Skip to content

Research separate from news

What could become important next

A preprint is a signal, not a finished product or an independently confirmed result. We therefore track research papers separately and show the actual state of supporting evidence for each one.

1381 published research events

Latest work

A significant claim from a single source is published only after further confirmation.

Research only one source so far

MIT researchers develop a method to automatically convert 2D designs into 3D CAD models

MIT researchers and partners developed a system that teaches vision-language models to automatically convert 2D images into CAD programs. It generates training data from the model's own mistakes, improving accuracy and reducing computational requirements. The method may accelerate prototyping and lower development costs…

MIT News – Artificial intelligence Original source ↗
Research only one source so far

MIT develops neural transparency to inspect AI chatbots during design

MIT Media Lab researchers (Pat Pataranutaporn, Anthony Baez, Sheer Karny) introduced ‘neural transparency’, a tool allowing users to visualize AI model behavior before launch. It compares internal activations for different traits (empathy, honesty, toxicity, hallucinations…

MIT News – Artificial intelligence Original source ↗
Research only one source so far

The mathematics of diffusion models' creative potential

Google Research published a paper on why diffusion models generate new data rather than memorize training data. Their creative ability results from smoothing the score function and interpolating between training data. The paper was presented at ICLR 2026.

Google Research Blog Original source ↗
Research only one source so far

Generative SNUPI uses AI to accelerate DNA origami design

Researchers from Seoul National University and Hanyang University created Generative SNUPI, a generative model using a diffusion model to automate DNA origami design. Instead of manually creating DNA sequences, researchers can now specify a target shape (such as a dog, a star or Mona Lisa), and the model…

IEEE Spectrum — Artificial Intelligence Original source ↗
Research only one source so far

A foundation model handles tabular data for enterprise forecasting

An MIT team led by Devavrat Shah developed a foundation model for structured and time-series data. Licensed to spin-off company Ikigai Labs, it learns continuously from enterprise data and provides demand forecasts and planning for electronics and pharmaceuticals.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

The JARVIS Challenge tests the role of AI copilots in turbine engine design

The JARVIS Challenge at MIT tested whether AI could accelerate the design and construction of a jet engine. Seven student teams had four weeks to build an engine producing 50–100 pounds of thrust. Professor Spakovszky stated that AI accelerates safety-critical engineering, but engineering judgment remains…

MIT News – Artificial intelligence Original source ↗
Research only one source so far

New constitutional documentation for the Claude model

Anthropic published a new constitution — documentation of Claude's values and behavior. The document is written primarily for the model itself and forms a central part of the training process. Released under the CC0 1.0 license for free use anytime and anywhere.

Anthropic News Original source ↗
Research only one source so far

MeMo: A framework for adding knowledge to language models

The Liquid AI research team introduced MeMo, a modular framework adding current, domain-specific knowledge to LLMs without changing model weights. It captures relationships between documents, resists retrieval noise, prevents catastrophic forgetting and works with both open and closed models.

Part of an overview of multiple AI topics; only this event has been covered.

Liquid AI (blog and model releases) Original source ↗