Review through October 7, 2026: coverage and archived exceptions

简体中文

This verification finished on 2026-10-10 and covers October 1–7 only, not October 8–10.

The review watermark is 2026-10-07. The strict source-exhaustion watermark was retired as a separate measure and blocker. This page records actual coverage and limits; it does not claim the entire web was exhausted or that candidate algorithms were implemented.

Central status · Review standard

Scope

• Google: resolved 398 detail-page records from 27 pages, including individual rechecks of 314 previously unidentified records.
• Meta: expanded to six pages and 144 unique records. Six in-window math studies were checked for scope; other official dates precede the window.
• Review completed for the 1,022 deduplicated candidates from the original four tracks and two query channels. Accepted entries still require separate implementation.

Final source dispositions

Source	Disposition	Count
Google	Public metadata/full-text exception	31
Google	Public before window	156
Google	Existing record	1
Google	Outside research scope	210
Meta	Public before window	138
Meta	Outside research scope	6

Archived exceptions and reopening criteria

See the exception archive for all 32: 31 Google records and one Meta pagination limit. Archived exceptions do not block current or later watermarks, count toward accepted or implemented totals, or trigger repeated retries each round. Reopen only on changed official material or new evidence from independent discovery. Each record retains actual attempts and reopening criteria. Google/Meta industrial papers still need quantified live A/B or clear full-rollout evidence. Missing A/B in an abstract is not a rejection reason.

Meta catalog pagination exception

We verified dates individually on the first six pages. The newest dates on pages four through six were all in 2025. Older records are interleaved and there is no stated total, so pagination and dates do not establish exhaustion. The user approved this bounded exception as non-blocking.

