{
  "title": "Daily Research Brief 2026-03-21/22",
  "url": "/en/posts/research-brief-2026-03-21-22/",
  "permalink": "https://hackcv.com/en/posts/research-brief-2026-03-21-22/",
  "date": "2026-03-20",
  "lastmod": "2026-03-20",
  "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-2026-03-21/22/1200/675",
  "readingTime": 3,
  "wordCount": 601,
  "content": "\u003ch1 id=\"daily-research-brief-2026-03-2122\"\u003eDaily Research Brief 2026-03-21/22\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\u003eWeekend issue covering 03-19–03-22: a strong CV/generation thread (VEGA-3D, Matryoshka Gaussian Splatting, Cubic Discrete Diffusion, EffectErase), LLM/Agent work (FinTradeBench, Nemotron-Cascade 2, F2LLM-v2), and a notable open-source surge — ClawTeam\u0026rsquo;s agent swarm (2,600+ stars in days). Note: Chinese community sources (Zhihu/Juejin) require login and were unavailable this period.\u003c/p\u003e\n\u003ch2 id=\"1-latest-arxiv-papers-03-19--03-22\"\u003e1. Latest arXiv Papers (03-19 ~ 03-22)\u003c/h2\u003e\n\u003ch3 id=\"cv--multimodal--generative-models\"\u003eCV / Multimodal / Generative Models\u003c/h3\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eVEGA-3D: Generation Models Know Space\u003c/strong\u003e — using the implicit 3D prior of video generation models for scene understanding; MLLMs are semantically strong but spatially blind.\u003cbr\u003e\n⭐ Worth reading: frontier fusion of CV + generative models + 3D scene understanding.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eMatryoshka Gaussian Splatting\u003c/strong\u003e — adjustable-fidelity scene rendering (Level of Detail) from a single model; highly significant for practical 3D GS deployment.\u003cbr\u003e\n⭐ Worth reading: engineering optimization + practical 3D reconstruction breakthrough.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eCubic Discrete Diffusion: Discrete Visual Generation\u003c/strong\u003e — combining discrete diffusion with high-dim representation tokens, unifying visual generation into the token-prediction paradigm of language models.\u003cbr\u003e\n⭐ Worth reading: new idea for unified multimodal token generation.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eEffectErase: Joint Video Object Removal and Insertion\u003c/strong\u003e — removing dynamic objects from video along with their visual effects (shadows, reflections) with high-quality restoration.\u003cbr\u003e\n⭐ Worth reading: practical video editing + CV engineering.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eSAMA: Factorized Semantic Anchoring and Motion Alignment\u003c/strong\u003e — tackles the balance between semantic modification and motion preservation in instruction-guided video editing.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch3 id=\"llm--agent--optimization\"\u003eLLM / Agent / Optimization\u003c/h3\u003e\n\u003col start=\"6\"\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eFinTradeBench: Financial Reasoning Benchmark for LLMs\u003c/strong\u003e — a financial decision-reasoning benchmark requiring synthesis of company fundamentals and heterogeneous signals.\u003cbr\u003e\n⭐ Worth reading: evaluation standard for LLM finance applications.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eNemotron-Cascade 2: Post-Training LLMs\u003c/strong\u003e — post-training via Cascade RL + multi-domain on-policy distillation.\u003cbr\u003e\n⭐ Worth reading: a new direction in LLM post-training.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eF2LLM-v2: Multilingual Embeddings (8 sizes, 80M-14B)\u003c/strong\u003e — multilingual general-purpose embedding models balancing inclusivity, performance and efficiency.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eDriveTok: 3D Driving Scene Tokenization\u003c/strong\u003e — a 3D driving-scene tokenization scheme unifying multi-view reconstruction and understanding.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eNot All Features Are Created Equal (VLA Models)\u003c/strong\u003e — mechanistic study of Vision-Language-Action models exposing feature-inequality issues.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch2 id=\"2-hot-github-open-source-past-week\"\u003e2. Hot GitHub Open Source (past week)\u003c/h2\u003e\n\u003ctable\u003e\n\t\u003cthead\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003cth\u003eProject\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003e⭐ Stars\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eNotes\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\u003eClawTeam / HKUDS\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e⭐ 2,602\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eAgent Swarm Intelligence — single command to full automation; HKU data-mining group\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\u003ewangziqi06/724-office\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e⭐ 530\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eSelf-evolving AI agent system — 26 tools, 3500 lines of pure Python, MCP / three-tier memory, self-repair\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\u003eNeoVertex1/nuggets\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e⭐ 315\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eFirst holographic-memory AI assistant\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\u003ehuggingface/hf-agents\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e⭐ 314\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eLocal coding agent driven by llama.cpp\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\u003emattprusak/autoresearch-genealogy\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e⭐ 856\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eStructured prompt templates for AI-assisted genealogy research\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/tbody\u003e\n\u003c/table\u003e\n\u003cblockquote\u003e\n\u003cp\u003e⭐ \u003cstrong\u003eWatchlist: ClawTeam\u003c/strong\u003e — multi-agent swarm framework, 2600+ stars, architecture worth studying.