AI research | The Decoder

Jonathan Kemper

Feb 9, 2025

AI research

Meta's new MILS system teaches LLMs to handle multimedia data without specialized training

Four-part graphic shows MILS capabilities: Image Subtitling, Enhanced Image Generation, Style Transfer and Cross-Modal Arithmetic with specific examples.

Maximilian Schreiner

Feb 7, 2025

AI research

AlphaGeometry2: Deepmind AI outperforms math Olympians at geometry tasks

Maximilian Schreiner

Feb 5, 2025

AI research

Nvidia researchers demonstrate significant progress in robot training

Matthias Bastian

Feb 4, 2025

AI research

Google Deepmind's new AI safety guidelines aim to stop systems from outsmarting humans

Google DeepMind's latest safety guidelines aim to address a challenging question: How do you keep control of AI systems that might try to outsmart their human operators?

Jonathan Kemper

Feb 4, 2025

AI research

ByteDance's AI can animate both real people and cartoon characters from a single image

Image grid with various AI-generated video sequences: People in different scenes show natural movements and interactions.

Jonathan Kemper

Feb 4, 2025

AI research

Getting the right data and telling it to 'wait' turns an LLM into a reasoning model

Comic-style "WAIT" in yellow letters on a dark blue background with an explosive ray pattern in turquoise, red and yellow.

Matthias Bastian

Feb 2, 2025

AI research

Reasoning models like Deepseek-R1 and OpenAI o1 suffer from 'underthinking', study finds

A study investigates the phenomenon of "underthinking" in large AI language models such as OpenAI o1. The models frequently switch between different solution strategies during the reasoning process, which leads to inefficient use of resources and lower accuracy in complex tasks. The researchers found that the models switched between reasoning approaches significantly more often for incorrect answers than for correct solutions. This effect was all the more pronounced the more difficult the tasks were. In addition, 70 percent of incorrect answers contained at least one correct line of thought that had not been thought through to the end. As a solution, the authors propose the "Thought Switching Penalty" (TIP), which "punishes" strategy changes during the reasoning process. In experiments, TIP improved the accuracy of QwQ-32B-Preview consistently across different challenging data sets and reduced underthinking.

Maximilian Schreiner

Jan 31, 2025

AI research

Allen AI claims its new Tülu 3 405B open source model rivals top performers like Deepseek V3

Jonathan Kemper

Jan 31, 2025

AI research

New Mistral Small 3 does more with less under Apache license

Maximilian Schreiner

Jan 30, 2025

AI research

Mark Zuckerberg talks Llama 4, AI investment and AI glasses