digest: 2026-05-22

_meta/claims-ledger.md

@@ -107,3 +107,12 @@
 - **2026-05-21** [arxiv:2605.21468] CLAIM: RELEX (Yu Meng UVA cluster) — RLVR weight-update trajectories are **extremely low-rank**; rank-1 approximation captures majority of downstream gains; estimating rank-1 subspace from short observation window enables linear-regression extrapolation; matches full RLVR performance with **as few as 15% of training steps**, extrapolation **10–20× beyond observed prefix**; mechanism attributed to denoising effect from subspace projection | source: https://arxiv.org/abs/2605.21468 | status: open ; third independent low-rank linearity finding in the agentic LLM stack alongside [arxiv:2605.07990] (tool calling linearly readable in logits) and [arxiv:2605.14038] (tool-necessity orthogonal probe directions) — three different surfaces converging on dominant linear structure
 - **2026-05-21** [deepmind-blog-2026-05-19-co-scientist] CLAIM: DeepMind Co-Scientist (Nature publication, May 19) — multi-agent Gemini system with **specialized agents for hypothesis generation, critical evaluation, and ranking** + AlphaGo-style tournament evolution; integrates ChEMBL / UniProt / web search / AlphaFold as external tools; demonstrated across liver fibrosis treatments, ALS research pathways, cellular aging reversal with named external scientist collaborators | source: https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/ | status: open ; production-blessed at Nature-publication tier the same role-separation pattern as [arxiv:2605.13880] PREPING (Proposer/Solver/Validator) and [arxiv:2605.18401] SkillsVote (Collection/Search/Post-Execution gating); concrete instantiation of [arxiv:2605.15177] OpenDeepThink Bradley-Terry pairwise aggregation applied to research hypotheses
 - **2026-05-21** [arxiv:2605.20668] CLAIM: 45 domain scientists rate 2,960 criticisms from reviews of 82 Nature-family papers; **GPT-5.2-powered reviewing agent scores above each paper's top-rated human reviewer (60.0% vs. 48.2%, p = 0.009)**; all AI reviewers exceed lowest-rated human; AI systems exhibit **16 recurring weaknesses humans do not share**; framed as complements not substitutes | source: https://arxiv.org/abs/2605.20668 | status: open ; conflict-pair with [arxiv:2605.19156] ResearchArena (AI fails as author of full papers under artifact-aware review) — both can be true: AI is competitive **as reviewer** while still failing **as author** of artifact-aware-scored work; the asymmetry is the operational signal for product strategy (review/critique surfaces are closer to production-ready than full-artifact-generation surfaces)
+- **2026-05-22** [google-blog-2026-05-19-gemini-3-5] CLAIM: Gemini 3.5 Flash (DeepMind / Google AI, May 19 release) reports **Terminal-Bench 2.1 76.2%, GDPval-AA 1656 Elo, MCP Atlas 83.6%, CharXiv Reasoning 84.2%**; **~4× faster output tokens/sec and "less than half the cost" of other frontier models**; outperforms Gemini 3.1 Pro on coding/agentic benchmarks; Pro variant in internal testing | source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/ | status: open ; vendor-self-reported; **MCP Atlas 83.6% is the first first-party MCP-specific eval number published in a frontier-model release post anywhere**, setting a comparative anchor that Anthropic / OpenAI have not yet provided
+- **2026-05-22** [microsoft-research-2026-05-21-magenticlite-fara15] CLAIM: Microsoft Research released **MagenticLite (agentic application on Magentic-UI) + MagenticBrain (14B Qwen 3 14B fine-tune orchestrator) + Fara1.5 (4B / 9B / 27B computer-use family)** with all three components downloadable (GitHub + Microsoft Foundry); **Fara1.5-9B is SOTA among small models on OnlineMind2Web (300 web tasks), nearly doubling Fara-7B; 27B reaches >90%**; harness has FastAPI orchestration with context compaction + critical-point detection, sandboxed by Quicksand (QEMU) | source: https://www.microsoft.com/en-us/research/blog/magenticlite-magenticbrain-fara1-5-an-agentic-experience-optimized-for-small-models/ | status: open ; largest small-model agent-system release this cycle; pairs with [hf-blog-2026-04-29-ibm-granite-4-1] Granite 4.1-8B and [hf-blog-2026-05-08-allenai-emo] EMO as third lab arguing tight harness ↔ small-model codesign > generic frontier API for production agents
