The DAIR Daily - AI Papers of the Week

[{"_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",{"_94":93,"_32":45,"_102":84,"_27":103,"_13":14,"_104":-5,"_11":105,"_106":107,"_108":109,"_110":111,"_112":113,"_20":-5,"_114":-5,"_115":-5,"_116":25,"_117":118,"_119":120,"_121":-5,"_122":-5,"_123":124},"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",2,"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","359bfc92-0f4d-11f1-9a28-42dde27851f2","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

\"All the Weights That's Fit to Print\"

\n \n \n

\n\n

SOCIETY OF THOUGHT: NEW REASONING MODELS SIMULATE INTERNAL DEBATE

\n\n

A breakthrough in Large Language Model (LLM) architecture has revealed that the most advanced reasoning models are no longer operating as simple predictive engines. Instead, researchers at DAIR.AI have identified a phenomenon dubbed the \"Internal Society of Thought.\"

The study reveals that these models simulate internal cognitive perspectives that debate and reconcile conflicting views before producing an output. This emergent behavior suggests that the future of AI reasoning lies not in monolithic logic, but in a form of computational collective intelligence that mimics human deliberation.

\n\n

BUSINESS & INDUSTRY

AI IDEs vs. Autonomous Agents: A longitudinal study of 4,700 resolved issues across LangChain and CrewAI reveals a startling trend: while autonomous agents speed up initial development, they often lead to an 18% increase in static-analysis warnings and a 35% jump in code complexity.

Physics Simulations Get a Voice: Researchers have introduced a self-correcting multi-agent framework that allows users to generate complex physics simulations using natural language, bridging the gap between abstract math and human intuition.

\n\n

IN BRIEF

\n Bio-Inspired Memory\n The Agent Cognitive Compressor (ACC) addresses \"cognitive drift\" in long workflows by replacing infinite transcripts with a bounded internal state.\n

\n Terminal-Bench 2.0\n A reality check for coding agents: even frontier models score below 65% on real-world CLI tasks. Visit tbench.ai.\n

\n Single-Agent Hypothesis\n New research suggests \"OneFlow\" algorithms can match multi-agent performance through iterative dialogue, saving significant KV cache resources.\n

\n\n

CLASSIFIEDS

LEARN TO VIBECODE: Master production-grade AI apps.
dair-ai.thinkific.com

\n\n \n

\n \n \n `;\n\n return new Response(html, {\n headers: { \"Content-Type\": \"text/html\" },\n });\n}\n","import { Hono } from \"npm:hono\";\n\nconst app = new Hono();\n\nconst papers = [\n {\n title: \"ALMA: Meta-Learning Memory for Agents\",\n headline: \"Jeff Clune's Group Unveils ALMA: The Meta-Agent That Designs Its Own Memory\",\n summary: \"Researchers have introduced ALMA, a system that automatically discovers memory designs for AI agents through open-ended code exploration. By searching over database schemas and retrieval strategies, ALMA consistently outperforms human-engineered memory modules across multiple benchmarks.\",\n category: \"AGENTIC SYSTEMS\",\n link: \"https://arxiv.org/abs/2602.07755\"\n },\n {\n title: \"LLaDA 2.1\",\n headline: \"Ant Group Breaks Speed Records with LLaDA 2.1 Discrete Diffusion\",\n summary: \"LLaDA 2.1 introduces Token-to-Token (T2T) editing, allowing diffusion language models to correct errors in real-time. The 'Flash' variant hits nearly 900 tokens per second, dramatically outpacing traditional autoregressive models while maintaining top-tier quality.\",\n category: \"MODEL ARCHITECTURE\",\n link: \"https://arxiv.org/abs/2602.08676\"\n },\n {\n title: \"InftyThink+\",\n headline: \"Infinite-Horizon Reasoning: InftyThink+ Solves the 'Lost-in-the-Middle' Problem\",\n summary: \"InftyThink+ optimizes iterative reasoning trajectories by letting models autonomously decide when to summarize and compress their thoughts. This approach led to a massive 21% accuracy jump on the AIME24 math benchmark.\",\n category: \"REASONING\",\n link: \"https://arxiv.org/abs/2602.06960\"\n },\n {\n title: \"Agyn\",\n headline: \"Software Engineering as an Organization: The Agyn Multi-Agent Framework\",\n summary: \"Agyn models software development as a team of specialized agents (Manager, Researcher, Engineer, Reviewer). Designed for production use, it resolves 72% of SWE-bench tasks by mimicking real-world organizational workflows.\",\n category: \"SOFTWARE ENGINEERING\",\n link: \"https://arxiv.org/abs/2602.01465\"\n },\n {\n title: \"EchoJEPA\",\n headline: \"EchoJEPA: A Foundation Model for the Heart Trained on 18 Million Scans\",\n summary: \"By predicting in latent space rather than pixels, EchoJEPA ignores ultrasound noise to focus on anatomical signals. It outperforms existing cardiac assessment models by 20% and shows incredible robustness across patient populations.\",\n category: \"HEALTHCARE AI\",\n link: \"https://arxiv.org/abs/2602.02603\"\n }\n];\n\nconst inBrief = [\n { title: \"SkillRL\", desc: \"A recursive skill-augmented RL framework that distills raw experience into reusable behavioral patterns.\" },\n { title: \"AdaptEvolve\", desc: \"Cuts agentic inference costs by 38% using confidence-driven routing between small and large models.\" },\n { title: \"Gaia2\", desc: \"Meta FAIR's new benchmark for agents in dynamic, independently changing environments.\" },\n { title: \"AgentArk\", desc: \"Distills complex multi-agent debate dynamics into a single, high-performance LLM.\" },\n { title: \"AgentSkiller\", desc: \"A 14B model that beats GPT-o3 on tool-use benchmarks through high-quality synthetic data.\" }\n];\n\napp.get(\"/\", (c) => {\n const html = `\n \n \n \n \n \n The DAIR Daily - Feb 15, 2026\n \n \n \n \n

