digest: 2026-06-12

_meta/claims-ledger.md

@@ -280,3 +280,11 @@
 - **2026-06-10** [arxiv:2606.11042] CLAIM: Workflow-GYM (Liya Zhu et al., 57 contributors; corresponds to Jingzhe Ding) — benchmark for **long-horizon GUI tasks in professional domains** with specialised software environments; state-of-the-art models reach **only slightly above 30%** success; failure analysis identifies four distinct modes: **workflow stage omission, error propagation, objective drift, insufficient understanding of professional software environments** | source: https://arxiv.org/abs/2606.11042 | status: open ; the **30% vs. 85% OSWorld-Verified (Fable 5) gap** is the *long-horizon + domain-specific* dimension OSWorld doesn't cover — closer to deployment reality for federal-data / specialised enterprise GUIs; four-failure-mode taxonomy is immediately useful as diagnostic vocabulary for long-running-agent logs; structurally the right shape of CI for long-running agents (fixed long-horizon trajectories with intermediate checkpoints + per-failure-mode regression tracking)
 - **2026-06-10** [arxiv:2606.11052] CLAIM: Attention Amnesia (X. Zhou / B. Zhu / Y. Xu / Z. Li / Y. Chen / H. Wang / Z. Guo) — identifies failure mode in hybrid LLM architectures (**HypeNet, Jet-Nemotron** class): **standard CoT-SFT biases attention gradients toward short-range patterns, disrupting query-key projections responsible for long-range routing**; **HypeNet-9B on NIAH-S2@256K collapses 67.2% → 9.4%** post CoT-SFT (58-pt drop); proposed **QK-Restore** surgically recovers and even improves (**HypeNet-5B S3@256K: 65.4% → 76.4%**) while preserving CoT-SFT reasoning gains | source: https://arxiv.org/abs/2606.11052 | status: open ; **first published evidence** that the "fine-tune for reasoning, deploy at long context" recipe (default for sub-quadratic hybrid architectures) silently breaks long-range recall; defensive requirement: **benchmark NIAH-S2/S3 @ 256K before AND after any reasoning-distillation fine-tune on hybrid checkpoints**; **open empirical question** whether the bias-gradient-toward-short-range mechanism applies to attention-only Transformers under CoT-SFT — paper only tests hybrids
 - **2026-06-10** [arxiv:2606.11164] CLAIM: ReasonAlloc (W. Liu / H. Shi / Y. Li / Weizhi Fei / Xiangyuan Wang / Mengzhe Ruan / Hanxu Hou / Peisong Wang / Linqi Song / Shuang Qiu) — two-level KV cache budget allocator for long-CoT reasoning: **offline layer-wise allocation** captures "Reasoning Wave" architectural demand patterns + **online head-wise reallocation** dynamically shifts budget to information-rich attention heads during decoding; **training-free**, **plug-and-play with R-KV / SnapKV / Pyramid-RKV** eviction policies, minimal runtime overhead; evaluated on **MATH-500 + AIME 2024** with DeepSeek-R1-Distill-Llama-8B / DeepSeek-R1-Distill-Qwen-14B / AceReason-14B; **greatest improvements at small budgets (128–512 tokens)** vs. uniform-budget baselines | source: https://arxiv.org/abs/2606.11164 | status: open ; **hierarchical-allocation axis** orthogonal to [arxiv:2605.19660] OScaR (per-token INT2 quantisation, 5.3× memory reduction) and [arxiv:2605.22106] ArborKV (tree-aware allocation, ~4× peak reduction) on the M5 Max 64 GB local-inference budget — three axes (token codec / structure-aware / layer×head importance) should stack on the same memory budget
+- **2026-06-12** [deepmind-blog-2026-06-10-diffusiongemma] CLAIM: DiffusionGemma (Google DeepMind, June 10 2026, Apache 2.0) — open-weight **26B-total / 3.8B-active MoE text-diffusion LM** built on Gemma 4 backbone with novel diffusion head + **bi-directional attention**; generates **256 tokens in parallel per forward pass**; **1000+ tok/s on single H100, 700+ tok/s on RTX 5090**; fits in **~18 GB VRAM at NVFP4 4-bit**; **MLX/vLLM/HF Transformers supported at launch**, llama.cpp queued; Google explicitly acknowledges **lower overall quality than autoregressive Gemma 4** as the trade-off | source: https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/ | status: open ; **first frontier-lab text-diffusion LM shipped with first-party MLX support at launch** — direct M5 Max sweet-spot candidate at same active-param class as Qwen3-30B-A3B but with fundamentally different decoding loop; **open empirical question whether the 4× speedup transfers from H100/RTX-5090 to M5 Max Metal kernels** — published numbers are GPU-only, no M5 Max benchmark yet; right experiment is direct head-to-head vs Qwen3-30B-A3B on agent inner-loop workloads (JSON tool-call bursts, code infill, structured rewrites)
