Unified roadmap and TODOs

简体中文

2026-11-04: New announcements and bilingual learning material

Eight-paper fixed scope: Google's XLA/TPU paper is reference-only due to cost. Meta-collaborative Strategist and WeSearch are prioritized, alongside ProLand, GAVAGAI, SpecOPD, RLiD and AgentHeal: seven L1 components, four with actual A100 three-seed validation and three CPU-only. All expensive full experiments are not reproduced (review DONE); no model-capability or Evolve evidence is claimed. Thirty-two new questions, seven deep-dive updates and reviewed English versions ship together. The latest observed announcement is November 3: seven first publications since the prior capture and one overlap backfill. Unreviewed leads remain separate, and the global watermark is unchanged.

2026-10-30: No new source entries; overlap follow-up

Ten scoped deliverables and source differences: Prioritize Google collaborations CONCUR and MERLIN; also cover Google's April CROP, two Pinterest papers with production A/B evidence, and five foundation-model, post-training, and agent components. Six CUDA items ran on a real A100 with three seeds, and four are CPU-only. Expensive full experiments were not reproduced and their reviews are DONE. The learning materials add 40 questions across four levels per paper and nine topic deep dives. October 28 remains the latest official announcement. There were no new identities in eight arXiv new listings, the shared recent dates, or the bounded Google/Meta/DeepMind catalogs. These are gap closures, not ten new first releases on the same day. The global watermark did not move.

2026-10-29: New announcements, priority institutions, and learning material

Scoped deliverables and source snapshots: Eight low-cost L1 components, of which four CUDA paths ran on a real A100 with three seeds and four are CPU-only. Google's Sleep and DeepMind's HOPE are historical finds triggered by official sites and source code, respectively. PailitaoMMSearch2.0's full experiments are documented only and not reproduced due to cost; Guided-RLVR is separately closed as high-cost. The existing EmbeddingGemma2 entry gains the formal paper and Meta RPL gains official source code; neither counts as a new model. Added 38 paper questions and updated nine deep dives. October 28 is the actual latest announcement; October 27 recent listings were also checked. Catalog entries, code releases, and first publications are counted separately. The global watermark did not move.

Google EMNLP: Closeout of 25 historical leads

All 25 decisions and acceptance evidence: 14 low-cost L1 components, eight high-cost full experiments not reproduced (review DONE), and three rejections. Ten CUDA paths passed three-seed verification on a real A100 and four are CPU-only. These are not evidence of complete model reproduction, capability gains, or Evolve integration. Each of the 25 metadata leads now has a method, cost, and source disposition rather than remaining unread. Four additional source exceptions from the conference page were outside the approved scope. The global watermark did not move.

2026-10-26: New announcements and priority institutions

Fixed scope and source record: Eight low-cost L1 components prioritize Google OPST and the DeepMind collaboration DetectorScope, plus HGE, Frontier-OPD, MART-GRPO, DT-JEPA, EditJev, and UniAction. Frozen source was verified with three seeds on CPU or A100. Eight expensive full experiments were documented only and marked review DONE without automatic retry. Count announcement identities, code-triggered finds, and two historical Google catalog finds separately. Do not treat 1,180 category rows as new papers or claim full-text review of all 835 identities. The global review watermark remains October 15.

2026-10-25: Priority institutions and historical gaps

Three scoped deliverables and source record: Meta minimal-edit OPD takes priority, alongside Inside-JEPA and historical Google/DeepMind collaboration GroupDPO. All three L1 components ran on a real A100 with three seeds. Eight arXiv categories still showed October 23 with no new announcement batch. Conference acceptance listings are not same-day first releases. GenIA and GRIP full experiments are not reproduced due to cost (review DONE). AmbiGUI is recorded only as an official code release without a canonical paper identity. Google/Meta catalog failures remain unknown. The global watermark remains October 15; full-text closure of the whole window is not claimed.

2026-10-24: Follow-up within the same announcement batch

Four scoped deliverables and source record: Prioritize Google's PLE offline engineering-artifact protocol, alongside precise components for TTR-Compressor, UserArena OPTD, and MOER. All three CUDA components ran on a real A100 with three seeds. Eight arXiv categories still displayed October 23 and unchanged listings, so all four are gap closures, not first releases on the day. Google/Meta catalog refetch timeouts are unknown, not zero additions. Expensive full experiments were not reproduced (review DONE). The global watermark remains October 15.

2026-10-23: Priority institutions and recent announcements

Scoped deliverables and source record: Eight low-cost components: the historical Google Gemma behavior protocol, finite-policy reverse-KL distillation with Google collaborators, OpenAttention, MoEZip, AOPD, GIFT, PAIR, and Jev-DLC. The expensive OneLatent full experiment was not reproduced. Acceptance covers specified L1 formulas, state, or protocols, not end-to-end model capability or unified Evolve integration. CUDA delivery requires a passing real A100 receipt. The global watermark remains October 15; October 23 announcements and official catalog additions do not close full-text review of the whole window.

2026-10-21: Follow-up on the same day's announcements

Fixed six-paper scope and source differences: Five low-cost components plus historical documentation for Google's ResidencyRL. Four CUDA components were verified on a real A100 with three seeds. There were zero new official URLs or announcement identities versus the preceding batch, so these are newly cataloged finds, not six newly released papers. High-cost full experiments were not reproduced (review DONE). The global watermark remains 2026-10-15.

