Recent paper scan and closeout: 2026-10-03¶
October 1–3 with September 29–30 late-index overlap. Tracks: industrial and LLM applications, foundation/multimodal, post-training, agents/RSI/Jev. Official arXiv API/HTML/PDF, bylines and author repositories; no DeepXiv. Google/DeepMind/YouTube and Meta/Instagram prioritized by title then actual affiliation/full text. Missing abstract A/B is not exclusion. arXiv succeeded, but five official-source transport/title reconciliation failures leave coverage_complete false and all-source watermark 2026-10-02. Continue overlap from October 1.
| Priority | Track | Paper | Local scope |
|---|---|---|---|
| P0 | Search/LLM recommendation | RPTune | Query–item scoring, trimming and suffix orchestration |
| P1 | Agent recommendation | AgentWebRec | Temporal semantic retrieval, confidence gate, privacy audit |
| P0 | Foundation/Jev | LLM2Jev | Finite joint option probabilities and auxiliary KL anchor |
| P1 | Multimodal embeddings | Omni-Embed-Mini | Shared geometry from dense caption teacher |
| P1 | Multimodal efficiency | MWOP | Separate modality paths and FFN pruning |
| P1 | Optimizer | AF-Muon | Support-aware capped directions for tied vocabulary |
| P1 | Long context | REA | Separate persistent instructions and episodic memory |
| P1 | Multimodal inference | HAWK | Multi-layer hidden mixtures and shifted teacher |
| P1 | Compression | IrekoGPT | Shared projection basis and nested width |
| P0 | Post-training/agentic RL | Sharpening Tax | Matched-budget coverage-loss diagnostic |
| P1 | Post-training | CARM | Response mask against signed-ratio cancellation |
| P1 | Post-training | GMC-GRPO | Asynchronous group mass capping |
| P1 | Preferences | GAW-PO | Rejected-token gradient alignment weights |
| P1 | Agentic RL | SHARPO | Segment-level self-distilled credit |
| P1 | Video RL | TVRL | Reward-sensitivity token credit |
| P1 | Multimodal OPD | LEGO-OPD | Language prior × grounding likelihood |
| P1 | VLA OPD | DriftOPD | Reverse KL and future-potential decomposition |
| P0 | Coding agent | AutoCompact | In-place decision, summary and next-action correction |
| P1 | Long-horizon agent | Explicit Belief State | Unresolved requirements and trapping |
| P1 | Agent safety | PACE | Pre-effect authorization by capability/provenance |
| P1 | RSI/coding | RuleEvolve | Validation-only selection after rule mutation |
| P1 | Jev/agent reasoning | JevSpawn | Feedback-weighted posterior and branch retention |
| P1 | Memory | MemFit | Append-only turns and hybrid search |
Details record paper metadata, original figures, defining code, three-seed L1 and limits. Operator evidence is not paper-scale capability, private data or A/B. RPTune's Google first author and seven real merchants do not establish quantified online A/B; academic/Evolve orchestration only. AgentWebRec has a user-approved academic exception but no industrial A/B. No new qualifying Meta industrial item found; source failures prevent claiming none were missed.
Exclude vertical LLM/VLM applications without transferable mechanisms, benchmarks/surveys/data/security audits without executable methods, unrelated RL/control, industrial papers without accepted production evidence or named exceptions, and keyword false positives. No placeholder adapters. Reconcile official sources and unmatched titles before advancing beyond October 3.