This study explores unsupervised methods, including principal component analysis and Isomap, alongside difference-in-means to investigate the refusal direction in LLMs. Experiments are conducted on Llama 8b 3.1 Instruct and Qwen2.5-7B-Instruct models using Alpaca and Beavertails datasets.
The chart will appear after repeat observations. The current metric comes from the source.
7 source pointsReal observations only. History before source connection is not reconstructed.
This research could help in understanding the mechanisms behind refusal in language models, which is important for improving their reliability and safety.
In a post by Nick Cammarata, it is discussed that neural networks may be understandable thanks to modern interpretability tools. The author claims that the majority of neurons in models with trillions of parameters are interpretable, and their circuits are beautiful. He also notes that neurons previously considered complex were misinterpreted.
6 days agoLessWrongThe post critiques the article on social reality in China, highlighting an unbalanced comparison and a lack of systematic analysis. The author emphasizes that Chinese intellectual communities have their own distinct characteristics, different from Western ones, and that values such as knowledge, truth, and technology ethics are also present in China.
6 days agoLessWrongThe author argues that owning large amounts of capital may lead to losing influence earlier than others, through mental breakdown. This is because people seek to manipulate the wealthy to gain resources, making it harder to maintain values and accurate beliefs.
2 days agoHacker NewsAn upcoming critical security update for PHP will address vulnerabilities CVE-2026-109618 and CVE-2026-109619. More information is available on the websites listed in the description.
yesterdayLobstersAn upcoming security release for the PHP programming language, likely including vulnerability fixes.
yesterdayHacker NewsMozilla's blog reports on the compromise of the eMudhra intermediate certificate authority. Additional details are available in the bug report.
yesterday