Continued overlap scan, sources incomplete: 2026-10-05¶
Following targeted October 5 review, overlap from 2026-10-01, last complete watermark 2026-10-02. Official API/full text/author repositories and Google Research, DeepMind, Meta pages; no DeepXiv.
| Source | Status | Limit |
|---|---|---|
| arXiv post-training/foundations | Batch paging done; some ID checks hit 429 | Window hits only, not affiliation coverage |
| arXiv agents/industrial | 429, paging incomplete | Not a four-track scan; retry |
| Google Research/Meta | Timeouts | Cannot rule out priority omissions |
| Google DeepMind | Readable, reconciliation incomplete | Recheck with other official sources |
Raw new counts include older month-ID and out-of-scope hits, not October 2 publications or qualified papers. Full text and evidence gate determine candidates. coverage_complete=false; next overlap October 1.
- Follow the Winners: FIFO replay, repeated minibatch top-\(K\) and KL-regularized CE projection; three-seed CPU bandit L1, not Search-R1/Sokoban or LLM capability.
- AdaStep: action-grouped variance attribution at anchor states and step credit; three-seed synthetic rewards, not genuine agent training or task comparison.
Both v1 submissions October 2, announced October 5. Paper results and local diagnostics are separate.
Meta AIMS has offline Recall/NDCG and industrial logs, but no randomized production A/B or full-traffic effect found in full text; not eligible for industrial implementation. Pivot-SD needs real LLaDA-style denoising and training, not sampling heuristics. RC-OPD has checkpoints/evaluations but pending training code; verify first-error repair, continuation and equal-budget controls. DepGPO needs terminal dependency graph, verifier reads, training and no gold leakage. Recursive Harness Self-Improvement needs a real generator, fixed verifier and same-cohort cost gate. KV² links author code; CUDA claims require sanitized A100/A30 receipt.
These are review leads, not implementations or qualified decisions. Retry incomplete arXiv and affiliation sources first.