CausalLM is not optimal for in-context learning