Daily Research Brief 2026-04-04

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

The paper thread today: DOne design-to-code generation, test-time scaling making overtraining compute-optimal, unified 3D reconstruction and generation (UniRecGen), and a new codec language model for TTS.

1. Latest arXiv Papers

1. DOne: A High-Fidelity Decoupled Framework for Design-to-Code Generation — https://arxiv.org/abs/2604.01226

2. Test-Time Scaling Makes Overtraining Compute-Optimal — https://arxiv.org/abs/2604.01411

3. UniRecGen: Unified Multi-View 3D Reconstruction and Generation — https://arxiv.org/abs/2604.01479

4. T5Gemma-TTS: An Encoder-Decoder Codec Language Model for TTS — https://arxiv.org/abs/2604.01760

5. DySCo: Dynamic Semantic Compression for Long-Horizon Time-Series Forecasting — https://arxiv.org/abs/2604.01261