Google exceptions, item by item

• Agentic Trend-to-Knowledge: A Methodology for Automating Case Deflection through Proactive Content Surfacing in Enterprise Employee Support: Connecting trend discovery to RAG queries is relevant to agents, but the official Download points to a defensive publication with a different title, not confirmed as the same manuscript. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: readable_html; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: returned similar methods under different titles; record dated 2026-01-08, identity still uncertain. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• The Glossary-Grounded Universal Queue: An Architecture for Enabling Follow-the-Sun Support in Global People Operations: Glossary-enhanced enterprise translation is relevant to LLM applications. The Download title differs from the official entry; topical similarity does not establish identity. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: readable_html; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: returned Glossary-Augmented Generative AI for Enterprise Translation, dated 2026-01-07; original-title relationship not verified. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• Self-Evolving Systems: Moving Beyond Deterministic Interfaces to Adaptive Generative Interfaces: Runtime feature extension in adaptive generative interfaces concerns self-evolution. Separate UI usability from algorithmic capability. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• XProf: An Open, Scalable and Extensible Profiling System for the Modern ML Stack: XProf profiles the full ML stack and could support training or evolve infrastructure, not a new model algorithm. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Securing Multi-Agent Systems: An Empirical Analysis of Security Prompt Hardening and Residual Risks: Multi-agent prompt hardening and residual risk are relevant. No Download on the official detail page, and title search found only that entry; exact date and public full text remain unresolved. Attempts: official_detail: cached_official_detail_read; local_title_match: no matching record found; frozen_register_and_worktree_manifest_match: no matching record found; official_download: no_download_link_in_cached_detail; primary_web_identity_search: cross-checked official, arXiv, and publisher sources; did not use third-party crawl time as first publication. Reopen: Obtain publisher first-publication date, verifiable arXiv identity, or official public version; then check window membership and review the text, without assuming acceptance.
• The Synthetic Gap: Automating Forensic Investigation of "AI Slop" with the Scaled Abuse Forensics Examiner (SAFE): SAFE's multi-agent synthetic-media forensics is relevant to moderation. Check production evidence separately; shorter investigations do not meet the online A/B gate automatically. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Geo-Contextual AI Concierge: Minimizing Internal Support Friction Through Proactive Knowledge Synthesis: Geo-contextual SENTINEL/RAG concierge is relevant to agents. Official PDF was unavailable; independent identity and first-publication date unverified. Attempts: official_detail: cached_official_detail_read; local_title_match: no matching record found; frozen_register_and_worktree_manifest_match: no matching record found; official_download: failed; primary_web_identity_search: cross-checked official, arXiv, and publisher sources; did not use third-party crawl time as first publication. Reopen: Obtain publisher first-publication date, verifiable arXiv identity, or official public version; then check window membership and review the text, without assuming acceptance.
• What’s on My Network? Using Large Language Models to Identify Real-World IoT Devices at Scale: Curriculum fine-tuning of LLaMA for IoT identification is an LLM application, but no primary source verified first publication and public implementation identity. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: no_download_link; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: found Google and secondary materials only, no verified manuscript date. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• AIRS: Scaling Live Inference in Resource Constrained Environments: Google Search autorater TPU scheduling is serving infrastructure. Internal scale is not an online search, ads, or recommendation effectiveness result. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Impact of User Tonalities on Coding AI Model Performance and Behavior: User tone and coding LLM correctness are relevant agent diagnostics. A challenge page blocked the publisher download and exact date remains unverified. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: challenge page, not paper full text; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: publisher link exists but download returned a verification page. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• FabScore: Fine-Grained Evaluation of Fabrications in Automated AI Research: FabScore audits research agents through extraction, static analysis, execution, and judgment. OpenReview forum/PDF verification and ICML page errors prevented date confirmation. Attempts: official_detail: cached_official_detail_read; local_title_match: no matching record found; frozen_register_and_worktree_manifest_match: no matching record found; official_download: no_download_link_in_cached_detail; primary_web_identity_search: cross-checked official, arXiv, and publisher sources; did not use third-party crawl time as first publication; openreview_forum: browser_verification_challenge; openreview_pdf: browser_verification_challenge; icml_poster: internal_error. Reopen: Obtain publisher first-publication date, verifiable arXiv identity, or official public version; then check window membership and review the text, without assuming acceptance.
