Daily Research Brief 2026-05-09

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Anthropic’s quarterly revenue soared 80x; a new SSA architecture claims 1000x less compute than Transformers at 5% of Opus cost; Node.js paused security bounties amid AI-generated fake reports; Anthropic cut agent runaway from 54% to 7% by onboarding AIs on employee handbooks first.

1. Latest arXiv Papers

1. CreativityBench: Evaluating Agent Creative Reasoning — https://arxiv.org/abs/2605.02910

2. Can RL Teach Long-Horizon Reasoning to LLMs? Experimental Evidence — https://arxiv.org/abs/2605.06638

3. Safety and Accuracy Follow Different Scaling Laws — https://arxiv.org/abs/2605.04039

4. Coherent Hierarchical Multi-Label Learning to Decompose — https://arxiv.org/abs/2605.02734

5. Perceptual Flow Network for Visually Grounded Reasoning — https://arxiv.org/abs/2605.02730

6. Augmenting Interface Usability Heuristics for Real-World Agents — https://arxiv.org/abs/2605.02729

7. SpecKV: Speculative Decoding with Key-Value Caching — https://arxiv.org/abs/2605.02888

8. BARD-VL: Bridging Autoregressive and Diffusion Vision-Language Models — https://arxiv.org/abs/2604.16514