2026-10-21: Recent announcements and priority-institution additions

Fixed deliverables and source receipts: Eight low-cost components and four high-cost records. Meta's Baseline and Incremental/RPL and Google collaboration OO-Harness take priority. Four CUDA components completed real A100 verification with three seeds. All full experiments are explicitly not reproduced (review DONE), with no automatic training TODO. This does not close the entire October 16–21 window; see the next section for the common watermark.

Reviewed through 2026-10-15: Coverage and archived exceptions

The unified status defines the current review watermark and exceptions. Review closeout records four tracks, two channels, Google/Meta/DeepMind sources, cost dispositions, and exceptions. The strict source-exhaustion watermark is retired. Only review watermark, coverage, and archived exceptions are shown, without promising exhaustive web coverage. Implementation progress is tracked separately. The batch retrieved 3,742 deduplicated identities: 794 unregistered first releases in-window, 1,320 historical or overlapping, and 1,628 already recorded. There were also two non-arXiv official identities. The latest revisions of 11 existing papers were reviewed separately, not counted as new. Historical triage may use identity, date, and abstract scope only and must not be described as paper-by-paper full-text algorithm review. Official coverage: Google 28 pages/412 entries, Meta six pages/144 entries, DeepMind nine pages/265 entries. Current exceptions are archived without blocking. Subsequent incremental scans overlap from October 14 to their new cutoff. Separate historical finds, first releases, revision updates, and independent official identities instead of double-counting them as same-day new papers. Keep relevant historical leads as candidates; reopen archived exceptions only with new evidence, not by repeating every scan.

Expensive items explicitly not reproduced count as review DONE. They are not compute TODOs and do not trigger automatic retry, implementation, or Evolve. Reopen only if the user explicitly requests it after seeing the cost. This MR does not train or implement algorithms. The review watermark does not substitute for GPU, algorithm, or capability acceptance. A small number of affordable review candidates remain unimplemented and require a separate fidelity scope.

The fixed October 12 scope retains the prior 16 implementations and their actual acceptance results: CUDA mechanism/infrastructure verification, negative results, and unavailable original environments all remain visible. Paper-scale results, online gains, and formal capability comparisons were not reproduced.

PR #195 candidate-review snapshot (source follow-up continued above)

The paper-by-paper review records the latest batch disposition: accepted pending implementation, deferred after review, rejected/out of scope, and already recorded. Accepted papers still need faithful implementation and verification of their defining algorithms. The frozen pool contains 1,022 deduplicated papers and zero unread: 494 accepted pending implementation, 113 deferred, 367 rejected/out of scope, and 48 already recorded. The original scan below preserves the historical snapshot. Date-window queries and independent monthly-ID checks corrected the earlier query ceiling.

Google's 2026 official catalog had been fetched across 27 pages with 398 detail pages, but exact publication dates and missing identities were not yet verified. Meta's catalog ordering could not establish window exhaustion. The review watermark then remained 2026-10-02; see the official source receipt for those gaps.

Originally, 84 of Google's 398 catalog entries had direct arXiv identities dated before the window by ID month, while 314 lacked direct identities. None of the detail pages gave an exact first-public date usable for the window. Those records were checked individually above and exceptions made public. Meta pagination remains uncertain; a successful catalog fetch does not prove strict exhaustiveness. This batch did not scan October 8–10 or bulk-implement accepted candidates. Deliver those independently.

2026-10-07: Paper implementations and source recheck

The scan and implementation record retains four-track queries, the full deduplicated title pool, and initial screening of 48 Agent titles from the preceding scan. Broad matches are not newly qualified papers. Google/Meta source failures and query ceilings remain, so the watermark did not advance.

Item	Current status	Remaining acceptance
FrugalEvo	Standalone controller and three-seed A100 validation complete	Mean best sum of radii 2.32222 versus 2.40000 for the weak model. The cost curve shows no mean gain; full paper and unified Evolve are incomplete.
Sentry	Standalone controller and three-seed A100 validation complete	Local EM 29.17% versus baseline 41.67%; retain the negative result. The paper's four environments and unified Evolve are incomplete.
Google TEE-FL · P0	September 25 historical find; full-text online evidence passed	Verify open TEE stack and reproducible scope of public training. Ordinary GPUs cannot replace remote attestation.
Airbnb SIFT · P0	New October 6 paper; PDF A/B table verified	Public data for journeys, filters, and ordered capacity labels, plus actual training.
RC-OPD / KV² · P1	Methods reviewed, unimplemented	Actual repair-distillation training or actual KV reconstruction and compression, respectively.
Dynamic Expert Pruning / Pivot-SD / Recursive Harness · P1	Methods and feasibility reviewed, unimplemented	Real MoE, diffusion, and fixed-verifier environments with separate end-to-end verification.
Source closeout	Partial	Retry five institutional pages, scan capped queries, and read remaining titles in full. Do not call 1,010 broad matches reviewed papers.

This table supersedes earlier OCT06-FRU/SEN/DEP-MOE/SRC statuses below; it does not imply all old rows are closed.

