November 4 update: Google/Meta priority and low-cost components¶
The latest observed arXiv announcement is November 3, 2026. This is a fixed eight-paper delivery, not an exhaustive full-text review of the web or announcement pool. First publication, overlap backfills, official-URL changes and revisions are separate.
Fixed scope¶
| Priority | Paper | Local delivery | Complete experiment not run |
|---|---|---|---|
| P0: Google | XLA from the Trenches | Review only; public tools distinguished from internal TPU systems | Multiple TPU generations, roughly 200,000 private graphs and internal tools; A100 is not a TPU substitute |
| P0: Meta collaboration | Strategist | CPU blueprint/AST, retrieval-score kernels, measurement cards and paired selection | No executor/meta-agent search or benchmark matrix |
| P0: industrial A/B | WeSearch | CPU cluster union, OPQ/ADC distance and fresh snapshots | No private billion-document retriever, trained FFN or production A/B |
| P1 | ProLand | CUDA thin-factor momentum, low-rank polar and Landing | No full LLM/model/rank matrix |
| P1 | GAVAGAI | CUDA capped OT with Nothing, detached alignment and prior | No full vision-language training; authors report four A100s and about 11 hours |
| P1 | SpecOPD | CUDA tree masks, predecessor readout and TV-prefix loss | No drafter/teacher checkpoints, rollouts or inference-speed claims |
| P1 | RLiD | CUDA legal-menu policy, calibration, PPO and credit-variance audit | No diffusion-LLM/environment RL |
| P1: backfill | AgentHeal | CPU external-evidence/provenance/compliance/abstention evaluator | No real browser actor or semantic judge |
All eight full experiments are not reproduced due to cost (review complete), without automatic retry. Seven components are L1 diagnostics; one paper is reference only. No complete capability reproduction or executable Evolve integration is claimed. Strategist's first-author institution remains Princeton despite Meta collaboration. AgentHeal was first published October 29 and is an overlap backfill; the other seven first appeared after the prior October 30 capture.
Actual execution and review¶
Preflight, frozen-source proof, raw CPU, raw CUDA, per-paper scope/cost.
CPU executes seven components across three seeds; the real A100 executes four CUDA components across the same seeds 42/43/44. Only small matrices and state contracts are used: no model/capability-data downloads or LLM/API calls. Each has its own metrics artifact, and four have separate GPU receipts. Runner time excludes research, transfer and startup; it is not full-experiment runtime.
Cross-review found that WeSearch reconstruction could expand individually bounded inputs into an enormous output. Combined output size is now checked before indexing, with a rejection test. CPU/A100 execution was repeated against repaired frozen source; earlier diagnostics are retained, not reused for changed code. GAVAGAI's eight OT iterations can leave row-sum residuals; these are preserved rather than hidden by post-hoc normalization.
Sources and deduplication¶
Source receipts and identities cover eight new listings, eight recent listings and 45 bounded institution pages. New listings contain 3,015 category entries and 2,168 distinct identities; recent listings contain 4,796 identities. These are recall counts, not qualified new papers or full-text reviews. The global review watermark remains October 15, 2026.
The parsed views contain 439 Google records, 144 Meta records across six pages and 266 distinct DeepMind URLs across nine pages. DeepMind's heading says 267; that discrepancy is explicit, not certified exhaustive. Priority review accounts for five added URLs: FabScore is an old title with a newly readable source but an unconfirmed original date; What Limits Us and EmbeddingGemma 2 are known identities. Subscription-entitlement consistency and chiplet manufacturing tests remain outside topic boundaries despite Google priority.
Primary receipts retain successful and failed fetches. Ten unfinished review leads do not count as this batch's implementations or review-DONE decisions. Unfinished reading is not relabelled high-cost nonreproduction. Future work can continue these leads without calling them another batch of today's new papers.
Learning and bilingual updates¶
Eight papers add basics, mechanism, comparison and implementation questions: 32 in total. Seven existing deep dives gain comparisons, analysis, self-tests and README-fingerprint-bound updates. Write-ups, Q&As, deep dives and this audit ship with reviewed English in the same PR.