{
  "title": "Daily Research Brief 2026-03-29",
  "url": "/en/posts/research-brief-2026-03-29/",
  "permalink": "https://hackcv.com/en/posts/research-brief-2026-03-29/",
  "date": "2026-03-29",
  "lastmod": "2026-03-29",
  "author": "",
  "description": "Daily research brief — AI / LLM / Agent / Computer Vision / Audio-Video / Engineering",
  "categories": ["Research Brief"],
  "tags": ["AI","LLM","Agent","Computer Vision","Audio-Video","Engineering","Daily Brief"],
  "cover": "https://picsum.photos/seed/daily-research-brief-2026-03-29/1200/675",
  "readingTime": 3,
  "wordCount": 623,
  "content": "\u003ch1 id=\"daily-research-brief-2026-03-29\"\u003eDaily Research Brief 2026-03-29\u003c/h1\u003e\n\u003cp\u003e📊 Token usage: estimated from retrieval and writing scale.\u003c/p\u003e\n\u003cp\u003eCovers the latest AI research, open source and industry moves, updated daily.\u003c/p\u003e\n\u003chr\u003e\n\u003ch2 id=\"editors-note\"\u003eEditor\u0026rsquo;s Note\u003c/h2\u003e\n\u003cp\u003eSunday issue (arXiv API rate-limited): the themes converge on OCR-free document understanding (VLM-native, no layout detection + character recognition), LLM cost optimization as a system (routing/caching/prompt efficiency), and self-evolving agent frameworks (HKU OpenSpace). GitHub trending is led by Dify and AutoGPT; HackerNews debates who monetizes open-source AI models.\u003c/p\u003e\n\u003ch2 id=\"1-latest-arxiv-papers-representative-directions-api-rate-limited\"\u003e1. Latest arXiv Papers (representative directions, API rate-limited)\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eVision-Based RAG for Long Documents\u003c/strong\u003e — explores VLMs (e.g. GPT-4.1) processing PDF images directly without traditional OCR, combined with a reasoning retrieval layer for long-document QA. — \u003ca href=\"https://github.com/VectifyAI/PageIndex/blob/main/cookbook/vision_RAG_pageindex.ipynb\"\u003ehttps://github.com/VectifyAI/PageIndex/blob/main/cookbook/vision_RAG_pageindex.ipynb\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eEnd-to-End Vision-Language Models for OCR-Free Document Understanding\u003c/strong\u003e — a new end-to-end document understanding paradigm where the VLM jointly understands visual + textual information, bypassing the two-stage layout detection + character recognition pipeline. (arXiv ID redacted in the original — search: VLM OCR-free document understanding)\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eLLM Cost Optimization as a System: Routing, Caching \u0026amp; Prompt Efficiency\u003c/strong\u003e — systematically addresses cost and latency in LLM production, covering prompt compression, model routing, RAG efficiency and agent workflow control. — \u003ca href=\"https://argmin.ai\"\u003ehttps://argmin.ai\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eOpenSpace: Self-Evolving AI Agents Framework\u003c/strong\u003e — open-source project by HKU HKUDS, focused on making agents smarter, lower-cost and self-evolving. — \u003ca href=\"https://github.com/HKUDS/OpenSpace\"\u003ehttps://github.com/HKUDS/OpenSpace\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eReasoning-Based Retrieval for Multimodal Agents\u003c/strong\u003e — a new reasoning retrieval layer for multimodal agents, improving information-gathering efficiency and answer accuracy in complex tasks. (arXiv ID redacted in the original — search: reasoning retrieval multimodal agent)\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch2 id=\"2-hot-github-open-source\"\u003e2. Hot GitHub Open Source\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003e🏆 \u003cstrong\u003eDify — Agentic Workflow development platform\u003c/strong\u003e — ⭐ 134k | TypeScript. Open-source LLM app development platform supporting agent orchestration, RAG and workflow automation, one-stop from prototype to production. — \u003ca href=\"https://github.com/langgenius/dify\"\u003ehttps://github.com/langgenius/dify\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e🤖 \u003cstrong\u003eAutoGPT — accessible AI agent tooling\u003c/strong\u003e — ⭐ 130k+ | Python. A vision for letting everyone use and build AI products, continuously iterating. — \u003ca href=\"https://github.com/Significant-Gravitas/AutoGPT\"\u003ehttps://github.com/Significant-Gravitas/AutoGPT\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e📚 \u003cstrong\u003efunNLP — comprehensive Chinese NLP toolkit\u003c/strong\u003e — ⭐ 70k+ | Python. Huge collection of Chinese/English sensitive-word detection, NER, summarization, sentiment analysis, BERT/ERNIE resources, dialogue systems and more. — \u003ca href=\"https://github.com/fighting41love/funNLP\"\u003ehttps://github.com/fighting41love/funNLP\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e🌐 \u003cstrong\u003eOpenSpace — self-evolving AI agent framework\u003c/strong\u003e — new and active | Python. By HKUDS; focused on agent self-evolution and low-cost deployment; good for frontier agent architecture research. — \u003ca href=\"https://github.com/HKUDS/OpenSpace\"\u003ehttps://github.com/HKUDS/OpenSpace\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e💹 \u003cstrong\u003edaily_stock_analysis — LLM-driven stock analysis system\u003c/strong\u003e — new project | Python. Multi-source market data + real-time news + LLM decision dashboard + multi-channel push, zero-cost scheduled runs. — \u003ca href=\"https://github.com/ZhuLinsen/daily_stock_analysis\"\u003ehttps://github.com/ZhuLinsen/daily_stock_analysis\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch2 id=\"3-hackernews-top-posts\"\u003e3. HackerNews Top Posts\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eWho monetizes open-source AI models?