Some questions to consider

[BradKML]

1. Would fixing LLMs with Dr. GRPO (which protects against overthinking AND giving wrong answers more frequently), enable LLMs to handle premises and context better? https://github.com/sail-sg/understand-r1-zero
2. Should the LLM fill in the premise with a variable, and turn the answer to the "broken" question into a function? That way, it won't just deal with checking dead ends, but also shows the way they answer "it depends" questions
3. What about logic puzzles and other linguistic problems? ZebraLogic for example, seemed like a good candidate for this https://github.com/WildEval/ZeroEval
4. How do you view false or inaccurate information? https://github.com/dannyallover/overthinking_the_truth