Unslop Skill Alters Chain Of Thought In LLMs
ML researcher tests unslop prompting skill on LLMs with Claude-generated questions.
Elie Bakouch posted about an experiment applying the unslop skill from a Cursor plugins repository. The skill aims to change chain-of-thought behavior in LLMs. Questions for the test were produced by Claude. Elvis Saravia replied that longer or more direct instructions improve adherence in some models while others overalign. The exchange centers on how this prompting technique performs across different LLMs.
Combined views
1.5K
2 posts, first seen 1d ago
Unslop Skill Alters Chain Of Thought In LLMs
ML researcher tests unslop prompting skill on LLMs with Claude-generated questions.
Elie Bakouch posted about an experiment applying the unslop skill from a Cursor plugins repository. The skill aims to change chain-of-thought behavior in LLMs. Questions for the test were produced by Claude. Elvis Saravia replied that longer or more direct instructions improve adherence in some models while others overalign. The exchange centers on how this prompting technique performs across different LLMs.
Reactions
2 postsEL 
elie@eliebakouchexperiment is using @poteto unslop skill experiment setting (claude generated the questions)
- likes: 10
- replies: 0
- bookmarks: 10
- reposts: 0
EL elvis@omarsar0How does the level of detail or length of the instructions in the skill affect results? I've found my agent follows skills better when they are more direct and clear. I also think this is model-dependent. Certain models I see tend to overalign on skill instructions. Interesting…
- likes: 0
- replies: 1
- bookmarks: 0
- reposts: 0
