October overlap follow-up and DepGPO implementation: 2026-10-06¶
Continue incomplete October 5 scan with 2026-10-01 overlap and last complete watermark 2026-10-02. Official arXiv API/full text and Google/Meta sources, no DeepXiv.
| Track | Coverage | Remaining boundary |
|---|---|---|
| Industrial arXiv | Paging succeeded; two unreviewed titles in window, both out of scope | Three Google Research and two Meta pages timed out |
| Agent arXiv | Paging succeeded; 48 unreviewed window hits | Title recall includes robotics, applications, safety and evaluations; not 48 implementations or full reviews |
| Foundation/post-training | Prior completed paging plus targeted official full text | No renewed full affiliation reconciliation |
| Google DeepMind/Google/Meta | DeepMind readable; directory timeouts for Google Research/Meta with some browser fallback | Targeted checks only; no overall advance |
Unreviewed counts include historical and late-index work. Use actual v1 submission dates, not time of a successful 429 retry. Local scan receipts carry machine details; page records review conclusions.
DepGPO, v1 October 2, implements structured read/write traces, verifier-related writes, supporting reads and step/token advantage normalization. Hand traces and property tests pass; L1 only, no terminal tracer, Qwen policy training or Terminal-Bench lift. Meta AIMS remains ineligible industrially: offline logs only, no verified production A/B/full traffic.
- FrugalEvo, author code: budgets, strong-model planning/weak coding, BA-AUC and shared-prefix cache suit Evolve. Need fixed public task/call costs, not formula-only.
- Sentry: conditional failure recovery, store only successful recoveries. Need real failure detection, isolated tasks and no reward leakage.
- Dynamic Expert Pruning: prompt-conditioned MoE pruning needs checkpoint, predictor training and quality/memory/latency comparison; CUDA requires A100/A30.
- RC-OPD, Recursive Harness Self-Improvement, Pivot-SD, KV²: retain full mechanism and data/GPU gates.
These leads are not implemented. Continue October 1 overlap and official-source checks; new-paper increments start from the last reviewed date, not PR date.