\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003ch2 id=\"3-hackernews-top-posts-03-20--03-22\"\u003e3. HackerNews Top Posts (03-20 ~ 03-22)\u003c/h2\u003e\n\u003ctable\u003e\n\t\u003cthead\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003cth\u003eHeat\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eTitle\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eLink\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🔥 116👍 24💬\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003ePatchwork – open-source framework automating dev chores\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://github.com/patched-codes/patchwork\"\u003egithub\u003c/a\u003e\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e10👍\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003ePomerium Agentic Access Gateway – dynamic auth for AI agents\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e—\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e6👍 2💬\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eCheevly – natural-language IDE for building collaborative AI agents\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e—\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e5👍\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003enanochat rewritten in C++ (ggml inference)\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://github.com/k-ye/nanochagg.ml\"\u003egithub\u003c/a\u003e\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e5👍 2💬\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eClaude Sonnet 4.5 free (ad-supported)\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e—\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e4👍\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eForge – 3MB Rust binary orchestrating multiple AI coding agents\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://github.com/nxtg-ai/forge-orchestrator\"\u003egithub\u003c/a\u003e\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2 id=\"4-deep-reads\"\u003e4. Deep Reads\u003c/h2\u003e\n\u003ctable\u003e\n\t\u003cthead\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003cth\u003ePriority\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eItem\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eDirection\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🌟\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003eVEGA-3D\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eCV + generative models + 3D scene understanding\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e🌟\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003eCubic Discrete Diffusion\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eunified multimodal token generation\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e🌟\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003eClawTeam\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003ereference multi-agent swarm project\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e💡\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003eFinTradeBench\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eLLM financial reasoning evaluation\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\t\t\u003ctr\u003e\n\t\t\t\t\t\u003ctd\u003e💡\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003cstrong\u003e724-office\u003c/strong\u003e\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eself-evolving AI agent engineering practice\u003c/td\u003e\n\t\t\t\u003c/tr\u003e\n\t\u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2 id=\"notes\"\u003eNotes\u003c/h2\u003e\n\u003cul\u003e\n\u003cli\u003eZhihu / Juejin Chinese communities require login for their APIs, so auto-collection was unavailable this period.\u003c/li\u003e\n\u003cli\u003eGitHub Trending direct connection is sometimes unstable; the fallback uses the GitHub API to search for newly popular projects over the past week.\u003c/li\u003e\n\u003cli\u003eHackerNews posts are filtered for AI/ML relevance via the Algolia API.\u003c/li\u003e\n\u003c/ul\u003e\n\u003chr\u003e\n",
  "summary": "Daily Research Brief 2026-03-21/22 📊 Token usage: estimated from retrieval and writing scale.\nCovers the latest AI research, open source and industry moves, updated daily.\nEditor\u0026rsquo;s Note Weekend issue covering 03-19–03-22: a strong CV/generation thread (VEGA-3D, Matryoshka Gaussian Splatting, Cubic Discrete Diffusion, EffectErase), LLM/Agent work (FinTradeBench, Nemotron-Cascade 2, F2LLM-v2), and a notable open-source surge — ClawTeam\u0026rsquo;s agent swarm (2,600+ stars in days). Note: Chinese community sources (Zhihu/Juejin) require login and were unavailable this period.\n"
}