Recent paper scan: 2026-09-27¶
September 20–27 overlap across four tracks with manifest/ledger deduplication. arXiv paging succeeded without cache; Google Research and Meta AI pages timed out. No cross-source completeness or all-source watermark advance. Month-wide IDs catch delayed indexing; new means not closed here, not published this week. Institution keywords are not affiliations. Per-query paging capped at 100; expansion to 200 stopped because of rate limits.
| Track | Unique hits | Not closed | Interpretation |
|---|---|---|---|
| Search/ads/recommendation | 143 | 92 | Includes institution false positives and delayed older work |
| Foundation | 600 | 532 | Broad hits need date and topic review |
| Post-training | 313 | 257 | Meta MaD-RL found on official site, not this arXiv candidate set |
| Agent | 548 | 491 | Unclosed is not implementable |
Full-text priority queue¶
| Priority | Paper | Evidence | Recorded status |
|---|---|---|---|
| P0 | OneTrans-V2 (Sep 23, ByteDance) | §6.3 user-level 50/50 A/B; shared causal context, three-stage training, decision-conditioned retrieval and ranking-to-preranking distillation | Three-seed public proxy; MovieLens lacks transactions, ads and funnel |
| P0 | X-Rec (Sep 24, TikTok) | §5 two vertical rollouts, vertical engagement +4.1484%, global +0.0111%; spherical flow matching and late-interaction DiT | Public mechanism vs SID-AR; matched 200 steps, 20×1 full-catalog and CPU throughput; negative low-budget result |
| P0 | CMRec (Sep 24, Alibaba International) | §4.2 May 10–20 2026, 10% traffic: revenue +1.77%, orders +2.64%; shared codebook, dual constraints, contextual weights | Pending; private ads plus public six-region Amazon M2, check author code |
| P0 | AgentX-Model (Sep 24, Kuaishou) | §4.3 five A/B tests; dual agents, Reproduce/Follow-up/Composition/Diagnose and dependent replay | Six-node MovieLens-1M result-blind replay, three seeds; no private 473-node graph, online gate or LLM agent |
P1 review leads, not implemented¶
- MaD-RL (Meta) (Sep 24): KL/JS distribution matching vs GRPO collapse. Found after source timeout; verify equations, protocol, arXiv and code, not entropy-only substitution.
- DeltaS: gated linear-attention drift for video KV eviction; author code linked, real hybrid checkpoint and equal-budget video tests needed.
- KITE: KV-invariant dual tower and prefill/decode cost; verify small trainable model, not just FLOPs arithmetic.
- PACT, DCRL: critic alignment and policy–reward coupling; review equations, permissions and matched generative baseline.
- Trajectory-graph advantages: step credit needs nongold multi-turn trajectories and a trainable policy.
- IterSynth: role-separated deep search needs real retrieval, independent judge and cost comparison.
Flexible recommendation for individuals/groups is IMT Atlantique/Lab-STICC, not Google/Meta, with synthetic experiments. LLM user profiling uses production data but compares four strategies offline; no eligible A/B found. From Interests to Semantic IDs is Amazon Reviews offline, retained outside industrial. Light Heads, Music Rationales, UNIQUE, MuSeR and Meta layered engagement are overlap duplicates; unfinished private paths remain so.
Complete timed-out sources, paging above 100 and author-code checks. Industrial eligibility is not implementation. Preserve metadata, figures, independent metrics, three seeds and fair split/baselines; run CUDA paths on A100/A30. Promote P1 only after formulas, public data and budget checks; no registry-only or gold-trajectory capability claims.