25 October 2026

AI · Protein design

RFD4-Proteina: protein design that thinks longer (preprint)

The Institute for Protein Design and NVIDIA released RFD4-Proteina, an open model that can spend more computing on hard design problems. It succeeds five to ten times more often, and runs five to ten times faster. With Manchester, it designed enzymes for making an HIV prevention drug. It is still a preprint.

• Researchers achieve major advance in protein design UW Medicine Newsroom · Primary · RFD4-Proteina, preprint
• Scientists demonstrate AI-designed enzyme for drug synthesis University of Manchester via Phys.org · 23 Oct 2026 · Primary · 23 Oct 2026; Corley et al., "Biomolecular design through latent reasoning with RFD4-Proteina", preprint

AI · Agents

Plain RAG beats complex agent memory

Do AI agents need complex memory? Researchers in China tested plain retrieval and reranking over the raw conversation history. It ranked first in six of seven benchmark settings, beating the strongest agentic memory system by up to 5.5 points, because it keeps more of the evidence.

• Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought arXiv · 22 Oct 2026 · Primary · Bao et al. (2026)

AI · Reasoning

Reasoning models inherit their 'wait'

Reasoning models often write 'wait' before rechecking a step, usually credited to reinforcement learning. A Hong Kong team traced the word through open models' training. Reinforcement learning from scratch almost never coins it; it arrives with training data. So counting 'wait' is no proof of learned reasoning.

• Inherited, Not Invented: Tracing the Provenance of Wait in Reasoning Models arXiv · 22 Oct 2026 · Primary · Kou et al. (2026)

AI · Weather

Machine learning doubles six-week forecasts

Forecasts two to six weeks ahead are notoriously hard. A team including MIT and Microsoft built a machine-learning layer that learns a forecast model's past errors and corrects them. In a European competition, it doubled the model's predictive power and beat climate averages. Published in Nature.

• Advancing subseasonal forecasting with machine learning Nature · 2026 · Primary · Guan et al. (2026); Nature
• An AI tool improves weather forecasts up to a month ahead Nature (Research Briefing)
• Researchers improve weather forecasts looking two to six weeks out MIT News · 2026

Idea · Language models

Lost in the middle

Lost in the middle. Language models use information best at the start or end of their context, and worst in the middle. In a 2023 test with twenty documents, accuracy dropped when the right one sat in the middle. That is one reason reranking, which puts the best evidence first, still matters.

Liu et al. 2023 · arXiv:2307.03172

• Lost in the Middle: How Language Models Use Long Contexts arXiv:2307.03172 · Jul 2023 · Primary · Liu et al. (2023)

Idea · Training

RL draws out, it does not create

Reinforcement learning with verifiable rewards makes a model likelier to find a correct reasoning path on its first try. But a 2025 study found those paths were already in the base model: given many attempts, it solves as many problems, or more. Training sharpens what is there, like today's 'wait'.

Yue et al. 2025 · arXiv:2504.13837

• Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? arXiv:2504.13837 · Apr 2025 · Primary · Yue et al. (2025)

Idea · Protein design

Designing proteins by removing noise

Diffusion models start from random noise and remove it step by step, the way image generators sharpen a picture. RFdiffusion applied this to protein backbones: start from scrambled coordinates, then denoise toward a structure that meets a goal, such as binding a target. It is an ancestor of today's RFD4-Proteina.

Watson et al. 2023 · Nature

• De novo design of protein structure and function with RFdiffusion Nature · 2023 · Primary · Watson et al. (2023)