+- **2026-06-12** [arxiv:2606.13681] CLAIM: EvoArena + EvoMem (Jundong Xu et al., 14 authors incl. Caiming Xiong, Bryan Hooi, Anh Tuan Luu — NTU + Salesforce AI Research + Singapore cluster) — benchmark models **environment change as sequences of progressive updates across terminal, software, social domains** + paired memory system tracking **memory evolution as structured update histories**; current agents avg **39.6% on EvoArena**; EvoMem yields **+1.5% EvoArena, +6.1% GAIA, +4.8% LoCoMo, +3.7% chain-level on consecutive evolutionary subtasks**; mechanistic analysis shows EvoMem improves "evidence capture, enabling better preservation of evolving environment states" | source: https://arxiv.org/abs/2606.13681 | status: open ; **first dedicated benchmark for "the world changed under your agent" axis** missing from HORMA/MemForest/MAGE/Infini Memory cluster (which all assume write-once-read-many); update-history primitive is **direct philosophical opposite of PROJECTMEM's append-only-immutable log** — both can be true depending on whether the supersedence event is preserved or destructive; open empirical question whether write-many-supersede-prior outperforms append-only when the underlying world is genuinely evolving
+- **2026-06-12** [arxiv:2606.13663] CLAIM: HyperTool (Yaxin Du et al., SJTU + collaborators) — MCP-style **code-block tool-invocation primitive** that bundles existing MCP tool calls through their schemas + value manipulation + local intermediate-result passing into a single execution unit; names underlying problem **"execution-granularity mismatch"** where deterministic tool workflows consume excessive context exposing low-level dataflow to the reasoning trace; **Qwen3-32B 15.69 → 35.29% (+19.6pp) and Qwen3-8B 9.93 → 33.33% (+23.4pp) on MCP-Universe**; surpasses GPT-OSS and Kimi-k2.5 baselines on average accuracy; **schema-compatible with existing MCP servers** (no protocol break) | source: https://arxiv.org/abs/2606.13663 | status: open ; **most directly portable MCP-stack artifact in recent months** — first named-and-reduced "MCP execution-granularity tax"; composes with [arxiv:2606.12320] Five-Plane governance (HyperTool reduces frequency of model-in-loop calls, Five-Plane controls which are allowed); +20pp deltas are unusually clean for a context-engineering primitive
+- **2026-06-12** [arxiv:2606.13003] CLAIM: "The Illusion of Multi-Agent Advantage" (Prathyusha Jwalapuram et al., Salesforce AI Research + UBC + NTU) — systematic head-to-head shows **automatically-generated MAS consistently underperforms Chain-of-Thought + Self-Consistency despite up to 10× cost** across reasoning datasets + interactive multi-step workflows (BrowseComp-Plus); **expert-architected MAS still outperforms auto-MAS** on diagnostic synthetic dataset designed to isolate MAS advantage; frames as "architectural bloat prioritizing superficial complexity" + indicts existing eval frameworks for ignoring compute cost | source: https://arxiv.org/abs/2606.13003 | status: open ; **NOT a contradiction of expert-architected MAS positive results** (Arbor [arxiv:2606.11926] 2.5× Claude Code on research, MUSE-Autoskill [arxiv:2605.27366], MasReplay-T1) — the resolution is that **auto-MAS design pipelines** are the failed category, not multi-agent systems per se; actionable rule: any new agent in a system should beat CoT-SC at the same compute budget or it's decoration; **first sharp public counter to the auto-MAS-pipeline hype wave** of 2026
