{
  "title": "Daily Research Brief 2026-03-30",
  "url": "/en/posts/research-brief-2026-03-30/",
  "permalink": "https://hackcv.com/en/posts/research-brief-2026-03-30/",
  "date": "2026-03-30",
  "lastmod": "2026-03-30",
  "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-2026-03-30/1200/675",
  "readingTime": 3,
  "wordCount": 725,
  "content": "\u003ch1 id=\"daily-research-brief-2026-03-30\"\u003eDaily Research Brief 2026-03-30\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\u003eToday\u0026rsquo;s papers: autoregressive 3D Gaussian scene generation (GaussianGPT), world-consistent video generation via 4D latent reward (VGGRPO), a perception-centric video reasoning benchmark (PerceptionComp), geometry-aware spatial reasoning for VLMs (GeoSR), and zero-shot depth-from-defocus (FOSSA). GitHub trending stays agent-centric (AutoGPT, ollama, transformers, Langflow, Dify); HackerNews highlights agent behavior caching (Muscle-Mem), vision-native E2E testing (Magnitude), and a 48-hour agent red-team methodology.\u003c/p\u003e\n\u003ch2 id=\"1-latest-arxiv-papers\"\u003e1. Latest arXiv Papers\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eGaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation\u003c/strong\u003e — a Transformer-based autoregressive 3D Gaussian scene generation model producing 3D Gaussians via next-token prediction. Uses a sparse 3D convolutional autoencoder + vector quantization to compress Gaussian primitives; supports scene completion, outpainting and controllable sampling. Complementary to diffusion models; natively supports context-aware 3D generation. — \u003ca href=\"https://arxiv.org/abs/2603.26661\"\u003ehttps://arxiv.org/abs/2603.26661\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eVGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward\u003c/strong\u003e — applies GRPO in latent space for geometry-consistent video post-training. A Latent Geometry Model (LGM) decodes scene geometry directly from latents, avoiding expensive VAE decoding; supports dynamic scenes with dual constraints (camera-motion smoothing reward + geometric reprojection consistency reward). — \u003ca href=\"https://arxiv.org/abs/2603.26599\"\u003ehttps://arxiv.org/abs/2603.26599\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003ePerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning\u003c/strong\u003e — a video benchmark requiring multi-time-span visual evidence + compositional logical reasoning (1,114 questions / 279 videos). The strongest model, Gemini-3-Flash, reaches only 45.96%; open models stay below 40% — perception-centric long-horizon reasoning remains a bottleneck. — \u003ca href=\"https://arxiv.org/abs/2603.26653\"\u003ehttps://arxiv.org/abs/2603.26653\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eGeoSR: Make Geometry Matter for Spatial Reasoning\u003c/strong\u003e — via Geometry-Unleashing Masking (masking 2D visual tokens to force reliance on geometry tokens) and Geometry-Guided Fusion (gated routing that adaptively amplifies geometry contributions), lets VLMs truly use 3D geometry for spatial reasoning; SOTA on static and dynamic benchmarks. — \u003ca href=\"https://arxiv.org/abs/2603.26639\"\u003ehttps://arxiv.org/abs/2603.26639\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eZero-Shot Depth from Defocus (FOSSA)\u003c/strong\u003e — a Transformer-based zero-shot focal-stack depth-from-defocus network; the core is a stack attention layer with focal-distance embeddings enabling efficient information exchange across the focal stack. Also releases the ZEDD benchmark (8.3x more scenes than its predecessor), cutting error by up to 55.7%. — \u003ca href=\"https://arxiv.org/abs/2603.26658\"\u003ehttps://arxiv.org/abs/2603.26658\u003c/a\u003e\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\u003ca href=\"https://github.com/Significant-Gravitas/AutoGPT\"\u003eAutoGPT\u003c/a\u003e ⭐ 182.9k — Python. Autonomous AI agent platform supporting agentic workflow building and deployment; very actively updated.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/ollama/ollama\"\u003eollama/ollama\u003c/a\u003e ⭐ 166.5k — Go. The standard tool for running LLMs locally; latest support for Kimi-K2.5, GLM-5, MiniMax, DeepSeek, Qwen and more.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/huggingface/transformers\"\u003ehuggingface/transformers\u003c/a\u003e ⭐ 158.6k — Python. SOTA model-definition framework covering text/vision/audio/multimodal, continuously integrating the newest models (Gemma3, GLM, Qwen\u0026hellip;).\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/langflow-ai/langflow\"\u003elangflow-ai/langflow\u003c/a\u003e ⭐ 146.4k — Python. Visual AI agent and workflow building platform; low-code drag-and-drop multi-agent apps.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/langgenius/dify\"\u003elanggenius/dify\u003c/a\u003e ⭐ 135k — TypeScript/Python. Production-grade agentic workflow platform with RAG, MCP and multi-model orchestration; active pushes today.