2026-10-07: Architecture refactors MR-A and MR-B

Batch	Status	Scope and evidence
MR-A	Merged · PR #192	Experiment integrity and runtime contracts: experiment identity, resume validation, GPU execution loop.
MR-B	Merged · PR #193	Full acceptance record: lazy registries and minimal install, per-command CLI, page-level scan checkpoints, authoritative specifications, stable algorithm modules; 1,251 tests, A100 regression, and sanitized receipts.

MR-B adds no papers or evidence-level promotions. Fixing scanner state does not finish official-source retries or full-text screening below. The cross-source watermark remains 2026-10-02. Continue reviewing candidates under the existing overlap policy.

2026-10-06: Overlap follow-up scan

The scan record completed arXiv pagination for search/ads/recommendations and Agent, but Google Research/Meta official catalogs still timed out and 48 titles in the broad Agent window had not been read in full. The cross-source watermark remained 2026-10-02; do not start the next incremental scan on the PR date.

ID	Priority	Status	Deliverable or remaining acceptance
OCT06-DEP	P1	DONE · L1 mechanism	DepGPO: read/write/output dependencies, contributing writes and supporting reads, conserved token-weighted credit. Full agent training still needs a tracer, verifier, and actual policy updates.
OCT06-FRU	P1	TODO · Full text and author code checked	FrugalEvo: real strong/weak-model program search, shared prefix, and BA-AUC on public tasks under fixed costs; the area formula alone is insufficient.
OCT06-SEN	P1	TODO · Candidate	Sentry: condition-triggered recovery, memory write after success, isolated held-out tasks.
OCT06-DEP-MOE	P1	TODO · Candidate	Dynamic Expert Pruning: compare quality, memory, and latency with real MoE weights and predictors. CUDA claims need A100/A30 receipts.
OCT06-SRC	P0	TODO · Source closeout	Retry Google/Meta official catalogs and screen broad Agent results in full. Advance only after closure.

See the scan record for the remaining full-mechanism gates for RC-OPD, Recursive Harness Self-Improvement, Pivot-SD, and KV². The October 5 snapshot does not supersede this updated queue.

2026-10-05: Targeted recheck and next scan watermark

This batch rechecked an October 1 overlap window and added LESSER output-layer gradient data selection as an L1 CPU mechanism with three-seed identity checks. Batch arXiv API requests stalled and neither all four domains nor official Google/Meta sources closed; the discovery watermark remains 2026-10-02. Resume high-recall search and full-text review from October 1. RC-OPD, DepGPO, Recursive Harness Self-Improvement, and AIMS remain pending review; names or simple proxies are not implementations.

2026-10-04: Seven full-text follow-up papers

These close gaps in papers announced October 2, not new papers from October 4. All seven were delivered in one PR at L1 mechanism or control-protocol scope. This does not complete full-paper benchmarks, GPU training, or Evolve capability integration.

Priority	Paper	Implementation and verification in this batch	Full experiment not covered
P0	Jev lookahead	Environment-isolated lookahead, legal successor actions, existing provider contract; CPU diagnostic	Actual Jev with ALFWorld or robots
P0	T2SPO	Historical successes, real frozen TabPFN on CPU with three seeds, step credit and gradients	Frozen text encoder and multistep LLM task RL
P0	Weakest-Link	Absorbing prefixes, exact vocabulary-level immediate gradients, and actual small-policy optimization	Teacher/student LLM reasoning distillation and benchmarks
P1	CMP	Balanced randomized exposure, effect and SE, risk decisions	Real query experiments. Cross-query aggregation is not solved by the paper.
P1	TESS	Sequential training of two models, detached value labels, selector optimization	LLM data selection followed by controlled retraining
P1	RLCPR	Strict CDF buckets, posterior/length gates, reward	Real posterior collection and full RL
P1	Mingbird	Schema compression, JSON rescue, loop correction, actual artifact acceptance	Real local-model LRAB/τ², Windows GUI and voice

For runnable entry points, source status, and common boundaries, see the batch report. New CUDA paths need a separate A100/A30 gate; these CPU receipts cannot substitute.

2026-08-26: Platform P0

ID	Status	Deliverable
INFRA-003	DONE	SQLite experiment index, queries, Pareto frontiers, and static HTML dashboard spanning paper reproductions, Evolve, post-training, Agent, and multimodal work.
INFRA-004	DONE	paper.yaml for all adapters, deterministic generation/validation, and scaffolding that refuses to overwrite.
EVOLVE-005	DONE	Paper-operator registry and model, slot, dependency, conflict, and resource-budget validation, integrated into candidate planning filters.

With these three P0 items complete, the standard path for new papers is: add an adapter and public experiments, sync declarations, map executable Evolve operators, then register metrics in the shared experiment index. Platform maintenance no longer relies on copying catalog counts or manually synchronizing metadata across pages.

