跳转至

2026 historical paper scan and implementation through August 24

简体中文

The requested frozen window was expanded to January 1–August 24, 2026 across industrial search/ads/recommendation and LLM applications, foundation models, LLM post-training and agents. The multi-query matrix and cross-source configuration deduplicate canonical arXiv IDs. Industrial work requires quantified online A/B or explicit full-traffic deployment; Google/Meta affiliation checks receive first priority.

Reproducible scan

Run once per track:

PYTHONPATH=src python scripts/discover_papers.py \
  --track <recommendation|foundation-model|post-training|agent> \
  --date-from 2026-01-01 \
  --date-to 2026-08-24 \
  --page-size 50 \
  --maximum-results-per-query 200 \
  --cross-source-config configs/paper-discovery-sources.json \
  --output paper-candidates.json
Track Deduplicated candidates New Full-text review
Industrial search/ads/recommendation and LLM applications 716 627 237
Foundation models 1,319 1,303 42
LLM post-training 1,121 1,092 73
Agents 1,245 1,221 57

Broad queries retrieve many adjacent papers. Across four tracks there were 4,401 hits, 4,050 canonical papers, 3,906 not yet closed and 404 selected for full-text review. Hits do not establish eligibility.

B00: First verified and implemented batch

Track Paper Eligibility evidence Local result
Industrial recommendation OneModel Quantified A/B on three Xiaohongshu surfaces Scene-aware MovieLens-1M proxy did not beat shared sequential baseline; negative result retained
Foundation RARE Independently testable MoE steering/routing decoupling Route agreement 82.37%→100%, with target steering retained
Post-training GCPO Rollout-RL subspace diagnostics and constraints, composable in Evolve Arithmetic proxy final accuracy 63.28%, primary-direction overlap removed 99.56%
Agent AUSO Action-level JSD for skill internalization/exploration/use; author code public PlanBench mini-suite retains success and reduces context cost 9.09%
Industrial recommendation ClockRoPE Quantified YouTube/Google DeepMind A/B in full text MovieLens-100K NDCG@10 −2.30%; negative result retained
Industrial recommendation OneShot Meta/Instagram full rollout with online metrics MovieLens-100K NDCG@10 −7.70%; negative result retained
Industrial recommendation NEXT Meta multi-week A/B with about 100 million users MovieLens-1M NDCG@10 +20.33% single-seed diagnostic
Agent AgentX Kuaishou three-week production self-iteration loop PlanBench-mini: 12 completed loops and 108 knowledge reuses

Detail pages record paper metadata, author code status, local adapter/path, original figures, equations and distinct paper vs fixed local results. OneModel illustrates why implementing a mechanism and reproducing production lift are separate claims.

Google/Meta priority checks

Affiliation queries also found 2608.18531, 2608.15780 and 2608.15424. Full-text or topic review did not establish eligibility under the industrial online-evidence standard. Missing A/B in an abstract is not a reason for rejection.

Fixed follow-up batches

B01–B06 cover industrial generative recommendation, multimodal methods, search, ads, mixed ranking, long-term value and SIDs. B07 covers foundations, long context, KV cache and A/B/n tooling. B08–B09 cover OPD, rubrics, external rollouts and multi-reward RL. B10–B11 cover agentic RL, long-horizon credit, memory, tool planning and self-evolution. Exact IDs are frozen in the batch plan.

The other 331 full-text candidates were not rejected by default: they stay in the full-text review backlog until evidence supports rejection, deferral or promotion to a fixed batch. At this recorded stage, incremental-only scanning was to resume after B01–B11 and that backlog closed.