\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\u003ca href=\"https://github.com/pig-dot-dev/muscle-mem\"\u003eShow HN: Muscle-Mem — behavior cache for AI agents\u003c/a\u003e ⭐ 226 points · 51 comments — caches agent tool-calling traces like a JIT compiler; repeated tasks take deterministic replay, switching back to agent mode only on edge cases. A pragmatic answer to the $40/hr token cost of pure-vision agents.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/magnitudedev/magnitude\"\u003eShow HN: Magnitude — AI-native web testing framework\u003c/a\u003e ⭐ 179 points · 44 comments — replaces set-of-marks with a pure-vision VLM (Moondream), dual-agent architecture (planner + executor); a faster, cheaper E2E testing framework than browser-use.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://www.ninjaflix.ai/\"\u003eShow HN: AI agents built and shipped an app in 36 hours for $270\u003c/a\u003e — hot post — 4 AI agents collaborated end-to-end from tech-stack choice to deployment, building a news-to-short-video platform (Sora 2 Pro + Veo 3.1); exposes real multi-agent problems: groupthink, hallucination, unstable video quality.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://tachyonicai.com/blog/how-to-red-team-ai-agent/\"\u003eShow HN: How to red-team your AI agent in 48 hours\u003c/a\u003e — 4-stage framework: reconnaissance → automated scanning → manual exploitation → verified report. Core insight: prompt injection → tool abuse → data exfiltration is the most common attack chain; indirect injection (RAG/web) is severely underestimated.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/liquidos-ai/Odyssey\"\u003eShow HN: Running AI agents across environments needs a proper solution\u003c/a\u003e — a Rust-built agent runtime addressing Python agent memory bloat, slow Docker startup and agent reuse; bundle-first packaging and cross-environment deployment.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\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\u003eType\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eTitle\u003c/th\u003e\n\t\t\t\t\t\u003cth\u003eWhy it matters\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\u003ePaper\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eVGGRPO\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003elatent-space GRPO for video geometric consistency; avoids VAE decode overhead; high engineering value\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://arxiv.org/abs/2603.26599\"\u003ehttps://arxiv.org/abs/2603.26599\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\u003ePaper\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eGeoSR\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003esystematic solution for VLM spatial reasoning; geometry-token utilization worth pondering\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://arxiv.org/abs/2603.26639\"\u003ehttps://arxiv.org/abs/2603.26639\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\u003ePaper\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003ePerceptionComp\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003enew video multimodal reasoning benchmark; strongest model only 46%, huge research space\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://arxiv.org/abs/2603.26653\"\u003ehttps://arxiv.org/abs/2603.26653\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\u003eProject\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eMuscle-Mem\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003epragmatic agent engineering; RPA + agent hybrid execution worth borrowing\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://github.com/pig-dot-dev/muscle-mem\"\u003ehttps://github.com/pig-dot-dev/muscle-mem\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\u003eArticle\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003eAI agent red-team methodology\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003esystematic framework for production-grade agent security; must-read for engineering\u003c/td\u003e\n\t\t\t\t\t\u003ctd\u003e\u003ca href=\"https://tachyonicai.com/blog/how-to-red-team-ai-agent/\"\u003ehttps://tachyonicai.com/blog/how-to-red-team-ai-agent/\u003c/a\u003e\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 2026-03-30 📊 Token usage: estimated from retrieval and writing scale.\nCovers the latest AI research, open source and industry moves, updated daily.\nEditor\u0026rsquo;s Note Today\u0026rsquo;s papers: autoregressive 3D Gaussian scene generation (GaussianGPT), world-consistent video generation via 4D latent reward (VGGRPO), a perception-centric video reasoning benchmark (PerceptionComp), geometry-aware spatial reasoning for VLMs (GeoSR), and zero-shot depth-from-defocus (FOSSA). GitHub trending stays agent-centric (AutoGPT, ollama, transformers, Langflow, Dify); HackerNews highlights agent behavior caching (Muscle-Mem), vision-native E2E testing (Magnitude), and a 48-hour agent red-team methodology.\n"
}