2026-08-27: Platform P1

ID	Status	Deliverable
INFRA-005	DONE · PR #128	Shared local/SSH/Slurm execution contract with logs, state, retries, hard timeouts, memory declarations, and cost ceilings.
EVAL-001	DONE · PR #128	Versioned fair evaluation across recommendations, foundation models, multimodal, post-training, and Agent; reject comparisons across incompatible protocols.
AUTO-001	DONE · PR #128	Auditable proposals from paper.yaml to hypotheses, operators, ablations, and search spaces, preserving provenance and requiring human confirmation.
EVAL-002	DONE · PR #128	Paired bootstrap, permutation tests, Holm correction, sequential seeds, and decisions under cost constraints.
MEMORY-001	DONE · PR #128	Context-exact negative-result knowledge base integrated into Evolve candidate filtering and per-round research memory.

See Auto Research platform P1 for usage, limits, and commands.

This page is the repository's single authoritative work queue. Domain lineage pages explain coverage and technical relationships, not a second TODO list that can drift. Each implementation MR must update status, acceptance evidence, and PR links here. Register newly discovered work before implementation.

Update baseline: 2026-09-30. The latest official-source retry and combined P0/P1 batch adds industrial P0 Snap GRP and eight P1 papers: LIFT, Triadic Linear Attention, OASIS, GRAFT, RIDE, PR-OPD, UserProxyBench, and UpliftMem. Date filters removed historical noise from the official pending queue, but undated and identity-unresolved titles remain, so the cross-source watermark does not advance.

The previous overlap scan and closeout completed official high-recall arXiv scanning across four domains from September 27–30 and implemented HELIX, STEPQuant, LeapQuant, Chinese-Jev, Dr. OPD, SIPO, ReMem, and Video-RSI. The high-recall pool and qualifying new papers are counted separately. Institutional entrances timed out during TLS, so only arXiv transport was complete; the cross-source watermark does not advance. The September 29 scan and RSI review covered four domains for September 20–29 and found industrial P0 Google/YouTube FLVM. Successful arXiv queries did not close official-source coverage: Google/Meta top-level entrances yielded zero IDs, so the discovery watermark does not advance. Earlier, the September 27 incremental audit ran paginated queries for four domains with overlap from September 20. The arXiv API recovered, but Google/Meta official publication pages timed out, so source exhaustiveness was not claimed and the watermark did not move. The previous scan verified four industrial P0s and one Meta evaluation P1 within the September 20–26 overlap. The queue below preserves completed and unfinished boundaries; placeholder adapters are not reproductions. The already merged PR #164 was not redone.

The September 28 expanded overlap scan followed arXiv API recovery, covering four domains and the top 50 per query, with original in-window releases separated from late indexing and historical unread material. Meta's top-level official page returned 500 and GitHub's institutional API 403, preventing cross-source advancement. New candidates pending acceptance: KuaFu (industrial P0), Meta Recursive OPD (P0), and ActKV, ToolSearcher, and Code Skills (P1). ActKV and ToolSearcher have testable mechanism kernels but lack real inference, original environments, and matched baselines. Code-Based Skills still needs NLE/MiniHack comparisons. None is a complete paper adapter; see the combined batch boundaries.

System One ran a fixed 200-item public slice and two real A100 Evolve rounds. Test and OOD were read only after selection, and the champion did not beat initialization. See the System One evaluation protocol for data hashes, metrics, and execution boundaries.

2026-09-26: New-paper execution queue

ID	Priority	Status	Work and completion criteria
SEP26-01	P0 · Google	DONE (this branch)	Light Heads: dynamic lightweight heads on a shared tower, reset and no-stop-gradient ablations; two explicit-rating tasks on MovieLens 100K, three seeds, isolated test, and retained negative results.
SEP26-02	P0 · Google	DONE (public-data concept demo)	YouTube Music rationale: Qwen2.5-7B/A100 asynchronous discovery profiles, Last.fm artist catalog/novelty/shared-tag checks, nomination-enhanced and explanation-only arms, CF fallback. Test on 100 users is 0.38 versus CF 0.40. No public rationale exposure data; online gains cannot be verified.
SEP26-03	P0	DONE (scaled mechanism)	UNIQUE: KuaiRand-Pure time-held-out exposures, planar quantization, early fusion, target-attention split, joint generation/ranking. Three-seed CTR-AUC, code recall, and ablations; retain negative results.
SEP26-04	P0	Partial (this branch)	MuSeR: trained multi-scale compression, multiple queries, candidate routing, and full-catalog retrieval controls on KuaiRand-Pure. Original ERNIE/BGE semantics, asynchronous caching, and production retrieval are incomplete; the adapter remains a concept demo.
SEP26-05	P1 · Meta	PROTOCOL (awaiting data)	Layered engagement evaluation: pre-exposure freeze checks, paired bootstrap, and CI-aware decisions implemented. The private scorer and paired A/B are unreleased, so 81.1% F1 is not reproduced.

2026-09-27: New-paper execution queue

This queue lists only industrial P0s with explicit online evidence, alongside the P1 follow-up shown below. X-Rec has a public-data core mechanism, but the full comparison remains incomplete. See the scan record for full-text evidence, domain P1 reviews, and exclusions.

