Targeted new-announcement review, incomplete coverage: 2026-10-05¶
Keep the October 2 watermark and October 1 overlap. October 5 announcements in cs.LG, cs.CL, cs.AI and cs.IR include v1s actually submitted October 2. No DeepXiv. Recommendation API stalled at the first batch and was stopped; other tracks and Google/Meta did not finish full paging. coverage_complete=false; targeted reading is not high-recall coverage.
LESSER implements output-layer analytic gradients, shared two-sided Rademacher projection and round-robin selection. Three-seed CPU autodiff identity checks are L1, not LLM post-training or paper speed. First author at University of Pennsylvania; no author code link found.
| Lead | Mechanism | Required checks |
|---|---|---|
| Recursive Harness Self-Improvement | RSI/agent data, fixed verifier, real generator, same-cohort cost gate | Verify code/model and public tasks; no static self-evolution |
| RC-OPD | Student first-error repair and continuation | Real teacher/student correction trajectories, matched baseline and A100/A30 before CUDA claim |
| DepGPO | Terminal read/write graph and actual verifier reads | Auditable trace, no gold leakage, real policy updates |
| AIMS | Industrial recommendation + LLM; abstract offline Recall/NDCG | Check full text for A/B/full traffic; otherwise exclude from industrial |
Google/Meta remain priority; incomplete source review cannot establish no qualifying papers since October 1. These priorities are not terminal decisions. Complete full-text/source checks without treating delayed announcements or omissions as new publications.