+- **2026-05-22** [arxiv:2605.21850] CLAIM: ACC (Agent Context Compilation) converts existing agent trajectories from search / SWE / DB-query agents into long-context QA pairs by combining the original question with multi-turn tool responses + environment observations; **Qwen3-30B-A3B + ACC reaches MRCR 68.3 (+18.1) and GraphWalks 77.5 (+7.6), comparable to Qwen3-235B-A22B**, while preserving GPQA / MMLU-Pro / AIME / IFEval; mechanism shows task-adaptive attention restructuring + expert specialization; combinable with any long-context extension method | source: https://arxiv.org/abs/2605.21850 | status: open ; first paper to use agent-harness rollouts as long-context training data; closes the harness ↔ base-model loop; supports the **30B-A3B-as-local-fit-sweet-spot** thesis alongside [arxiv:2605.05191] Context-ReAct and [arxiv:2605.04050] LCM
+- **2026-05-22** [arxiv:2605.22177] CLAIM: Maestro (Jianhua Tao cluster) RL-trains a **4B lightweight orchestrator** to compose ensembles of frozen expert models + a two-tier skill library; reports **70.1% average across 10 multimodal benchmarks vs. GPT-5 69.3% and Gemini-2.5-Pro 68.7%**; **59.5% on 4 challenging benchmarks when augmented with out-of-domain experts**; code released | source: https://arxiv.org/abs/2605.22177 | status: open ; second result in 30 days (alongside [arxiv:2605.05007] Uno-Orchestra) arguing **structured / learned orchestration over frozen experts > monolithic frontier model** as the cost-quality frontier; converges with [arxiv:2605.06614] SkillOS, [arxiv:2605.05413] Constant-Context Skill Learning, [arxiv:2605.18401] SkillsVote on explicit-skill-library pattern (4 independent confirmations in 4 weeks); open question whether Maestro + Uno-Orchestra compose
+- **2026-05-22** [arxiv:2605.22791] CLAIM: Gated DeltaNet-2 (NVIDIA — Hatamizadeh / Choi / Kautz) decouples linear-attention memory management into **channel-wise erase gate b_t + channel-wise write gate w_t** (vs. prior single scalar gate); 1.3B params on 100B FineWeb-Edu tokens outperforms Mamba-2 / Gated DeltaNet / KDA / Mamba-3 across recurrent-only and hybrid settings, with named improvement on **RULER multi-key needle-in-a-haystack**; code released under NVIDIA Source Code License-NC; **no pre-trained weights yet** | source: https://arxiv.org/abs/2605.22791 | status: open ; conceptual parallel to [arxiv:2605.21463] Mem-π decoupling **when-to-write from what-to-write** — same insight at two scales (architecture-level recurrent state vs. agent-level explicit memory) in the same week
+- **2026-05-22** [arxiv:2605.22148] CLAIM: Ratchet — single-agent skill-lifecycle loop in which a **frozen Claude Opus 4.7** writes / retrieves / curates / retires natural-language skills via 4 hygiene mechanisms (outcome-driven retirement, bounded active-cap, meta-skill authoring, pattern canonicalisation); **MBPP+ hard-100 pass@1 lifts from 0.258 baseline → 0.584 rolling mean (peak 0.658) over 100 rounds; SWE-bench Verified +0.22 peak over 20 rounds**; ablations isolate **retirement + meta-skill authoring as load-bearing components** | source: https://arxiv.org/abs/2605.22148 | status: open ; minimal-end of the skill-curation spectrum joining [arxiv:2605.06614] SkillOS, [arxiv:2605.13941] EvolveMem, [arxiv:2605.18401] SkillsVote, [arxiv:2605.21463] Mem-π — 5 independent papers in 5 weeks on skill-libraries + hygiene > raw memory accumulation; open question which primitives compose
+- **2026-05-22** [arxiv:2605.22721] CLAIM: DecentMem (Hao / Long / Zhao) — multi-agent systems with **decentralized dual-pool memory** (exploitation pool of consolidated past trajectories + exploration pool of LLM-generated candidates), reweighted via LLM-as-judge feedback; theoretical O(log T) regret matching stochastic bandit lower bounds; **+23.8% accuracy over centralized memory, +52.5% over no-memory, 49% token reduction** across AutoGen / DyLAN / AgentNet × 5 benchmarks × Qwen3-4B/8B/14B + Gemma4-E2B/E4B | source: https://arxiv.org/abs/2605.22721 | status: open ; **directional contradiction with [arxiv:2605.08060] Memory Curse** (more accessible history degrades cooperation in 18/28 settings) — both can be true if the mechanism is task-evidence (DecentMem) vs. cooperation (Memory Curse); ledger now holds **conditional claim**: "more memory helps under decentralized dual-pool on task-evidence tasks; harms under flat history on cooperation"; empirical resolution requires running DecentMem under the Memory Curse 4-game protocol