ID	Priority	Status	Work and completion criteria
SEP27-01	P0 · ByteDance	Public proxy mechanism implemented · PR #169	OneTrans-V2: shared causal backbone, three-stage proxy losses, DCGR, and ranking-to-pre-ranking distillation, with per-stage three-seed metrics. MovieLens has no real transaction, advertising, or funnel labels; this is a concept demo.
SEP27-02	P0 · TikTok	Public core mechanism and fair control implemented · PR #169	X-Rec: SID-AR control with the same 200 updates, 20×1 full-catalog recall, and CPU generation throughput measured. Local low-budget results are negative; no production ANN or online reproduction.
SEP27-03	P0 · Alibaba	Unimplemented	CMRec: first-author affiliation verified as Alibaba International Digital Commerce Group. Verify author code and cross-domain data; ablate the shared semantic codebook, doubly constrained code mixing, and context-weighted loss across domains.
SEP27-04	P0	Public mechanism verification implemented	AgentX-Model: two-role proposal/execution review, four action categories, outcome-blind dependency replay, and a six-node, three-seed MovieLens-1M control ran. The private 473-node graph, LLM agents, and online gains were not reproduced, so this remains a concept demo.
SEP27-05	P1 · Meta	Small-model and three-seed A100 mechanism diagnostics ran; formal reproduction incomplete	MaD-RL: unit tests cover the four original distribution rewards. Qwen3-4B received equal-budget updates with L2, forward KL, and JSD on five-choice tasks for three seeds; validation and test did not distinguish rewards. Original math/code free generation, verifiable original figures, and paper-scale baselines are missing.

2026-09-28: Scaled mechanisms and pending verification

DeepXiv was unreachable and all four arXiv API track queries returned HTTP 406. This was targeted work using the registered September 27 queue, arXiv full text, and author pages, not a full September 28 scan. The discovery watermark did not move; rerun the overlap when access recovers.

ID	Priority	Status	Work and completion criteria
SEP28-01	P1 · Agent	L2.1 diagnostic ran; formal reproduction pending	GRAFT: experience trajectory graph, Bellman values, and Graph GAE. A gold-free tool environment with three-seed tabular policy controls found no gain. Still needs an LLM policy and paper task environment.
SEP28-02	P1 · Foundation models	Small-model compute and CUDA latency diagnostics ran; formal reproduction pending	KITE: two-stage dual towers, KV invariants, and three-seed WikiText-2 diagnostic. Approximate FLOP matching and small-model A100 prefill/decode ran without clear CUDA speedup. Original scales, long context, and throughput remain.
SEP28-03	P1 · Multimodal	L1 mechanism implemented; formal reproduction pending	DeltaS: state drift and bucketed KV eviction. Still needs actual hybrid-model states, public video, and matched-budget GPU verification.
SEP28-04	P1 · Post-training/Agent	Full text and data protocol pending	DCRL and IterSynth: no public low-budget path yet preserves the defining algorithms with fair evaluation. No name-only registration or gold-trace substitute.
SEP28-05	P0 · Alibaba	Data blocked	CMRec: Amazon-M2 unavailable. Do not fabricate cross-region data or metrics; retain SEP27-03.

2026-09-29: Combined Agent and post-training batch

The evidence boundaries and remaining acceptance are maintained together. “Mechanism” or “diagnostic” below does not mean a complete paper adapter.

ID	Priority	Current status	Next formal gate
SEP29-01	P1 · Agent inference efficiency	ActKV action-attention LRFU and dynamic-budget unit tests complete	Original hybrid-model KV, paged cache compaction, matched baseline, and A100/A30 end-to-end latency and memory.
SEP29-02	P1 · Agent retrieval	ToolSearcher first-discovery credit, gating, and masked policy loss unit tests complete	Pinned StableToolBench/tool index, actual grouped Qwen rollout, and isolated AppWorld evaluation.
SEP29-03	P1 · Code-Based Skills	Real MiniHack smoke tests in three modes and three seeds using author CodeHack and custom NLE	Compare primitive, skill, and mixed with the same policy and decision/environment-step budgets. The current random controller is not a paper-performance baseline.
SEP29-04	P0 · Meta post-training	Recursive OPSD low-budget official GSM8K three-arm, three-seed A100 comparison completed; negative	Exact OpenThoughts split, math benchmarks, original OPSD, and budget closer to the paper.
SEP29-05	P1 · Meta post-training	MaD-RL Qwen3-4B three rewards and three seeds in an A100 five-choice diagnostic; no difference	Original math/code free generation and fair baselines. Synthetic choices cannot be promoted to formal reproduction.
SEP29-06	P1 · Foundation models	KITE small-model approximate FLOP matching, local latency, and A100 CUDA prefill/decode microbenchmark complete; no clear speedup	Paper-scale tiers, long context, and throughput. A random-weight small-model microbenchmark is not the paper's speed claim.
SEP29-07	P1 · Agent/multimodal	GRAFT and DeltaS remain mechanism diagnostics	GRAFT needs an LLM policy and original tasks. DeltaS needs actual hybrid states and public-video GPU controls.
SEP29-08	P0 · Industrial recommendation	MuSeR remains a public KuaiRand tag proxy	Auditable product text and multimodal semantics plus cache refresh; public data lack these contents.
SEP29-09	P0 · Google/YouTube industrial recommendation	FLVM is a public KuaiRand three-seed concept demo. Like AP moves slightly, long viewing declines, sparse negative feedback varies, and semantics do not separate stably.	No YouTube satisfaction surveys, production pre-ranking, or online traffic. Do not claim reproduced online gains or generalize this as an Evolve gain operator.
SEP29-10	Cross-domain RSI	Recursive self-improvement topic separates executed mechanisms and pending candidates	RRSI/RSIBench-Data still need real tasks, multiround inheritance, fair costs, and OOD acceptance. The topic is not a new code domain.
SEP29-11	Discovery closeout	Zero IDs at official entrances and PDF misreads made explicit	Traverse detail pages or add feeds, review institutional full text, and close each historical pending candidate before advancing.
SEP29-12	P0 · KuaFu fair control	Independent three-seed A100 training on MRQA completed; no compression arm beats the equal-token control	Match training FLOPs, execute effective hallucination RL updates, use original task data and larger evaluation. This remains a public small-sample diagnostic.