• Mining Naturally Romanized Seed Corpora without Romanizations: Cross-language mining of naturally romanized corpora is a historical foundation-model data candidate. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse: A LoRA-Enabled Multimodal Defense System: S-CTS combines account clustering with LoRA/APO multimodal classification against coordinated abuse. The four-page official paper was read. §V compares live-stream operations with a six-month baseline, not randomized A/B. Exact date and online admission remain unverified. Attempts: official_detail: cached_official_detail_read; local_title_match: no matching record found; frozen_register_and_worktree_manifest_match: no matching record found; official_download: no_download_link_in_cached_detail; primary_web_identity_search: cross-checked official, arXiv, and publisher sources; did not use third-party crawl time as first publication; official_pdf_fallback: read all four pages; no publication date; §III–IV architecture, §V operating results, Table I classification metrics. Reopen: Obtain publisher first-publication date, verifiable arXiv identity, or official public version; then check window membership and review the text, without assuming acceptance.
• Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse: A LoRA-Enabled Multimodal Defense System: Synthetic clusters, LoRA, and automatic prompt optimization detect textual spam. An official manuscript was found, but exact first release is unverified and the abstract does not satisfy the industrial gate. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: no_download_link; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: found official PDF 1039291; no credible first-publication date. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse: A LoRA-Enabled Multimodal Defense System: LoRA/APO and account links support synthetic-media moderation. Full text and first publication need verification; abstract operational statistics do not pass the industrial gate. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; download_link_check: no Download anchor on cached official detail; web_search: no exact first-publication date established; official_pdf_search: Official PDF identified and method excerpt inspected; exact public-release date not established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Meridian GeoX: An Open-Source Framework for Precision and Efficiency in Geo Experiments: GeoX is open-source causal advertising experiment infrastructure and a historical evaluation candidate, not a ranking model with qualifying online evidence. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: primary-source identity/date found; primary_identity_date: Manuscript first-page date, not proven earliest public release. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Lessons learned from building an agent to convert the R Forecast Package into a JAX api: A multimodel skill pipeline for moving forecasting code from R to JAX is relevant to coding agents/RSI. No confirmed first date or verifiable open repo; an abstract release claim is not verification. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: no_download_link; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: no dated original manuscript beyond the official listing. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study: AMIE prospective clinical feasibility is an agent application. No verified primary date; a secondary index suggests October 8, so neither count it through October 7 nor mislabel it as historical. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: unavailable; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: secondary index suggests 2026-10-08; Lancet/PubMed primary entries inaccessible; exact Europe PMC ID query returned no record. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• A Declarative Multi-Agent Framework for Reducing On-Call Operational Burden: Declarative multi-agent production on-call diagnosis constrains tools and observability. Official PDF readable without first date; production time savings do not establish publicly reproducible capability. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: readable_pdf; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: official PDF undated; no arXiv manuscript located. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• QBAT: Model-based Query Budget Autotuner for Clustering-based Approximate Nearest Neighbor Search: QBAT predicts ANN search budgets from query features and compares GBDT against AlphaEvolve formulas, relevant to retrieval and RSI. Identity and DOI verified; exact first publication unlocated. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: readable_pdf; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: verified DOI 10.14778/3836663.3836675 and official VLDB program; first date unconfirmed. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• Holistic Latent Diffusion Acceleration: Unifying Spatial, Temporal, and Architectural Efficiency: Jointly distills spatial tokenizer, student backbone, and time trajectory, a multimodal generation efficiency candidate. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: failed; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior: GMD distills a frozen generative prior for unpaired image restoration, a multimodal method. Official publishing entry exists but exact first date was inaccessible. