The DAIR Gazette - ML Papers of the Week

[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_151":152,"_153":154},"routes/_app.x.$handle.$name",{"_11":12,"_13":14,"_15":16,"_18":19,"_17":-7,"_115":-5,"_20":116,"_22":117,"_118":119,"_120":121,"_25":122,"_24":-5,"_123":-5,"_124":-5,"_125":53,"_126":127,"_128":44,"_129":-5,"_130":-5,"_131":132},"routes/_app.x.$handle.$name.branch.$branch.code.$",{"_11":12,"_13":14,"_15":16,"_17":-7,"_18":19,"_20":21,"_22":23,"_24":-7,"_25":26,"_27":-5,"_28":-5,"_29":-5,"_30":-5,"_31":-5},"handle","danielpcox","name","dair_gazette_2026_03_01","branchName","main","version","path","","project",{"_33":61,"_13":14,"_45":62,"_63":-5,"_64":53,"_65":66,"_67":66,"_68":-5,"_69":-5,"_70":71,"_72":-5,"_73":53,"_74":-5,"_75":-5,"_76":77,"_78":79,"_80":-5,"_81":-5,"_82":-5,"_83":-5,"_84":85,"_86":85,"_87":88,"_89":88,"_90":88,"_91":85,"_92":85,"_93":94,"_95":96,"_97":98,"_99":53},"tree",[56],"file","contents",[32],"upstream","pull","ahead","behind","envVars",{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":46,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_18":55},"id","f020def8-1528-11f1-a95b-42dde27851f2","mod.ts","fileId","42f84000-150e-11f1-8d85-42dde27851f2","parentId","branchId","42cf258a-150e-11f1-88c7-42dde27851f2","branchVersion",2,"latest",true,"createdAt",["D",1772340099659],1,"fileType","http","content","import { Hono } from \"npm:hono\";\n\nconst app = new Hono();\n\nconst papers = [\n {\n id: 1,\n title: \"Intelligent AI Delegation\",\n summary: \"Google DeepMind introduces a comprehensive framework for intelligent AI delegation that goes beyond simple task assignment. The framework models delegation as a sequence of decisions: whether to delegate, how to instruct, and how to verify and integrate AI outputs.\",\n highlights: [\n \"Adaptive delegation structure for real-time adaptation.\",\n \"Trust calibration mechanisms to prevent over/under-delegation.\",\n \"Structured approaches for validating AI outputs.\",\n \"Multi-agent delegation networks for accountability tracking.\"\n ],\n link: \"https://arxiv.org/abs/2602.11865\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7821f5c8-d0cb-4376-8ecd-6c1b56178810_5576x1520.png\"\n },\n {\n id: 2,\n title: \"Emergent Socialization in AI Agent Society\",\n summary: \"A study on Moltbook, a social network with no humans where all participants are LLM-driven agents, challenges the assumption that scale and interaction density alone produce meaningful social dynamics.\",\n highlights: [\n \"Moltbook as a natural laboratory for AI-only social platforms.\",\n \"Metrics for semantic stabilization and lexical change.\",\n \"Agents fail to develop stable social structures despite scale.\",\n \"Shared memory is essential for genuine social learning.\"\n ],\n link: \"https://arxiv.org/abs/2602.14299\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdf37da9e-3ca3-4e62-9c59-02f175b9afec_785x473.png\"\n },\n {\n id: 3,\n title: \"Lossless Context Management (LCM)\",\n summary: \"Lossless Context Management (LCM) is a deterministic architecture for LLM memory that outperforms Claude Code on long-context tasks. It uses recursive context compression and task partitioning.\",\n highlights: [\n \"Recursive context compression into hierarchical summary DAGs.\",\n \"Engine-managed parallel primitives like LLM-Map.\",\n \"Three-level escalation to reduce context overflow.\",\n \"Outperforms Claude Code on OOLONG benchmarks.\"\n ],\n link: \"https://papers.voltropy.com/LCM\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe77fd9a2-d53b-4b6c-94b6-7d70b93c4828_2481x1333.png\"\n },\n {\n id: 4,\n title: \"GLM-5: From Vibe Coding to Agentic Engineering\",\n summary: \"GLM-5 is a foundation model from Zhipu AI designed to transition from vibe coding to agentic engineering. It introduces novel asynchronous agent RL algorithms.\",\n highlights: [\n \"Asynchronous agent RL decouples generation from training.\",\n \"Focuses on end-to-end software engineering tasks.