2026-09-30: Combined P0/P1 batch

ID	Priority	Status	Deliverable and boundary
SEP30-01	P0 · Industrial recommendation	DONE (this branch)	HELIX three token streams, one-way user cache, and MPTF. TikTok full-traffic A/B is cited; public tensor diagnostics are not an online reproduction.
SEP30-02	P1 · Foundation models	DONE (this branch)	Defining STEPQuant and LeapQuant state quantization, three-seed diagnostics, and A100 CUDA receipts. Specialized-kernel throughput was not reproduced.
SEP30-03	P1 · System One	DONE (this branch)	Chinese-Jev candidate decision head and CE/RLCD in the System One index. No ten-million-example model training.
SEP30-04	P1 · Post-training	DONE (this branch)	Dr. OPD and SIPO token credit, three-seed diagnostics, and A100 CUDA receipts. No paper-scale rollout training.
SEP30-05	P1 · Agent/RSI	DONE (this branch)	ReMem temporal memory and Multi-Memory GRPO, plus Video-RSI Pareto acceptance. No real browser/video agent environment; not executable Evolve operators.
SEP30-06	Discovery closeout	PARTIAL	Four arXiv tracks transferred completely. Official institutional entrances timed out during TLS; RECAP lacks online A/B and SelfSearch/VACE need real environments. Watermark unchanged.
SEP30-07	P0 · Industrial recommendation	DONE (this branch)	Snap GRP block-wise SID, detached MHP, and mGRPO. Full-text online A/B is industrial evidence, not a local online reproduction.
SEP30-08	P1 · Foundation models	DONE (this branch)	LIFT top-k state feedback and Triadic third-order linear attention, both L1 mechanisms. No large-model training or GPU kernel performance claim.
SEP30-09	P1 · Post-training	DONE (this branch)	Defining selection, objective, or representation mechanisms for OASIS, GRAFT, RIDE, and PR-OPD with three-seed diagnostics. No paper-scale rollouts.
SEP30-10	P1 · Agent	DONE (this branch)	UserProxyBench UFS/leakage audit and UpliftMem paired uplift/EVSI, without gold answers or plans.
SEP30-11	Discovery closeout	PARTIAL (this branch)	Official entrances recovered and historical dated titles/pagination numerals no longer contaminate the queue. Retain undated or identity-unresolved items; watermark unchanged.

Priority and status

Label	Meaning
P0	Highest-priority newly found paper meeting hard gates, ahead of P1 on this page. In industrial research, Google and Meta come first.
P1	Capability work with a reliable implementation path and public data that can ship in batches.
EVIDENCE	Important direction lacking public evidence, data, or fair evaluation. Do not substitute placeholder adapters.
DEFERRED	Explicitly postponed by the user or dependent on a complex external environment. Keep out of execution until resumed.
DONE	Complete with code, docs, metrics, and test evidence.

Google (including DeepMind and YouTube) and Meta (including Instagram) are the only automatically prioritized institutions. Netflix, ByteDance, Alibaba, Kuaishou, and Pinterest still enter high-recall search and normal review. An institutional query hit only triggers full-text review; it does not replace affiliations on the paper's front page or evidence in its body.

Dynamic P0: Paper discovery closeout

ID	Status	Work	Completion criteria
DISC-001	DONE	Daily multiquery search across four domains, pagination, and canonical arXiv ID deduplication	GitHub Actions publishes four candidate artifacts.
DISC-002	DONE	Automatic diffs against manifest and historical ledger	JSON and Actions summaries distinguish new candidates, implementations, and audits.
DISC-003	DONE	Automatic priority alerts for new Google/Meta candidates	Only these two trigger warnings; Netflix and others remain standard candidates.
DISC-004	DONE	Cross-source retrieval	Reconcile dynamic detail pages to arXiv, snapshot/diff 906 static conference titles, and separate transport, pagination, and identity states. Official pending verification queue for this window is zero.
DISC-005	DONE · PR #113	Automatic write-back of final batch decisions	Require a final decision for every new candidate; strict audit reconciles pending artifacts and ledger.

Discover industrial candidates by institution and topic, then read the full PDF or HTML. Admit to implementation only with quantified online A/B or statistically significant full-traffic deployment explicitly accepted by the user. Missing A/B wording in the abstract is not a rejection reason.

