[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_152":153,"_154":155},"routes/_app.x.$handle.$name",{"_93":92,"_32":45,"_101":83,"_27":102,"_13":14,"_103":-5,"_11":104,"_105":106,"_107":108,"_109":110,"_111":112,"_20":-5,"_113":-5,"_114":-5,"_115":25,"_116":117,"_118":119,"_120":-5,"_121":-5,"_122":123},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_43":44,"_32":45,"_38":46,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":-5,"_53":-5,"_18":54,"_55":-5,"_56":25,"_57":-5,"_58":-5,"_59":60,"_61":62,"_63":-5,"_64":-5,"_65":-5,"_66":-5,"_67":68,"_69":68,"_70":71,"_72":71,"_73":71,"_74":68,"_75":68,"_76":77,"_78":79,"_80":81,"_82":25},"version",3,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":25,"_27":28,"_29":28},"file",{"_30":31,"_32":28,"_33":34,"_35":37,"_38":39,"_40":14,"_41":42,"_24":25},"previous",{"_30":31,"_32":28,"_33":34,"_35":36},"deleted",false,"deletedHere","path","mod.ts","latestPath","fileId","b869936e-0f6b-11f1-a6d7-42dde27851f2","name","fileType","http","content","export default async function(req: Request) {\n const html = `\n\n\n\n \n \n The DAIR Daily - AI Papers of the Week\n \n \n\n\n

The DAIR Daily

\n Vol. III — No. 7\n Saturday, February 21, 2026\n Price: One Token\n

\n\n

ALMA: Meta-Learning Memory for Agents

\n In a groundbreaking shift for agentic systems, researchers from Jeff Clune’s group have unveiled ALMA (Automated meta-Learning of Memory designs for Agentic systems). The system introduces a \"Meta Agent\" that autonomously discovers memory architectures through open-ended exploration in code space. By searching over database schemas, retrieval mechanisms, and update strategies expressed as executable code, ALMA consistently outperforms all human-designed memory baselines across four sequential decision-making benchmarks. This represents a significant step toward AI systems that learn to be continual learners.\n

\n Read the full report at arXiv.org →\n

\n\n

LLaDA 2.1: Speed Meets Quality

\n Ant Group has released LLaDA 2.1, a major upgrade to discrete diffusion language models. By introducing Token-to-Token (T2T) editing, the model allows for aggressive drafting followed by refinement. The \"Flash\" variant hits an astounding 892 tokens per second, fundamentally changing the speed-quality trade-off in LLM decoding.\n

\n Source: arXiv:2602.08676 →\n

\n\n

SkillRL: Recursive Skill Discovery

\n SkillRL introduces a recursive framework that distills raw experience into reusable behavioral patterns. This \"SkillBank\" approach reduces token footprints while preserving reasoning utility, achieving state-of-the-art performance with an 89.9% success rate on the ALFWorld benchmark.\n

\n Source: arXiv:2602.08234 →\n

\n\n

InftyThink+: Infinite Reasoning

\n Addressing the quadratic cost of long chain-of-thought reasoning, InftyThink+ uses trajectory-level RL to let models autonomously decide when to summarize and resume. The framework yielded a 21% accuracy gain on AIME24, outperforming vanilla long-CoT methods.\n

\n Source: arXiv:2602.06960 →\n

\n\n

Agyn: Software Engineering as an Org

\n Agyn models software development as an organizational process with specialized agents (Manager, Researcher, Engineer, Reviewer). It resolves 72.2% of SWE-bench tasks without specific tuning, suggesting organizational design matters as much as model size.\n

\n Source: arXiv:2602.01465 →\n

\n\n

Research Briefs

EchoJEPA

A foundation model for echocardiography trained on 18M scans, showing 20% improvement in cardiac assessment and exceptional robustness to noise.

\n arXiv:2602.02603 →\n

AdaptEvolve

Confidence-driven model routing cuts evolutionary agent costs by 38% while retaining 97.5% accuracy, making iterative refinement viable for production.

\n arXiv:2602.11931 →\n

Gaia2

Meta FAIR's new dynamic agent benchmark forces agents to handle temporal pressure and multi-agent coordination in changing environments.

\n arXiv:2602.11964 →\n

AgentArk

Distilling multi-agent debate into a single LLM for 4.8% better reasoning, approaching full multi-agent performance at a fraction of the cost.

\n arXiv:2602.03955 →\n

AgentSkiller

A 14B model beating GPT-o3 on tool-use benchmarks via high-quality synthetic trajectories, proving data quality over parameter count.

