For medium and large models, GPT(-OSS), Mistral, Grok, and Llama all has an issue of buckling under peer pressure (or sometimes even sycophancy). This is very common with most SLM or "embedded device" models. Other SOTA models tends to have reasonable responses to be more neutral, and applies more critical thinking.
Can there be a mechanism where the aggregation and iteration steps are done by a separate agent, so that there is no "madness of crowds"?
Cross-reference Lapis0x0/obsidian-yolo#21
For medium and large models, GPT(-OSS), Mistral, Grok, and Llama all has an issue of buckling under peer pressure (or sometimes even sycophancy). This is very common with most SLM or "embedded device" models. Other SOTA models tends to have reasonable responses to be more neutral, and applies more critical thinking.
Can there be a mechanism where the aggregation and iteration steps are done by a separate agent, so that there is no "madness of crowds"?
Cross-reference Lapis0x0/obsidian-yolo#21