Ask HN: How do you avoid / get out of LLMs local minima?
3 comments
This is conversational momentum due to the autoregressive nature of models. Each token is drawn from a probability distribution conditioned on the preceding tokens. The best method I know of working around this is to request and curate a markdown export package from a conversation that is used to prime a new conversation in a clean context.
Having a separate planning / research phase helps with this. Make the LLM curate a plan by gathering internal and external context. Then execute the plan in another fresh session. Of course if the planning phase itself ends up in the local min then I would just start a new planning session with the learnings.
I have found that in such situations all one needs to do is ask "are you sure x is y" atleast as far as most Claude models go. That usually results in and apology and and escape from your local min.
What's your trick to get the LLM out of its own beliefs ?
I often find myself having to kill the entire context, which is obviously not the best...