https://arxiv.org/pdf/2410.05229

Key idea:

• Does LLM really know how to do formal reasoning or is it just a pattern-matching?
• The paper therefore makes several different variations of GSM8K, a widely used math benchmark.

Variations:

• GSM-Symbolic: Change the names and numbers using the template

  

• GSM8K with different difficulties

  

  • GSM-Symbolic-M1: one fewer clause
  • GSM-Symbolic-P1: one more clause
  • GSM-Symbolic-P2: two more clauses

• GSM-NoOp: Adding a clause that looks relevant but actually irrelevant.

  

Results:

GSM-Symbolic

• The performances of all the models dropped.
• No big change when only names are changed. However, performance dropped significantly when the numbers changed.

MY CONCERN:

• LLMs are innately bad at arithmetic. What we should check is if it is making calculation mistake or logical mistake.
• I guess the mistakes they committed are mostly calculation mistakes, as the performance never changed that much when only the names are changed.