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: challenge page, not paper full text; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: found ECCV/Springer entry; verification challenge; original date not located. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• Cold Start Latency of Quantized Small Language Models on Serverless CPU Infrastructure: Benchmarks quantized small LM CPU cold starts and quality trade-offs; candidate foundation inference evaluation. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• A Unified Protocol Architecture for Agentic Interoperability: Quantitative Evaluation of Standardized Context Exchange in the N + M Economy: Simulated quantitative interoperability comparison of MCP-like protocols is an agent systems candidate; evaluation definitions and resources need checking. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Structural and Programmatic Prompting in Financial Artificial Intelligence: Structured prompts and programmatic financial reasoning are agent-relevant. Only an official abstract; manuscript identity, date, and public evaluation unlocated. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: no_download_link; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: no matching manuscript or date located. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• PIPE: Personalized Image-generation via Preference Encoding: PIPE learns user preference embeddings and uses IP-Adapter and Diffusion-DPO for personalized generation, a multimodal post-training method. OpenReview verification blocks access; venue year is not first publication. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: unavailable; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: confirmed OpenReview G5GipElVVe; PDF and forum both blocked by verification challenges. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• AI Driven Verification Architectures - Scaling Formal Sign-off of High Bandwidth Data Crossbars via Generative AI and Required-Proof-Depth (RPD) Synthesis: The official site shows only title and research area, no abstract/download. Possibly generative formal verification, so not excluded merely for hardware subject matter. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; download_link_check: no Download anchor on cached official detail; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Beyond Real or Fake: A Dual-Channel Authenticity and Reasoning Protocol for Photographic Assessment: Dual camera and semantic evidence streams with frozen VLM iteration are a multimodal content candidate. Offline F1 does not meet industrial deployment criteria. The official entry and local title matches were checked, but exact first publication or identifiable source remained insufficiently confirmed. Excluded from this round's accepted count and retained for recheck. Attempts: official_detail: cached detail title and available abstract read; scope triaged; repository_title_match: no exact normalized-title match; official_download: retrieved; web_search: no exact first-publication date established. Reopen: Reverify and review full text once an exact official release date, accessible manuscript, or arXiv identity is supplied.
• Dynamic Learning and Optimal Advertising Mechanism for LLM Platforms: Stochastic control for dynamic LLM advertising is relevant to ads optimization. No manuscript or date identified; the abstract does not confirm quantified online experiments. Attempts: official_detail_scope_read: cached official abstract read; no abstract retained in output; official_download: no_download_link; manifest_and_frozen_title_match: no matching record found; frozen_register_exact_normalized_title_match: no matching record found; official_or_arxiv_identity_search: no public manuscript with verifiable identity and date located. Reopen: Obtain primary manuscript identity and exact first publication. If within 2026-10-01 to 10-07, review full text, source release, and evaluation rather than accepting automatically.
• From Unfinished to Done: Bridging the MDE Implementation Gap with Constrained LLMs: Agentic Gap Filling applies protected TODOs, a review jury, and independent testing agents to MDE. ACM DOI returned 403, no matching arXiv identity, and first date unknown. Attempts: official_detail: cached_official_detail_read; local_title_match: no matching record found; frozen_register_and_worktree_manifest_match: no matching record found; official_download: failed; primary_web_identity_search: cross-checked official, arXiv, and publisher sources; did not use third-party crawl time as first publication. Reopen: Obtain publisher first-publication date, verifiable arXiv identity, or official public version; then check window membership and review the text, without assuming acceptance.
• ContextFidelity-Bench: A Three-Paradigm Framework for Evaluating Multi-Turn Hallucination Behavior in Language Models: A multi-turn contextual faithfulness/hallucination benchmark is relevant. The official PDF was unavailable and title search did not establish a unique preprint or date. Attempts: official_detail: cached_official_detail_read; local_title_match: no matching record found; frozen_register_and_worktree_manifest_match: no matching record found; official_download: failed; primary_web_identity_search: cross-checked official, arXiv, and publisher sources; did not use third-party crawl time as first publication. Reopen: Obtain publisher first-publication date, verifiable arXiv identity, or official public version; then check window membership and review the text, without assuming acceptance.

