a hypothesis for the mechanism of reasoning

research hypothesis llm

• the direction representing opposite concepts (e.g. A is true and A is false) are not represented by opposite directions
• Hence, they can co-occur and the model can detect this either by
  • an overlapping (non-zero scalar projection) of the 2 directions
  • or a “logical conflict” direction which is the difference of the 2 said directions (I prefer this hypothesis)
• then it would trigger a “rethink” direction that makes the model “backtrack” its reasoning