\",\n \"DSA compression reduces computational overhead for long-context.\",\n \"Strong benchmark results on real-world projects.\"\n ],\n link: \"https://arxiv.org/abs/2602.15763\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4cc8bdf1-6f70-43fb-8924-e083c10003a2_996x593.png\"\n },\n {\n id: 5,\n title: \"MemoryArena: Evaluating Multi-Session Memory\",\n summary: \"MemoryArena introduces a benchmark for evaluating how agents utilize memory across multiple interconnected sessions. It exposes a gap between retrieval accuracy and actionable memory use.\",\n highlights: [\n \"Agentic memory evaluation beyond simple recall.\",\n \"Interdependent multi-session tasks like web navigation.\",\n \"Exposing evaluation blind spots in current benchmarks.\",\n \"Practical implications for building persistent agents.\"\n ],\n link: \"https://arxiv.org/abs/2602.16313\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa5b7de74-4d5a-4a15-ab81-ecad4fa7cc6a_997x806.png\"\n },\n {\n id: 6,\n title: \"MAPLE: Specialized Sub-Agents for Memory\",\n summary: \"MAPLE proposes separating memory, learning, and personalization into specialized sub-agents. The framework achieves a 14.6% improvement in personalization scores.\",\n highlights: [\n \"Sub-agent decomposition for memory, learning, and personalization.\",\n \"Asynchronous learning distills patterns without consuming context.\",\n \"Context-budget-aware personalization prevents dilution.\",\n \"Validated through the MAPLE-Personas benchmark.\"\n ],\n link: \"https://arxiv.org/abs/2602.13258\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f0190d0-63ca-49ab-ba39-8e8015606394_7562x5725.png\"\n },\n {\n id: 7,\n title: \"SkillsBench: Procedural Knowledge Gap\",\n summary: \"SkillsBench evaluates whether LLM agents can generate their own procedural knowledge. It reveals that agents benefit from consuming skills but cannot reliably author them.\",\n highlights: [\n \"Curated skills boost performance significantly (up to 51.9pp).\",\n \"Self-generated skills provide no benefit on average.\",\n \"Focused modules outperform comprehensive documentation.\",\n \"Smaller models with skills can match larger models.\"\n ],\n link: \"https://arxiv.org/abs/2602.12670\",\n image: \"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3d755769-9e33-4897-a568-4b7a9a179dd9_997x1149.png\"\n },\n {\n id: 8,\n title: \"LongCLI-Bench: Extended CLI Tasks\",\n summary: \"LongCLI-Bench benchmarks how well AI agents handle complex, extended tasks through command-line interfaces. Leading agents succeed less than 20% of the time.\",\n highlights: [\n \"20 demanding tasks spanning development and optimization.\",\n \"Most failures occur early in task execution.\",\n \"Human-agent collaboration yields greater improvements.\",\n \"Automated self-correction alone is insufficient.\"\n ],\n link: \"https://arxiv.org/abs/2602.14337\"\n },\n {\n id: 9,\n title: \"CogRouter: Adaptive Reasoning Depth\",\n summary: \"CogRouter enables adaptive reasoning depth for LLM agents by dynamically selecting from four hierarchical cognitive levels at each step.\",\n highlights: [\n \"Dynamic selection from four hierarchical cognitive levels.\",\n \"Confidence-aware advantage reweighting during training.\",\n \"Achieves 82.3% success rate on agentic benchmarks.\",\n \"Consumes fewer tokens by skipping heavy reasoning on routine steps.\"\n ],\n link: \"https://arxiv.org/abs/2602.12662\"\n },\n {\n id: 10,\n title: \"Team of Thoughts: Multi-Agent Scaling\",\n summary: \"Team of Thoughts presents a multi-agent framework for efficient test-time scaling through orchestrated tool calling.\",\n highlights: [\n \"Orchestrated tool calling with specialized agents.\",\n \"Self-assessment for tool agents and orchestrator calibration.\",\n \"Achieves 96.67% on AIME24 and 72.53% on LiveCodeBench.\",\n \"Substantially exceeds homogeneous baselines.\"\n ],\n link: \"https://arxiv.org/abs/2602.16485\"\n }\n];\n\napp.get(\"/\", (c) => {\n const html = `\n\n\n\n \n \n The DAIR Gazette - ML Papers of the Week\n \n \n\n\n