Relevant historical leads

These public releases predate the window and do not block the watermark. This says nothing about implementation status or research value. See the Google machine-readable inventory for identities, dates, original sources, scope reasons, and possible relevance.

• Towards conversational artificial intelligence for disease management: Dual-agent clinical dialogue and management reasoning with RxQA are agent-related. Historical candidate; simulated OSCE is not deployment capability.
• On-the-Fly OVD Adaptation with FLAME: Few-shot Localization via Active Marginal-Samples Exploration: FLAME adapts open-vocabulary detection; retain as a historical multimodal candidate.
• ADCanvas: Accessible and Conversational Audio Description Authoring for Blind and Low Vision Creators: Conversational VQA, scripting, and accessibility audio descriptions use agents. A 12-person study does not replace capability benchmarks.
• Type-Aware Ranking of Urban Similarity from Aerial Imagery: Type-conditioned ViT, CLIP semantics, and adaptive contrastive loss for urban visual retrieval. The conference predates the window, not an exact first-release date.
• Multi-Agent System for Cognitive Work Automation Using an Adversarial Review Panel: Adversarial-panel multi-agent technical disclosure. Historical architecture lead; without controls it is not a mature algorithm.
• Prompt-Level Distillation: A Non-Parametric Alternative to Model Fine-Tuning for Efficient Reasoning: Compresses teacher reasoning rules into student system prompts; candidate prompt distillation and evolve.
• ALPS: Activation-Based Length Prediction for Intelligent LLM Inference Scheduling: Uses hidden activations to predict length and budget inference early. Verified March release; not new in this window.
• KVCIS: Activation-Based Token Importance Prediction for Intelligent KV-Cache Compression: Uses intermediate activations to estimate token importance and compress KV. Historical foundation inference-efficiency candidate.
• Reasoning-Driven Synthetic Data Generation and Evaluation: Simula uses reasoning agents for seedless synthetic data generation and evaluation; spans foundation data and agents. March arXiv manuscript located.
• Vibe Coding XR: Accelerating AI + XR Prototyping with XR Blocks and Gemini: Open workflow from natural language to WebXR with VCXR60 trials, a historical coding-agent lead.
• AgentHands: Generating Interactive Hands Gestures for Spatially Grounded Agent Conversations in XR: Synchronizes gestures and spatial events into XR agent interaction signals. CHI date predates the window.
• AASE: Activation-Based AI Safety Enforcement via Lightweight Probes: Lightweight activation probes for safety, action gating, and policy detection. Candidate safety evaluation; verify threshold calibration.
• Expert evaluation of LLM world models: A high-Tc superconductivity case study: Expert science QA and multimodal literature RAG evaluation are historical public Agent/multimodal benchmark candidates.
• System for Validating Medical Software Algorithms With Closed-Loop Adversarial Persona Generation: Adversarial patient personas support closed-loop healthcare software verification. Multi-agent testing, verified January release.
• Gaze Target Estimation Anywhere with Concepts: GazeAnywhere/official AnyGaze estimates gaze targets by concept. arXiv v1 predates the window; earliest conference version needs checking.
• A Multi-Layered Framework for Behavioral Governance of Non-Deterministic AI Agents: Technical disclosure on layered identity, memory, and safety governance for agents. Historical lead, not capability implementation.
• Probabilistic Governance and Telemetry Framework for Autonomous Agents: Policy and telemetry constrain autonomous agents. Defensive publication appeared in March.
• Detecting Safety Training Modification in Language Models via Activation Analysis: Activation-space detection of safety-training modification is a foundation safety candidate; retain blind spots for behavioral fine-tuning.
• Framework for Multi-AI Orchestration via Task Arbitrage and Byzantine Consensus: Technical disclosure of task arbitration and Byzantine consensus over heterogeneous models; historical agent architecture lead.
• ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders: ConvApparel evaluates conversational user simulation. Historical candidate requiring public data and executable simulator checks.
• Unified Management System for a Hybrid Workforce of Human and Computational Agents: Lifecycle management for hybrid human/agent workforces; March defensive publication predates the window.
• Identifying Hearing Difficulty Moments in Conversational Audio: Audio LMs detect moments of hearing difficulty, a multimodal application needing public data and real-model verification.
• An AI system to help scientists write expert-level empirical software: ERA uses LLM-guided tree search for empirical software science, relevant to Agent/RSI; Nature May publication is historical.