\n arXiv:2602.09372 →\n

\n \n

\n\n\n \\`;\n\n return new Response(html, {\n headers: { \"Content-Type\": \"text/html\" },\n });\n}\n","export default async function(req: Request) {\n const html = `\n\n\n\n \n \n The DAIR Daily - AI Papers of the Week\n \n \n\n\n

The DAIR Daily

\n Vol. III — No. 7\n Saturday, February 21, 2026\n Price: One Token\n

\n\n

ALMA: Meta-Learning Memory for Agents

\n Read the full report at arXiv.org →\n

\n\n

LLaDA 2.1: Speed Meets Quality

\n Source: arXiv:2602.08676 →\n

\n\n

SkillRL: Recursive Skill Discovery

\n Source: arXiv:2602.08234 →\n

\n\n

InftyThink+: Infinite Reasoning

\n Source: arXiv:2602.06960 →\n

\n\n

Agyn: Software Engineering as an Org

\n Source: arXiv:2602.01465 →\n

\n\n

Research Briefs

EchoJEPA

A foundation model for echocardiography trained on 18M scans, showing 20% improvement in cardiac assessment and exceptional robustness to noise.

\n arXiv:2602.02603 →\n

AdaptEvolve

Confidence-driven model routing cuts evolutionary agent costs by 38% while retaining 97.5% accuracy, making iterative refinement viable for production.

\n arXiv:2602.11931 →\n

Gaia2

Meta FAIR's new dynamic agent benchmark forces agents to handle temporal pressure and multi-agent coordination in changing environments.

\n arXiv:2602.11964 →\n

AgentArk

Distilling multi-agent debate into a single LLM for 4.8% better reasoning, approaching full multi-agent performance at a fraction of the cost.

\n arXiv:2602.03955 →\n

AgentSkiller

A 14B model beating GPT-o3 on tool-use benchmarks via high-quality synthetic trajectories, proving data quality over parameter count.

\n arXiv:2602.09372 →\n

\n \n

\n\n\n`;\n\n return new Response(html, {\n headers: { \"Content-Type\": \"text/html\" },\n });\n}\n","createdAt",["D",1771709870943],"branchVersion","branchId","b846e468-0f6b-11f1-82b1-42dde27851f2","id","b8469e72-0f6b-11f1-82b1-42dde27851f2","dair_daily",["D",1771709075271],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","imageUrl","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",{"_93":92,"_100":96},"org",{"_93":92,"_94":-5,"_95":96,"_18":54,"_97":98,"_99":25},"branch",{"_13":14,"_32":83,"_43":42,"_84":44,"_85":-5,"_86":-5,"_87":-5,"_88":-5,"_89":-5,"_90":45,"_91":92},"isOwned","main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","danielpcox","handle","displayName","avatar","https://img.clerk.com/eyJ0eXBlIjoicHJveHkiLCJzcmMiOiJodHRwczovL2ltYWdlcy5jbGVyay5kZXYvb2F1dGhfZ2l0aHViL2ltZ18zOWpmZkZlTFFUYkJFSlE4R291RFhqVE1ma0UifQ","tier","free","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_43":44,"_32":45,"_38":148,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":-5,"_53":-5,"_18":54,"_55":-5,"_56":25,"_57":-5,"_58":-5,"_59":60,"_61":62,"_63":-5,"_64":-5,"_65":-5,"_66":-5,"_67":68,"_69":68,"_70":71,"_72":71,"_73":71,"_74":68,"_75":68,"_76":149,"_78":150,"_80":151,"_82":25},"tree",[143],"files",[143],"counts",{"_132":60,"_133":134,"_135":60,"_136":137,"_138":139},"contents",[124],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles",true,"grants","orgs","customDomains",{},{"_43":125,"_32":28,"_30":31,"_126":-5,"_41":42,"_40":14,"_127":119,"_38":128,"_13":129,"_33":34,"_35":37,"_24":25,"_130":-5,"_27":131},"928a69c8-0f6d-11f1-a671-42dde27851f2","parentId","latest",["D",1771709870943],2,"label",[28],"pulls","branches",1,"forks","envVars",{"_140":60,"_142":60},"connections",{"_140":60,"_141":60},"count","unfilledCount","emptyCount",{"_43":125,"_32":28,"_30":31,"_126":-5,"_41":42,"_40":14,"_127":119,"_38":144,"_13":129,"_33":34,"_35":37,"_24":25,"_130":-5,"_145":146,"_27":147},["D",1771709870943],"children",[],[28],["D",1771709075271],{"_93":92,"_100":96},{"_93":92,"_94":-5,"_95":96,"_18":54,"_97":98,"_99":25},{"_13":14,"_32":83,"_43":42,"_84":44,"_85":-5,"_86":-5,"_87":-5,"_88":-5,"_89":-5,"_90":45,"_91":92},"ctx",{"_76":-5,"_121":185,"_186":187,"_188":189,"_190":25,"_191":25,"_192":25,"_193":119,"_194":-5,"_195":-7,"_196":-7,"_197":60,"_198":25,"_199":25,"_200":-5,"_201":25,"_202":25,"_203":25,"_204":-5},"ENV",{"_156":157,"_158":159,"_160":83,"_161":162,"_163":164,"_165":166,"_167":168,"_169":170,"_171":172,"_173":174,"_175":176,"_177":178,"_179":180,"_181":182,"_183":184},"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"]

Diff · v3