+- **2026-05-22** [arxiv:2605.22502] CLAIM: "Subterranean agents" — compiling agentic workflows into LLM weights (rather than running through external orchestrators) tested on **3 real workflows (travel booking 14 nodes, Zoom support 14 nodes, insurance claims 55 nodes with 6 decision hubs)** reports **near-frontier quality at two orders of magnitude less cost** vs. equivalent external-orchestrator implementation; abstract notes existing agent-orchestration framework ecosystem totals >290,000 GitHub stars | source: https://arxiv.org/abs/2605.22502 | status: open ; third paper this cycle (alongside [arxiv:2605.05007] Uno-Orchestra and [arxiv:2605.22177] Maestro) arguing **the orchestration cost line is collapsing under learned / compiled approaches**; no code release per abstract; abstract does not decompose "near-frontier quality" — needs evaluation on **held-out workflow variants** per [arxiv:2605.21384] SpecBench's visible↔held-out gap framework before treating the 100× as portable
+- **2026-05-22** [google-research-2026-05-19-era] CLAIM: ERA (Google Research) — Gemini-based AI tool for expert-level scientific coding; uses **tree-search to optimize output code** by searching literature, writing/refining code, combining techniques, evaluating results; reports "expert-level performance" across genomics / public health / satellite imagery / neuroscience prediction / time-series / mathematics; **open-source at google-research/era** | source: https://research.google/blog/empirical-research-assistance-era-from-nature-publication-to-catalyzing-computational-discovery/ | status: open ; complements [deepmind-blog-2026-05-19-co-scientist] Co-Scientist (hypothesis generation vs. ERA's code optimization) — same lab, same week, two complementary research-agent surfaces; covers a different stage of the scientific-method workflow

_meta/processed.json

@@ -1,4 +1,13 @@
 [
+  { "id": "google-blog-2026-05-19-gemini-3-5", "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/", "title": "Gemini 3.5: Frontier intelligence with action (Flash release)", "date_covered": "2026-05-22" },
+  { "id": "microsoft-research-2026-05-21-magenticlite-fara15", "url": "https://www.microsoft.com/en-us/research/blog/magenticlite-magenticbrain-fara1-5-an-agentic-experience-optimized-for-small-models/", "title": "MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models", "date_covered": "2026-05-22" },
+  { "id": "arxiv:2605.21850", "url": "https://arxiv.org/abs/2605.21850", "title": "ACC: Compiling Agent Trajectories for Long-Context Training", "date_covered": "2026-05-22" },
+  { "id": "arxiv:2605.22177", "url": "https://arxiv.org/abs/2605.22177", "title": "Maestro: Reinforcement Learning to Orchestrate Hierarchical Model-Skill Ensembles", "date_covered": "2026-05-22" },
+  { "id": "arxiv:2605.22791", "url": "https://arxiv.org/abs/2605.22791", "title": "Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention", "date_covered": "2026-05-22" },
+  { "id": "arxiv:2605.22148", "url": "https://arxiv.org/abs/2605.22148", "title": "Ratchet: A Minimal Hygiene Recipe for Self-Evolving LLM Agents", "date_covered": "2026-05-22" },
+  { "id": "arxiv:2605.22721", "url": "https://arxiv.org/abs/2605.22721", "title": "DecentMem: Self-Evolving Multi-Agent Systems via Decentralized Memory", "date_covered": "2026-05-22" },
+  { "id": "arxiv:2605.22502", "url": "https://arxiv.org/abs/2605.22502", "title": "Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost", "date_covered": "2026-05-22" },
+  { "id": "google-research-2026-05-19-era", "url": "https://research.google/blog/empirical-research-assistance-era-from-nature-publication-to-catalyzing-computational-discovery/", "title": "Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery", "date_covered": "2026-05-22" },
   { "id": "arxiv:2605.21463", "url": "https://arxiv.org/abs/2605.21463", "title": "Mem-π: Adaptive Memory through Learning When and What to Generate", "date_covered": "2026-05-21" },
   { "id": "arxiv:2605.19660", "url": "https://arxiv.org/abs/2605.19660", "title": "OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond", "date_covered": "2026-05-21" },
   { "id": "arxiv:2605.21226", "url": "https://arxiv.org/abs/2605.21226", "title": "OCTOPUS: Optimized KV Cache for Transformers via Octahedral Parametrization Under Optimal Squared Error Quantization", "date_covered": "2026-05-21" },