Daily Research Brief 2026-03-21/22

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’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’s agent swarm (2,600+ stars in days). Note: Chinese community sources (Zhihu/Juejin) require login and were unavailable this period.

1. Latest arXiv Papers (03-19 ~ 03-22)

CV / Multimodal / Generative Models

1. VEGA-3D: Generation Models Know Space — using the implicit 3D prior of video generation models for scene understanding; MLLMs are semantically strong but spatially blind.
   ⭐ Worth reading: frontier fusion of CV + generative models + 3D scene understanding.

2. Matryoshka Gaussian Splatting — adjustable-fidelity scene rendering (Level of Detail) from a single model; highly significant for practical 3D GS deployment.
   ⭐ Worth reading: engineering optimization + practical 3D reconstruction breakthrough.

3. Cubic Discrete Diffusion: Discrete Visual Generation — combining discrete diffusion with high-dim representation tokens, unifying visual generation into the token-prediction paradigm of language models.
   ⭐ Worth reading: new idea for unified multimodal token generation.

4. EffectErase: Joint Video Object Removal and Insertion — removing dynamic objects from video along with their visual effects (shadows, reflections) with high-quality restoration.
   ⭐ Worth reading: practical video editing + CV engineering.

5. SAMA: Factorized Semantic Anchoring and Motion Alignment — tackles the balance between semantic modification and motion preservation in instruction-guided video editing.

LLM / Agent / Optimization

6. FinTradeBench: Financial Reasoning Benchmark for LLMs — a financial decision-reasoning benchmark requiring synthesis of company fundamentals and heterogeneous signals.
   ⭐ Worth reading: evaluation standard for LLM finance applications.

7. Nemotron-Cascade 2: Post-Training LLMs — post-training via Cascade RL + multi-domain on-policy distillation.
   ⭐ Worth reading: a new direction in LLM post-training.

8. F2LLM-v2: Multilingual Embeddings (8 sizes, 80M-14B) — multilingual general-purpose embedding models balancing inclusivity, performance and efficiency.

9. DriveTok: 3D Driving Scene Tokenization — a 3D driving-scene tokenization scheme unifying multi-view reconstruction and understanding.

10. Not All Features Are Created Equal (VLA Models) — mechanistic study of Vision-Language-Action models exposing feature-inequality issues.

2. Hot GitHub Open Source (past week)

⭐ Watchlist: ClawTeam — multi-agent swarm framework, 2600+ stars, architecture worth studying.

3. HackerNews Top Posts (03-20 ~ 03-22)

4. Deep Reads

Notes

• Zhihu / Juejin Chinese communities require login for their APIs, so auto-collection was unavailable this period.
• GitHub Trending direct connection is sometimes unstable; the fallback uses the GitHub API to search for newly popular projects over the past week.
• HackerNews posts are filtered for AI/ML relevance via the Algolia API.