2026-10-02: Combined P0/P1 batch

Priority	Domain	Papers/methods delivered	Acceptance status
P0	Industrial LLM/search/ads/recs	Gemini for Google, GEAR	Full-text online evidence, dedicated pages, executable L1 mechanisms.
P0	Post-training	DARS, Range-GRPO	Dependency reward and range GRPO invariants passed.
P0	Agent/RSI	ActiveSaddler, Safety Must Survive Self-Improvement, VeriHarness	Curriculum, verification, and rollback protocol tests passed.
P1	Foundation models/multimodal	TACO, VETO	CUDA mechanisms implemented; real A100 receipt is the merge gate.
P1	Agent/post-training	Mem++, DeFA, My FAULT, FloWright, Where-OPD	State, gradient, credit, and information-isolation contracts passed.
P1	Search/ads/recs	Effective Training Time, GrIS, REPAIR	Exceptions are in metadata and not included in industrial A/B conclusions.

See the 2026-10-02 scan closeout for the complete scan, primary evidence, and limits.

2026-10-03: Combined P0/P1 batch

Priority	Domain	Papers/methods delivered	Acceptance status
P0	Search/Jev	RPTune, LLM2Jev	Google full text prioritized; defining orchestration, decision distribution, and KL anchoring are executable.
P0	Post-training	Sharpening Tax	Fixed-budget coverage-loss diagnostics and three-seed L1 receipt.
P0	Agent	AutoCompact	Correction contract over compression decision, summary, and next action.
P1	Search/ads/recs and foundation models	AgentWebRec, Omni-Embed-Mini, MWOP, AF-Muon, REA, HAWK, IrekoGPT	Dedicated pages, key paper figures, mechanism code, and boundaries.
P1	Post-training	CARM, GMC-GRPO, GAW-PO, SHARPO, TVRL, LEGO-OPD, DriftOPD	Objective, mask, credit, and teacher-composition invariants passed.
P1	Agent/RSI/Jev	Explicit Belief State, PACE, RuleEvolve, JevSpawn, MemFit	State, permissions, validation selection, branch posterior, and append-only memory contracts passed.
Discovery closeout	Four domains	PARTIAL	arXiv succeeded; five official cross-source failures leave the 2026-10-02 watermark unchanged.

See the 2026-10-03 scan closeout for the complete scan, rejection boundaries, and next overlap start.

2026-10-04: Historical Meta official-source follow-up

Priority	Domain	Papers/methods delivered	Acceptance status
P0 · Meta	Agent retrieval	SIRA	DF gate and one-pass weighted BM25 run; no reproduction of the paper's LLM or large index.
P0 · Meta	AI research agents	AIRA2	Asynchronous steady-state scheduling and HCE signal isolation run; no claim of reproducing 8×H200 throughput.
P0 · Meta	Personalized agents	PAHF	Invariants cover pre-action clarification, explicit memory, and post-action drift correction.
P0 · Meta	RSI/open-ended evolution	HyperAgents	Parent selection and bounded archive step run; generated-code execution is prohibited.
Discovery closeout	Official-source reconciliation	PARTIAL	Historical gap closure, not new releases. Batch arXiv returned 429; watermark remains 2026-10-02.

See the 2026-10-04 Meta source follow-up for sources, dates, code status, and reproduction boundaries.

P1 execution queue

Follow the order in each table. Aim to make each ID an independently mergeable MR.

Foundation models and multimodal

ID	Status	Work	Minimum acceptance
FM-001	DONE · PR #113	Test-time compute, verifier, and dynamic reasoning budget	Pin SmolLM2. Report accuracy, tokens, latency, and call cost on GSM8K/arithmetic curves across budgets.
FM-002	DONE · PR #114	Multi-budget scaling-law infrastructure	Four default budget points for model size, data, and steps. Record actual parameter count, tokens seen, FLOP proxy, residuals, RMSE/R², and extrapolation limits.
MM-001	DONE · PR #115	Video multimodal	Pin SmolVLM2; use Video-MME-v2 Parquet/JSONL and MP4; resume per item and report three seeds, intervals, and subset scope.
MM-002	DONE · PR #115	Audio multimodal	Pin CLAP; use actual ESC-50/ESC-10 WAV; report zero-shot top-1/top-5 and text embedding cache fingerprint.
MM-003	DONE · PR #116	Embodied and large-scale multimodal post-training	Real SmolVLA/LeRobot training entry, data manifest, and explicit simulator versus hardware success boundaries.
FM-003	DONE · PR #137	TwinKV fixed-budget KV eviction repair	Exact formulas, actual Qwen3/Qwen2.5 KV runner, public WikiText-2 long context, matched quality/latency/memory, and A30 receipt.
MM-004	DONE · PR #137	PACE vision-token compression and extraction	APC/DDAE, default Qwen2.5-VL + RealWorldQA, actual SmolVLM2 + POPE verification, quality/token/latency/memory, and A30 receipt.
FM-004	DONE	SAS end-to-end sparse attention ranking	Actual selector training on WikiText-2, continuous log-gate gradients, Top-K routing, and isolated three-seed test. Triton performance is out of scope.