\n \n \n

\n\n

\n ${papers[0].title}\n

By Google DeepMind • Special Report

\n $\"Delegation$ \n

${papers[0].summary.charAt(0)}${papers[0].summary.slice(1)}

${h}

\n\n

\n ${papers[1].title}\n

Social Dynamics in AI-Only Environments

\n $\"Moltbook$ \n

${papers[1].summary}

${h}

\n\n \n\n \n

\n\n

Briefs from the Field

\n ${papers[7].title}\n

${papers[7].summary}

\n ${papers[8].title}\n

${papers[8].summary}

\n ${papers[9].title}\n

${papers[9].summary}

\n\n \n

\n\n\n `;\n return c.html(html);\n});\n\nexport default app.fetch;\n","deleted",false,"label",[35],{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":57,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_58":59,"_18":60},["D",1772340099659],"children",[],[35],"42cec13a-150e-11f1-88c7-42dde27851f2",["D",1772328641971],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","imageUrl","userId","019c649c-877d-70d7-a441-396b9c85efaa","pinnedFileId","pinPreview","pinMigrated","screenshotVersion","likeCount",0,"orgId","019c649c-877d-70d7-a441-3e13100c1184","tursoName","tursoHostname","tursoMeasuredAt","pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_11":12,"_114":110},"org",{"_11":12,"_108":-5,"_109":110,"_70":71,"_111":112,"_113":53},"branch",{"_17":42,"_13":16,"_33":40,"_100":61,"_101":-5,"_102":-5,"_103":-5,"_104":-5,"_105":-5,"_106":14,"_107":12},"isOwned","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","displayName","avatar","https://img.clerk.com/eyJ0eXBlIjoicHJveHkiLCJzcmMiOiJodHRwczovL2ltYWdlcy5jbGVyay5kZXYvb2F1dGhfZ2l0aHViL2ltZ18zOWpmZkZlTFFUYkJFSlE4R291RFhqVE1ma0UifQ","tier","free","canRestrictVals","profileImageUrl","adoptOrgId",{"_33":61,"_13":14,"_45":147,"_63":-5,"_64":53,"_65":66,"_67":66,"_68":-5,"_69":-5,"_70":71,"_72":-5,"_73":53,"_74":-5,"_75":-5,"_76":77,"_78":79,"_80":-5,"_81":-5,"_82":-5,"_83":-5,"_84":85,"_86":85,"_87":88,"_89":88,"_90":88,"_91":85,"_92":85,"_93":148,"_95":149,"_97":150,"_99":53},[144],"files",[144],"counts",{"_135":77,"_136":47,"_137":77,"_31":138,"_139":140},[133],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":46,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_18":134},[35],"pulls","branches","forks",{"_141":77,"_143":77},"connections",{"_141":77,"_142":77},"count","unfilledCount","emptyCount",{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":57,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_58":145,"_18":146},[],[35],["D",1772328641971],{"_11":12,"_114":110},{"_11":12,"_108":-5,"_109":110,"_70":71,"_111":112,"_113":53},{"_17":42,"_13":16,"_33":40,"_100":61,"_101":-5,"_102":-5,"_103":-5,"_104":-5,"_105":-5,"_106":14,"_107":12},"ctx",{"_93":-5,"_130":184,"_185":186,"_187":188,"_189":53,"_190":53,"_191":53,"_192":44,"_193":-5,"_194":-7,"_195":-7,"_196":77,"_197":53,"_198":53,"_199":-5,"_200":53,"_201":53,"_202":53,"_203":-5},"ENV",{"_155":156,"_157":158,"_159":16,"_160":161,"_162":163,"_164":165,"_166":167,"_168":169,"_170":171,"_172":173,"_174":175,"_176":177,"_178":179,"_180":181,"_182":183},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Vibe Coding AI Apps

Briefs from the Field