+- **2026-06-12** [arxiv:2606.13662] CLAIM: EurekAgent ("Agent Environment Engineering is All You Need for Autonomous Scientific Discovery", Amy Xin et al., Tsinghua KEG) — position-plus-system arguing **bottleneck for autonomous scientific discovery shifts from prescribing workflows to designing environments** as agent capability rises; engineers environments along **4 dimensions — permissions (bounded execution), artifacts (filesystem/Git), budgets (exploration awareness), human oversight**; SOTA across math + kernel engineering + ML tasks; **new SOTA 26-circle packing result at ~$11 total API cost**; open-sourced | source: https://arxiv.org/abs/2606.13662 | status: open ; **executable counterpart to [arxiv:2606.12191] Agentic Environment Engineering survey from June 11** (survey identified 4 co-evolution pathways, EurekAgent demonstrates orchestration-centric one end-to-end on verifiable benchmark at hobbyist cost); $11 SOTA-on-verifiable-optimization is the calibration number — says well-designed environment + stock frontier model can match specialized methods on verifiable problems; structurally same long-lived-coordinator + scoped-executors + persistent-artifact pattern as Arbor [arxiv:2606.11926]
+- **2026-06-12** [arxiv:2606.12908] CLAIM: SENTINEL (Ziyi Wang et al., NEU + collaborators) — RL training framework converting agent failures into targeted training tasks via **Controller–Proposer–Solver loop**: Controller mines failure patterns from unsuccessful trajectories, Proposer generates new executable tasks targeting those weaknesses, Solver trains on curated tasks; replaces fixed task distributions (which become uninformative as agent improves) with **failures-as-curriculum**; **Qwen3-4B-Thinking-2507 Pass@1 66.4 → 74.9% (+8.5pp) on Tau2-Bench Retail**; outperforms standard RL across Pass@k on general synthetic tasks | source: https://arxiv.org/abs/2606.12908 | status: open ; **failure-as-curriculum primitive generalizes beyond RL** to any agent self-improvement loop (failure log → propose-similar-task → run-and-grade → add to test suite); PROJECTMEM's append-only event log [arxiv:2606.12329] is natural input substrate; composes with HERO [arxiv:2606.11559] from June 11 (HERO converts each observation into turn-level diagnostic; SENTINEL converts trajectory-level failure into new task) — together they are the two halves of "learn from your own runs without privileged terminal rewards"
+- **2026-06-12** [arxiv:2606.13115] CLAIM: G-Long (Minjun Choi et al., Sungkyunkwan University) — long-dialogue agent memory via **fine-tuned small LM doing structured triplet extraction + associative graph retrieval** + **attention-aware importance scoring** using T5-summarizer signals to identify salient info; **+9.8% response quality on MSC, 40.8% retrieval recall on LME**; substantially lower compute than baselines | source: https://arxiv.org/abs/2606.13115 | status: open ; the interesting thing is **small-fine-tuned-LM-as-memory-curator as separable tier** in the planner/executor/curator topology (distinct from both planner and executor); adds **associative-graph axis** to the multi-architecture memory thread (HORMA file-system tree, MemForest temporal tree, EvoMem update history, G-Long associative graph) — field is rapidly exploring orthogonal axes with **NO head-to-head head-to-head benchmark across the four**, becoming increasingly urgent gap
+- **2026-06-12** [arxiv:2606.13177] CLAIM: MemRefine (Minjae Kim, Jinheon Baek, Soyeong Jeong, Sung Ju Hwang — KAIST) — memory-compaction framework that **uses similarity only to *propose* candidate entry pairs, then defers delete/merge/preserve decisions to LLM judge based on factual content**; iterates until target budget while preserving downstream performance; names underlying problem that **surface similarity correlates poorly with factual importance** so similarity-only memory pruning systematically deletes the wrong things; beats rule-based baselines under tight budgets across multiple memory frameworks | source: https://arxiv.org/abs/2606.13177 | status: open ; **counterpart to EvoMem at the deletion end** (EvoMem grows monotonically with supersedence events; MemRefine is the bounded-budget compactor that decides which historic entries survive); LLM-judge-over-similarity-proposals primitive is **reusable outside memory** for any "score and keep top-K" pruning where surface similarity is a poor proxy for content importance

