Cool results! Some of these are good student project ideas for courses and such.
The “Let’s think step by step” result about the Hindsight neglect submission to the Inverse Scaling Prize contest is a cool demonstration, but a few more experiments would be needed before we call it surprising. It’s kind of expected that breaking the pattern helps break the spurious correlation.
1. Does “Let’s think step by step” help when “Let’s think step by step” is added to all few-shot examples? 2. Is adding some random string instead of “Let’s think step by step” significantly worse?
Cool results! Some of these are good student project ideas for courses and such.
The “Let’s think step by step” result about the Hindsight neglect submission to the Inverse Scaling Prize contest is a cool demonstration, but a few more experiments would be needed before we call it surprising. It’s kind of expected that breaking the pattern helps break the spurious correlation.
1. Does “Let’s think step by step” help when “Let’s think step by step” is added to all few-shot examples?
2. Is adding some random string instead of “Let’s think step by step” significantly worse?