LLM post-training

ID	Status	Work	Minimum acceptance
PT-001	DONE · PR #113	Switch L2 to a downloadable pretrained causal LM	Pin SmolLM2; unrestricted GSM8K generation with three seeds.
PT-002	DONE · PR #115	Complete CoBA-RL teacher path	Pin Qwen2.5 teacher; pass@k and teacher caches, actual call rate, token costs, and before/after capability-boundary curves.
PT-003	DONE · PR #113	Public preference data	Pin UltraFeedback revision and MIT metadata; matched-budget DPO/ORPO with actual models.
PT-004	DONE · PR #113	GPU training integrity	Batch, gradient accumulation, mixed precision, safe checkpointing, and optimizer resume; retain CPU/Mac paths.

Agent

ID	Status	Work	Minimum acceptance
AG-001	DONE · PR #116	Connect Agent Lightning to a trainable LLM policy	SmolLM2 pairwise update, explicit transition credit, real test executor, and token/tool costs.
AG-002	DONE · PR #116	Fair matrix for real agent executors	Same repository fixtures and subprocess caps; three-seed controller-policy comparison.
AG-003	DONE · PR #131	Isolated Agent L2.1 evaluation and nine-axis Evolve	Remove guide/oracle; isolate train/validation/test. Six methods, five ablations, three-seed confidence intervals. Select champions on validation and run test only after the generation process ends.

Auto Research / Evolve

ID	Status	Work	Minimum acceptance
EV-001	DONE · PR #113	Add FM-001 inference-budget operator to unified genome	reasoning-checkpoint searches sampling budget, self-consistency verifier, and stop thresholds, with per-generation reports.
EV-002	DONE · PR #116	Add PT-001 through PT-004 to the post-training genome	Data/objective/teacher/rollout/accumulation/precision become composable inherited axes and appear in generation reports.
EV-003	DONE · PR #117	Add AG-001/AG-002 to the Agent genome	Composable memory/planner/tool/critic/policy/recovery; measurable cross-episode reuse and failure recovery.
EV-004	DONE · PR #117	Evolve for GenRec-style generative recommendation	Real full-catalog MovieLens-1M head, context/reward/distillation controls, and shared ID-catalog baseline. This does not grant Netflix paper priority.
INFRA-001	DONE · PR #111	GPU dependency safeguards	pip dry run prevents silent PyTorch replacement. Linux CPU contract tests and existing A30 critical-path regression coverage remain.
INFRA-002	DONE · PR #117	Multi-seed promotion for key methods	Shared three seeds, intervals, per-seed failures, and resume across recommendation/foundation adapters, post-training, and Agent.
EV-005	DONE · PR #137	PACE and TwinKV checkpoint/architecture operators	Trace paper IDs to operators. PACE enters VLM checkpoint recipes; TwinKV runs in micro-LLM attention and shares repair with actual checkpoints.
EV-006	DONE	EvoSkillRec typed skill genome	The controller type-checks, runs, validates, promotes, and reuses candidates. RankMixer Evolve trains and evaluates rankmixer_evoskill; this is not a registry-only label.

Awaiting public evidence: No placeholder implementations

ID	Status	Gap	Resume when
EVD-001	EVIDENCE	Moderation, cheating, fraud, ad compliance, and risk control	Public labeled data and precision/recall, false-positive, and guardrail protocols.
EVD-002	EVIDENCE	Private large-scale auction and conversion logs	Public replacement data and fair offline protocols without business-data leakage.
EVD-003	EVIDENCE	Public end-to-end LLM recommendation product reproduction	Legal public data for users, catalog, generation, ranking, and feedback.
EVD-004	EVIDENCE	AIGQ, RaG, RoleGen, LCU	Public rewards, video/user feedback or conversion trajectories, and clear licenses.
EVD-005	EVIDENCE	Industrial candidates including RecoChain and DIG	Verify quantified online A/B or user-approved full-traffic deployment in full text.

Deferred by the user

ID	Status	Work	Resume when
DEF-001	DEFERRED	Official SWE-bench Lite	Official containers, image cache, execution budget, and long CI/development windows are available.
DEF-002	DEFERRED	Full official ToolHop	Confirm data/evaluation dependencies, model call budget, and reproducible runner.
DEF-003	DEFERRED	Real browser-agent environment	Isolated sandbox, credential and network policies, and failure replay.
DEF-004	DEFERRED	OneLA fused GPU large-beam decode	Run defining kernels on A100/A30 and submit memory, latency, and correctness receipt.
DEF-005	DEFERRED	CanvasAnneal diffusion LM curriculum RL	Pin public DLM/teacher, training budget, and fair MATH/tool-use protocol.
DEF-006	DEFERRED	GAUGE grounded-agent evaluation	Public transcripts and blind ratings become available, with ranking and construct-validity protocols.
DEF-007	DEFERRED	AMDKernelVault ROCm/HIP	AMD CDNA/ROCm environment and passing kernel correctness/latency gate.

Update contract for future MRs

1. Register ID, priority, status, and acceptance criteria here before starting. Insert dynamic P0 above P1.
2. Paper implementations must meet common metadata, Chinese explanation, key paper figure, code, metrics, and test requirements.
3. For Evolve integration, state whether an operator comes from an implemented paper, a live search, or a new combination hypothesis. Keep these distinct.
4. Mark complete work DONE with PR and key evidence. Assign a separate ID to unfinished work.
5. Do not duplicate TODO tables. Lineage pages link here so multiple lists of remaining work cannot contradict one another.