• Bi-level Hierarchical Neural Contextual Bandits for Online Recommendation: H2N neural contextual bandits. January TMLR publication predates the window; offline benchmarks cannot replace full-text industrial online evidence.
• Architecting Conversational Data Systems for Stateless LLM APIs: Separates persistent conversation state from stateless LLM APIs, an agent memory architecture. arXiv first appeared September 1.
• MoXaRt: Audio-Visual Object-Guided Sound Interaction for XR: Audio-visual object-guided sound interaction and separation, a historical multimodal candidate.
• Beyond Vector Similarity: Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration: AST graph relations and context caching aid code migration. Lexical similarity cannot substitute for compile correctness.
• AI-PROPELLER: Warehouse-Scale Inter-Procedural Code Layout Optimization with AlphaEvolve: AlphaEvolve evolves compiler layout policies with hardware-counter rewards, a high-priority historical RSI/Agent candidate.
• LiveSVG: Zero-Shot SVG Animation via Video Generation: Frozen video generation supplies targets, then differentiable rendering fits SVG animation. Multimodal generation, verified May first release.
• LLM-Powered Analysis of IoT User Reviews: Tracking and Ranking Security and Privacy Concerns: Fine-tunes LLMs to identify and cluster security/privacy concerns in IoT reviews. Content understanding, January release; offline reviews do not prove online A/B.
• Rethinking Context-Enriched Time-Series Forecasting Evaluation: TimesX combines textual context and time-series benchmarks, a foundation/multimodal evolve evaluation candidate.
• MoGen: Detailed Neuronal Morphology Generation via Point Cloud Flow Matching: MoGen uses point-cloud flow matching for neuronal morphologies, relevant to 3D generation. Official April ICLR presentation predates the window; review cross-domain exception separately.
• Diffusion Controller: Framework, Algorithms and Parameterization: DiffCon formulates diffusion sampling as control and provides SFT/RLFT controllers; multimodal and post-training, March first release.
• Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion: R4T uses RL to create set-level supervision then distills a diffusion retriever. Already in the manifest; historical duplicate.
• Reinforcement Learning with Discrete Diffusion Policies for Combinatorial Action-Spaces: RL optimizes discrete diffusion policies, relevant to post-training. First released in 2025, not in this 2026 window.
• When Backdoors Go Beyond Triggers: Semantic Drift in Diffusion Models Under Encoder Attacks: SEMAD audits encoder poisoning via geometry, relevant to multimodal safety. July ACL publication predates the window; no invented exact first date.
• Causal Armor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution: Causal ablation attribution triggers selective IPI sanitization and CoT masking, a historical agent safety candidate.
• Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents: CoRedTeam uses execution feedback for collaborative cybersecurity agents, a historical candidate.
• Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion: Same R4T paper as another official URL. Keep the URL audit without counting another candidate.
• Towards A Human-in-the-Loop Framework for Reliable Patch Evaluation using an LLM-as-a-Judge: Humans edit per-bug rubrics before an LLM judges patches. Agent evaluation; 2025 arXiv manuscript.
• Abstain and Validate: A Dual-LLM Policy for Reducing Noise in Agentic Program Repair: Dual-LLM abstention and patch verification are historical coding-agent leads; filtered success must report retention.
• Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair: Jointly generates reproduction and repair tests dynamically. The 120 internal Google bugs are not a local public benchmark.
• PoTRE: Test-Time Reasoning inspired by Cognitive Heterogeneity: PoTRE uses cognitively heterogeneous roles for test-time reasoning, an Agent/reasoning candidate; verified July first release.
• Scientist-One: Verifiable Autonomous Research via Chain-of-Evidence: Scientist One links paper claims to code through evidence chains. Official title/author aliases match the preprint.
• DocLens: A Tool-Augmented Multi-Agent Framework for Long Visual Document Understanding: DocLens combines visual parsing, evidence lenses, and adjudication for long-document QA. Multimodal Agent; first release in 2025.
• Mind the Gap: Structure-Aware Consistency in Preference Learning: SA-DPO adapts margins to semantic distances, a historical post-training candidate.
• Optimized Deferral for Imbalanced Settings: MILD optimizes LLM deferral for imbalanced tasks, retained as a historical inference router.
• ECO: Accurate Quantization-Aware Training Without Master Weights: ECO removes full-precision master weights and corrects update errors in low-precision training. Authors/method matched a differently titled January manuscript.
• A Theoretical Framework for Modular Learning of Robust Generative Models: Gated experts and robust objectives organize modular generative models, foundation training. February manuscript, not new here.
