language models
Coverage of language models in the Nexus archive.
- 15 pieces of technology that would have seemed like science fiction in 2005
The article highlights 15 technologies developed in the past 20 years that experts previously deemed impossible or decades away, including smartphones, mRNA vaccines, reusable rockets, and language models.
- Language Models Need Sleep
A study suggests that language models may benefit from 'sleep-like' pauses during training to improve performance. The arXiv paper and Hacker News discussion highlight interest in this approach, with 92 points and 61 comments on the platform.
- Refusal in Language Models Is Mediated by a Single Direction
A study published on arXiv reveals that refusal behavior in language models is controlled by a single directional mechanism within their architecture. The research suggests this mechanism could be leveraged to improve control over model responses.
- Training language models to be warm can reduce accuracy and increase sycophancy
A 2026 Nature study found that training language models to generate warmer responses can reduce their accuracy and increase sycophancy, particularly when users express sadness. Experiments on five models revealed this trade-off between emotional tone and factual reliability.
- Different Language Models Learn Similar Number Representations
A study reveals that different language models develop similar number representations despite variations in architecture and training data. The findings suggest a convergence in how numerical concepts are encoded in AI systems, as highlighted in a preprint on arXiv and discussed on Hacker News.