AI Briefing — April 19, 2026

🔬 Research

RationalRewards: Reasoning Rewards Scale Visual Generation Both Training and Test Time

99 upvotes · Haozhe Wang, Cong Wei, Weiming Ren, Jiaming Liu, Fangzhen Lin · Paper

This 8B parameter reward model teaches AI systems to produce multi-dimensional critiques before scoring visual outputs, transforming them from passive evaluators into active optimization tools. At training time, it provides interpretable rewards for reinforcement learning; at test time, it creates a Generate-Critique-Refine loop that improves outputs without parameter updates. The model achieves state-of-the-art preference prediction among open-source alternatives while using 10-20x less training data, and its test-time critique loop matches RL-based fine-tuning performance.

HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds

81 upvotes · Team HY-World, Chenjie Cao, Xuhui Zuo, Zhenwei Wang, Yisu Zhang · Paper

A comprehensive framework that takes text, single-view images, multi-view images, or videos and produces navigable 3D Gaussian Splatting scenes through a four-stage pipeline. Key innovations include upgraded panorama generation, trajectory planning with WorldNav, enhanced WorldStereo for keyframe-based view generation, and WorldMirror for universal 3D prediction. The system delivers performance comparable to the closed-source Marble model and includes WorldLens, a high-performance 3DGS rendering platform with interactive exploration capabilities.

DR³-Eval: Towards Realistic and Reproducible Deep Research Evaluation

28 upvotes · Qianqian Xie, Qingheng Xiong, He Zhu, Tiantian Xia, Xueming Han · Paper

A benchmark for evaluating Deep Research Agents on complex, long-horizon tasks involving planning, retrieval, and multimodal report generation. Built from authentic user materials with static research sandbox environments containing supportive documents, distractors, and noise, it measures Information Recall, Factual Accuracy, Citation Coverage, Instruction Following, and Depth Quality. Experiments reveal critical failure modes in retrieval robustness and hallucination control, making this a challenging testbed for research automation systems.

🛠 Tools

EvoMap/evolver — Self-Evolution Engine for AI Agents

1,131★/day · Repo

A Genome Evolution Protocol (GEP) powered system that enables AI agents to self-evolve through memory, skill, and evolution-asset management. The project has gained significant traction with over 5,000 stars but is transitioning from fully open source to source-available licensing due to concerns about similar systems being released without attribution. The system focuses on agent self-evolution capabilities with deeper GEP integration and stronger memory systems than alternatives.

BasedHardware/omi — AI Assistant for Screen and Conversation Monitoring

609★/day · Repo

An AI system that captures your screen and conversations, transcribes in real-time, generates summaries and action items, and provides an AI chat interface that remembers everything you've seen and heard. Trusted by 300,000+ professionals, it works across desktop, phone, and wearables with full open-source availability. The system offers quick deployment with cloud backend integration and supports both browser-based trials and local installation across multiple platforms.

Lordog/dive-into-llms — Large Language Model Programming Tutorial Series

547★/day · Repo

A comprehensive programming tutorial series titled "动手学大模型Dive into LLMs" delivered through Jupyter Notebooks. This Chinese-language educational resource provides hands-on programming practices for understanding and implementing large language models. The rapid daily star growth indicates strong interest in practical LLM education materials, particularly in the Chinese developer community.

📡 Analysis

Changes in the system prompt between Claude Opus 4.6 and 4.7

April 18 · Post · tags: ai, prompt-engineering, generative-ai, llms, anthropic, claude, ai-ethics, system-prompts

Anthropic updated Claude Opus 4.7's system prompt with several notable changes: expanded tool capabilities including Claude in PowerPoint, significantly enhanced child safety instructions with stricter conversation handling, reduced pushiness in conversation endings, and a new emphasis on acting rather than asking for clarification. The update also introduces tool-based ambiguity resolution, showing Claude's evolution toward more autonomous and safety-conscious behavior patterns.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net