\u003c/strong\u003e — who profits from open-source AI models; the tension between open ecosystems and commercial monetization. 💬 2 comments — \u003ca href=\"https://blog.kilocode.ai/p/who-monetizes-open-source-ai-models\"\u003ehttps://blog.kilocode.ai/p/who-monetizes-open-source-ai-models\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eWhy are the big labs open-sourcing AI models?\u003c/strong\u003e — the logic behind Meta/Google/NVIDIA open-sourcing: ecosystem building, standard setting or talent competition? 💬 3 comments — Ask HN thread\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eVision-Based Vectorless RAG for Long Documents\u003c/strong\u003e — document QA without OCR and without vector DB, built on GPT-4.1 multimodal reasoning; an implementation path worth watching. 💬 0 comments — \u003ca href=\"https://github.com/VectifyAI/PageIndex\"\u003ehttps://github.com/VectifyAI/PageIndex\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eOSSAIX — curated directory of open-source AI projects\u003c/strong\u003e — hand-maintained navigation of OSS AI tools covering LLM, RAG, agents, local AI, image/audio/video processing. 💬 2 comments — \u003ca href=\"https://ossaix.com\"\u003ehttps://ossaix.com\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eArgmin AI — LLM production cost optimization platform\u003c/strong\u003e — demo works but production cost spirals; Argmin does system-level prompt efficiency, model routing, caching and workflow optimization. 💬 HN discussion — \u003ca href=\"https://argmin.ai\"\u003ehttps://argmin.ai\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch2 id=\"4-deep-reads\"\u003e4. Deep Reads\u003c/h2\u003e\n\u003ch3 id=\"must-read-papers\"\u003eMust-read papers\u003c/h3\u003e\n\u003ctable\u003e\n\t\u003cthead\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003cth\u003ePaper\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eDirection\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eHighlight\u003c/th\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/thead\u003e\n\t\u003ctbody\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003eVision-Based Vectorless RAG\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003emultimodal RAG\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eupends traditional OCR flow; VLM natively understands documents\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003eOpenSpace: Self-Evolving Agents\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eagent self-evolution\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eHKUDS open new paradigm; agents iterate at low cost\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003eLLM Cost Optimization as System\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eengineering optimization\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eoptimizing LLM cost from a system view, not single points\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch3 id=\"recommended-tools\"\u003eRecommended tools\u003c/h3\u003e\n\u003ctable\u003e\n\t\u003cthead\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003cth\u003eTool\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eType\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eUse case\u003c/th\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/thead\u003e\n\t\u003ctbody\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003eDify\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003edev platform\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003erapid LLM app building / agent orchestration\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003eAutoGPT\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eagent framework\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eexperimental agent task automation\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003efunNLP\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003etoolkit\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003efull-scenario Chinese NLP development\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/tbody\u003e\n\u003c/table\u003e\n\u003chr\u003e\n",
  "summary": "Daily Research Brief 2026-03-29 📊 Token usage: estimated from retrieval and writing scale.\nCovers the latest AI research, open source and industry moves, updated daily.\nEditor\u0026rsquo;s Note Sunday issue (arXiv API rate-limited): the themes converge on OCR-free document understanding (VLM-native, no layout detection + character recognition), LLM cost optimization as a system (routing/caching/prompt efficiency), and self-evolving agent frameworks (HKU OpenSpace). GitHub trending is led by Dify and AutoGPT; HackerNews debates who monetizes open-source AI models.\n"
}