In a study presented at COLM 2026, researchers explore how reasoning in large language models (LLMs) can unlock simple factual knowledge. They find two mechanisms: using generated reasoning tokens to perform latent computation and generating related facts to prime correct answer recall. This phenomenon is driven by computational buffer effects and factual priming, enhancing parametric knowledge recall for simple, single-hop questions.
The National Institute of Standards and Technology (NIST) has released technical findings on the cause of the partial collapse of Champlain Towers South, which resulted in the deaths of 98 people. The investigation found that the collapse began in early June 2021 due to two connections between garage columns and the pool deck failing. This led to cracks growing and loads redistributing, ultimately causing the larger catastrophic collapse on June 24. The findings highlight the importance of margins against failure in building design and construction.
Michelangelo Conserva and Charlotte Stanton from Google Research developed a deep learning framework to analyze fine-scale ecological features like hedgerows and copses on working lands. This precise vector data offers a new pathway to address climate and biodiversity crises without compromising food security. Forests are critical for carbon sequestration and biodiversity, but increasing forest habitat is a global priority. The challenge lies in balancing land use to mitigate climate change and halt biodiversity loss without compromising food security or causing
Google Research··1 min read
LESS NOISE. MORE SIGNAL.
A little further into the unknown.
Follow our open news feed. No inbox clutter. Just the latest dispatches, in your reader.
NRKZ Signal · Signal illustration