Daily Research Brief 2026-05-06

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

OpenAI shipped GPT-5.5 Instant; Zhipu’s Kimi K2.6 beat Claude, GPT-5.5 and Gemini in a coding challenge; DeepClaude cut cost to 1/17 of native Claude Code; DeepSeek V4 is ‘almost frontier’; Harvard found AI ER diagnosis at 67% accuracy, first time above human doctors.

1. Latest arXiv Papers

1. NEURON: A Neuro-Symbolic System for Grounded Clinical Reasoning — https://arxiv.org/abs/2605.01189

2. LLMs Should Not Yet Be Credited with Decision Expertise — https://arxiv.org/abs/2605.01164

3. Arithmetic in the Wild: Llama Uses Base-10 Addition — https://arxiv.org/abs/2605.01148

4. CLEAR: Revealing How Noise and Ambiguity Degrade Reasoning — https://arxiv.org/abs/2605.01011

5. Can AI Debias the News? LLM Interventions Improve Political Neutrality — https://arxiv.org/abs/2605.01006

6. ADAPTS: Agentic Decomposition for Automated Prototype Search — https://arxiv.org/abs/2605.03212

7. Model Organisms Are Leaky: Perplexity Differences in LLM Evaluation — https://arxiv.org/abs/2605.00994

8. SubQ 1M-Preview: The World’s First Sub-Quadratic LLM — https://arxiv.org/abs/2605.02897