Daily Research Brief 2026-03-29

📊 Token usage: estimated from retrieval and writing scale.

Covers the latest AI research, open source and industry moves, updated daily.

Editor’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.

1. Latest arXiv Papers (representative directions, API rate-limited)

1. Vision-Based RAG for Long Documents — explores VLMs (e.g. GPT-4.1) processing PDF images directly without traditional OCR, combined with a reasoning retrieval layer for long-document QA. — https://github.com/VectifyAI/PageIndex/blob/main/cookbook/vision_RAG_pageindex.ipynb

2. End-to-End Vision-Language Models for OCR-Free Document Understanding — 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)

3. LLM Cost Optimization as a System: Routing, Caching & Prompt Efficiency — systematically addresses cost and latency in LLM production, covering prompt compression, model routing, RAG efficiency and agent workflow control. — https://argmin.ai

4. OpenSpace: Self-Evolving AI Agents Framework — open-source project by HKU HKUDS, focused on making agents smarter, lower-cost and self-evolving. — https://github.com/HKUDS/OpenSpace

5. Reasoning-Based Retrieval for Multimodal Agents — 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)

2. Hot GitHub Open Source

1. 🏆 Dify — Agentic Workflow development platform — ⭐ 134k | TypeScript. Open-source LLM app development platform supporting agent orchestration, RAG and workflow automation, one-stop from prototype to production. — https://github.com/langgenius/dify

2. 🤖 AutoGPT — accessible AI agent tooling — ⭐ 130k+ | Python. A vision for letting everyone use and build AI products, continuously iterating. — https://github.com/Significant-Gravitas/AutoGPT

3. 📚 funNLP — comprehensive Chinese NLP toolkit — ⭐ 70k+ | Python. Huge collection of Chinese/English sensitive-word detection, NER, summarization, sentiment analysis, BERT/ERNIE resources, dialogue systems and more. — https://github.com/fighting41love/funNLP

4. 🌐 OpenSpace — self-evolving AI agent framework — new and active | Python. By HKUDS; focused on agent self-evolution and low-cost deployment; good for frontier agent architecture research. — https://github.com/HKUDS/OpenSpace

5. 💹 daily_stock_analysis — LLM-driven stock analysis system — new project | Python. Multi-source market data + real-time news + LLM decision dashboard + multi-channel push, zero-cost scheduled runs. — https://github.com/ZhuLinsen/daily_stock_analysis

3. HackerNews Top Posts

1. Who monetizes open-source AI models? — who profits from open-source AI models; the tension between open ecosystems and commercial monetization. 💬 2 comments — https://blog.kilocode.ai/p/who-monetizes-open-source-ai-models

2. Why are the big labs open-sourcing AI models? — the logic behind Meta/Google/NVIDIA open-sourcing: ecosystem building, standard setting or talent competition? 💬 3 comments — Ask HN thread

3. Vision-Based Vectorless RAG for Long Documents — document QA without OCR and without vector DB, built on GPT-4.1 multimodal reasoning; an implementation path worth watching. 💬 0 comments — https://github.com/VectifyAI/PageIndex

4. OSSAIX — curated directory of open-source AI projects — hand-maintained navigation of OSS AI tools covering LLM, RAG, agents, local AI, image/audio/video processing. 💬 2 comments — https://ossaix.com

5. Argmin AI — LLM production cost optimization platform — demo works but production cost spirals; Argmin does system-level prompt efficiency, model routing, caching and workflow optimization. 💬 HN discussion — https://argmin.ai

4. Deep Reads

Must-read papers

Recommended tools