/short interest routing, constrained SIDs, and exposure-aware loss	Kuaishou views +0.527%, watch time +0.432%
Training and serving	MPFormer	Multitask retriever and resource sharing	Watch time +0.426%; training resources −60%, serving −66.7%
Coarse ranking	HAP	Heterogeneous examples and dynamic compute budget	Toutiao for nine months; duration +0.4%, active days +0.05%
Generative recommendation	OnePiece	Context engineering, blockwise latent reasoning, and progressive MTL	Shopee GMV/UU more than +2%, ad revenue +2.90%
Content understanding	IntSR	Joint search/recommendation intent and point-of-interest understanding	Amap GMV +9.34%, POI CTR +2.76%
Reranking and blending	CDM	Diversity as a controllable industrial blending objective	Kuaishou main app watch time +0.406%, clustering coefficient −0.957%

CWM also reports online MWT +2.9%, VV +2.5%, and CTR +0.3%. Its counterfactual long-term-value ranking merits a fair reproduction in the same P0 batch.

Implemented P1 and evidence threshold

• SIM, TWIN-V2, and CRSD are implemented. Preserve the online tables rather than treating “significant improvement” in an abstract as numerical evidence. Later candidates must pass the same bar.
• Moderation and risk: No public candidate satisfied both industrial scope and quantitative online A/B or full rollout. This is an audited subtopic with no qualifying candidates. Admitting public moderation benchmarks would require a separate change to the criteria.

Foundation models

P0: Fair local reproduction and evolve integration

Subtopic	Paper	Variable to test
Long context	RoPE / RoFormer	Absolute positions versus RoPE length extrapolation at matched parameter count
Long context	ALiBi	Train-short/test-long attention-bias comparison
Architecture	GQA	Joint quality, KV cache, and throughput metrics for MHA, MQA, and GQA
Architecture	Hymba	Parallel attention and SSM heads, not a configuration placeholder
Architecture	MoBA	Differentiable block routing and real sparse computation
Pretraining data	DoReMi	Proxy-loss-driven domain-weight updates
Pretraining data	Data Mixing Laws	Predict data mixtures from multiple low-budget domain curves, not one run
Architecture / tokenizer	Byte Latent Transformer	Entropy patching, byte encoders and decoders, and same-FLOP comparison

Implemented P1

• Multimodal anchors: CLIP and LLaVA.
• Inference: Speculative Decoding, AWQ, and Medusa.

LLM post-training

Many DPO/GRPO/OPD objective names hid genuine gaps in three lineages.

Implemented P0

Lineage	Paper	Missing mechanism
AI feedback	RLAIF	Principles and preference labels, AI preference model, and policy update in one loop
Process rewards	Let's Verify Step by Step	Fair step-level comparison of outcome and process rewards
Process rewards	Math-Shepherd	Automatic step supervision and a step verifier
Self-rewarding	Self-Rewarding Language Models	One model judges, generates preferences, and updates iteratively
Reasoning RL	LUFFY	Correction that preserves on-policy behavior with offline data
Test-time RL	TTRL	Reward estimation and online adaptation on unlabeled test examples
Self-play	Absolute Zero	Jointly generated tasks, verifier, and curriculum
Self-confidence reward	INTUITOR	Self-certainty reward without an external verifier
Stability	CISPO	Actual gradient path for token importance sampling and clipping
Self-play	SPIRAL	Language-game curriculum and two-role updates
Stability	ConSPO	Sequence consistency constraint

PPO, GAE, and DeepSeek-R1/GRPO are not missing adapters because existing implementations already cover their mechanisms. Technical reports or pure theory without new executable objectives are lineage notes, not additional algorithm counts.

Agents

Implemented P0

Subtopic	Paper	Missing mechanism
Deep research	DeepResearcher	Long search/browse/answer trajectory RL and citation reward
Tool RL	ReTool	RL curriculum for dynamic tool calls during reasoning
Tool RL	ToolRL	Multi-tool selection, argument generation, and execution-feedback training
Skill library	SAGE	RL-driven generation, updating, and reuse of skills
Memory and skills	MemSkill	Consolidating episodic memories into executable skills
Memory and skills	Memento-Skills	Experience extraction, skill retrieval, and cross-task transfer
Agentic RL	SEARL	Self-evolving trajectory pool and policy-improvement loop
Multi-agent systems	Agent0	Self-generated tasks and a multi-agent curriculum

P1 Agent-R1, CAMEL, ToolBench/ToolLLM, and GAIA are implemented. Each write-up separately states the deterministic mini-suite limits.

Executed results

1. Each industrial P0 records quantitative online evidence, upstream code status, fair local baselines, and negative results.
2. RoPE, ALiBi, GQA, Hymba, MoBA, BLT, DoReMi, and Data Mixing Laws are executable adapters; architecture and data-mixture operators entered evolve.
3. Eleven post-training methods use a common reward, KL, length, and seed protocol. TTRL and INTUITOR did not improve in the non-shifted mini-suite; this is a fidelity limit, not a hidden failure.
4. Eight agent methods ran 120 episodes with method-specific telemetry.
5. Future omission checks append a scope_kind: global ledger batch or clearly declare bounded scope. CI rejects missing subtopic records in all-domain audits.