_meta/processed.json

@@ -1,4 +1,52 @@
 [
+  {
+    "id": "deepmind-blog-2026-06-10-diffusiongemma",
+    "url": "https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/",
+    "title": "DiffusionGemma — Google DeepMind Open-Weight 26B-Total / 3.8B-Active MoE Text-Diffusion LM Built on Gemma 4 Backbone; 256 Tokens in Parallel Per Forward Pass + Bi-Directional Attention; 1000+ tok/s H100 / 700+ tok/s RTX 5090; ~18 GB at NVFP4 4-bit; MLX/vLLM/HF Transformers at Launch (llama.cpp Coming); Apache 2.0; Lower Quality vs AR Gemma 4 Acknowledged",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.13681",
+    "url": "https://arxiv.org/abs/2606.13681",
+    "title": "EvoArena + EvoMem — Benchmark Modeling Environment Change as Progressive Updates Across Terminal/Software/Social Domains + Memory System Tracking Structured Update Histories; Current Agents Avg 39.6% on EvoArena; EvoMem +1.5% EvoArena, +6.1% GAIA, +4.8% LoCoMo, +3.7% Chain-Level on Consecutive Evolutionary Subtasks",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.13663",
+    "url": "https://arxiv.org/abs/2606.13663",
+    "title": "HyperTool — MCP-Style Code-Block Tool Invocation Bundling Existing MCP Tool Calls + Value Manipulation + Local Intermediate-Result Passing into a Single Reasoning-Trace Unit; Names 'Execution-Granularity Mismatch'; Qwen3-32B 15.69→35.29% (+19.6pp) and Qwen3-8B 9.93→33.33% (+23.4pp) on MCP-Universe; Surpasses GPT-OSS + Kimi-k2.5",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.13003",
+    "url": "https://arxiv.org/abs/2606.13003",
+    "title": "The Illusion of Multi-Agent Advantage — Auto-Generated MAS Consistently Underperforms Chain-of-Thought + Self-Consistency Despite Up to 10× Cost Across Reasoning Datasets + BrowseComp-Plus; Expert-Architected MAS Beats Auto-MAS on Diagnostic Synthetic Dataset; Indicts Automated MAS Design as 'Architectural Bloat Prioritizing Superficial Complexity'",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.13662",
+    "url": "https://arxiv.org/abs/2606.13662",
+    "title": "EurekAgent ('Agent Environment Engineering is All You Need for Autonomous Scientific Discovery') — Engineers Environments Along 4 Dimensions (Permissions/Artifacts/Budgets/Human Oversight) Instead of Prescribing Workflows; SOTA on Math/Kernel/ML Tasks Incl. New SOTA 26-Circle Packing Result at ~$11 Total API Cost; Open-Sourced",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.12908",
+    "url": "https://arxiv.org/abs/2606.12908",
+    "title": "SENTINEL — Failure-Driven RL Training Framework with Controller–Proposer–Solver Loop That Mines Failure Patterns, Generates Targeted Tasks, Trains on Curated Curriculum (Failures-as-Curriculum); Qwen3-4B-Thinking-2507 Pass@1 66.4→74.9% (+8.5pp) on Tau2-Bench Retail; Outperforms Standard RL Across Pass@k on General Synthetic Tasks",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.13115",
+    "url": "https://arxiv.org/abs/2606.13115",
+    "title": "G-Long — Graph-Enhanced Long-Term Dialogue Memory via Fine-Tuned Small-LM Structured Triplet Extraction + Associative Graph Retrieval + Attention-Aware Importance Scoring (T5 Summarizer Signals); +9.8% Response Quality on MSC, 40.8% Retrieval Recall on LME; Substantially Lower Compute vs Baselines",
+    "date_covered": "2026-06-12"
+  },
+  {
+    "id": "arxiv:2606.13177",
+    "url": "https://arxiv.org/abs/2606.13177",
+    "title": "MemRefine — LLM-Guided Long-Term-Memory Compaction Using Similarity Only to Propose Candidate Pairs, Then Defers Delete/Merge/Preserve Decisions to LLM Judge Based on Factual Content; Iterates Until Budget Met While Preserving Downstream Performance; Beats Rule-Based Baselines Under Tight Budgets Across Multiple Memory Frameworks",
+    "date_covered": "2026-06-12"
+  },
   {
     "id": "arxiv:2606.12329",
     "url": "https://arxiv.org/abs/2606.12329",