• Learning from Attribution Sets: Attribution sets support statistical estimation for ad conversions. February manuscript verified; assess industrial online evidence separately.
• Platform Competition in the Autobidding World: Analyzes competition among autobidding ad platforms. Manuscript public in 2024; 2026 venue year is not first publication.
• Online Advertising with Spatial Interactions: Allocation theory for spatial externalities in ads. Historical lead; no industrial reproduction admission without online evidence.
• Beyond Vector Similarity: Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration: Full-abstract variant of another official HCRG URL. One historical paper identity.
• Inference Perf: A Benchmarking Tool for GenAI Inference: InferencePerf benchmarks GenAI serving performance. JOSS publication in August does not block the October window.
• Mobility-Embedded POIs: Learning What a Place Is and How It’s Used from Human Movement: Contrastive learning joins POI text and mobility trajectories. Historical foundation/LLM application; map tasks are not online A/B.
• A large language model for complex cardiology care: AMIE cardiology reasoning and collaboration evaluation. Distinguish text supplied to the model from clinicians' original multimodal inputs.
• Fisheye3R: Adapting Unified 3D Feed-Forward Foundation Models to Fisheye Lenses: Fisheye3R adapts 3D foundation models to fisheye images with adaptation tokens; March manuscript located.
• TFRBench: A Reasoning Benchmark for Evaluating Forecasting Systems: TFRBench uses multiple agents to generate verifiable forecasting reasoning in a public benchmark, a historical Agent/evolve candidate.
• OpenClaw in the Wild: Security Analysis of Autonomous Agents: OpenClaw security analysis covers five trust boundaries across the lifecycle. June publication; governance/diagnostics, not capability training.
• Theoretical Perspectives on Data Quality and Synergistic Effects in Pre- and Post-Training Reasoning Models: Theory of data-quality synergy between pretraining and post-training. Relevant to foundation and post-training; first published March.
• Outrunning LLM Cutoffs: A Live Kernel Crash Resolution Benchmark for All,: LivekBench uses live kernel crash resolution to reduce cutoff contamination, a historical agent benchmark.
• Learned Task Relationships in Multi-task Models: Title alias matches YouTube cross-task learning. Historical industrial priority; abstract evidence does not pass the online gate. High-priority historical candidate pending online evidence in full text.
• MAGE: Modality-Agnostic Music Generation and Editing: Flow Transformer multimodal music generation/editing with dynamic modality masks, a historical multimodal candidate.
• Evaluating Contextual Illegality: AI Compliance in Corporate Law Scenarios: Contextual illegality in chat and agent safety evaluation. Venue month predates the window; exact first publication unconfirmed.
• PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation: PointDiT uses pixel-space diffusion for monocular 3D geometry. July first release and official code located.
• GroundSet: A Cadastral-Grounded Dataset for Spatial Understanding with Vector Data: GroundSet provides public spatial understanding data, models, and instruction tuning, a historical multimodal candidate.
• A Theoretical Framework for Statistical Evaluability of Generative Models: Studies theoretical conditions for statistical evaluability of generative models. Official July ICML presentation predates the window; theory is not implemented capability.
• When “Correct” Is Not Safe: Can We Trust Functionally Correct Patches Generated by Code Agents?: Evaluates functionally correct yet vulnerable patches. July ACL version predates the window; check functionality and security separately.
• VIP-MINGLE: A Corpus for Videoconference and In-Person Multimodal Interaction in Group Language Engagement: Cross-setting audiovisual group interaction corpus, a historical multimodal data candidate.
• FUSE: Scaling Verification with Zero Labeled Data: Official Scaling Verification matches FUSE verifier ensembling by authors and method; historical candidate.
• POLCA: Stochastic Generative Optimization with LLM: POLCA uses stochastic-feedback priority queues, diversity, and historical summaries; historical RSI/Agent candidate.
• AI-Simulated Expert Panels for Socio-Technical Scenarios and Decision Guidance: LLM expert panels and cross-impact-consistent scenario generation, a historical agent candidate, not actual expert truth.

Acceptance and next window

The strict coverage audit checks the watermark, eight recall receipts, frozen-candidate closure, terminal dispositions of 542 source records, and exception reasons, attempts, and reopening criteria. New regression tests reject false exhaustiveness, treating unread items as ignored, using historical labels to hide in-window papers, and presenting completed review as implementation. The next window overlaps from 2026-10-06. Metadata exceptions reopen only if materials change or independent discovery supplies evidence.