\n \n \n

\n\n

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

${papers[0].headline}

${papers[0].summary}

The discovery represents a significant shift away from hand-engineered memory modules. By allowing a Meta-Agent to explore the code space of database schemas and retrieval mechanisms, the system can tailor memory architectures to specific task domains automatically.

According to the paper from Jeff Clune's group, ALMA achieved a 12.3% average success rate with GPT-5-nano, significantly outperforming human-designed baselines. Read the full report »

\n\n

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

${papers[1].headline}

${papers[1].summary} [ArXiv]

\n ${papers[2].category}\n

${papers[2].headline}

${papers[2].summary} [ArXiv]

\n ${papers[3].category}\n

${papers[3].headline}

${papers[3].summary} [ArXiv]

\n ${papers[4].category}\n

${papers[4].headline}

${papers[4].summary} [ArXiv]

\n\n \n

\n \n \n `;\n return c.html(html);\n});\n\nexport default app.fetch;\n","createdAt",["D",1771696222348],"branchVersion","branchId","357618f6-0f4d-11f1-943b-42dde27851f2","id","3575a70e-0f4d-11f1-943b-42dde27851f2","dair_daily_newspaper",["D",1771695970894],"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",{"_94":93,"_101":97},"org",{"_94":93,"_95":-5,"_96":97,"_18":54,"_98":99,"_100":25},"branch",{"_13":83,"_32":84,"_43":42,"_85":44,"_86":-5,"_87":-5,"_88":-5,"_89":-5,"_90":-5,"_91":45,"_92":93},"isOwned",4,"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",{"_133":60,"_134":130,"_135":60,"_136":137,"_138":139},"contents",[125],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles",true,"grants","orgs","customDomains",{},{"_43":126,"_32":28,"_30":31,"_127":-5,"_41":42,"_40":14,"_128":25,"_38":129,"_13":130,"_33":34,"_35":37,"_24":25,"_131":-5,"_27":132},"cb5976b0-0f4d-11f1-b8e6-42dde27851f2","parentId","latest",["D",1771696222348],1,"label",[28],"pulls","branches","forks","envVars",{"_140":60,"_142":60},"connections",{"_140":60,"_141":60},"count","unfilledCount","emptyCount",{"_43":126,"_32":28,"_30":31,"_127":-5,"_41":42,"_40":14,"_128":25,"_38":144,"_13":130,"_33":34,"_35":37,"_24":25,"_131":-5,"_145":146,"_27":147},["D",1771696222348],"children",[],[28],["D",1771695970894],{"_94":93,"_101":97},{"_94":93,"_95":-5,"_96":97,"_18":54,"_98":99,"_100":25},{"_13":83,"_32":84,"_43":42,"_85":44,"_86":-5,"_87":-5,"_88":-5,"_89":-5,"_90":-5,"_91":45,"_92":93},"ctx",{"_76":-5,"_122":185,"_186":187,"_188":189,"_190":25,"_191":25,"_192":25,"_193":120,"_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":84,"_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 · v2