I really feel out of my depth because 2 out of the 3 methods here seem like they shouldn’t work?

> To evaluate comprehensibility quantitatively, we employ an LLM-as-a-Judge framework

This isn’t the worst idea, but it’s still a bit incestuous. Adding an LLM judge to check for hallucinations creates two new kinds of problems: false positives, where your judge hallucinates an incorrect fact, and false negatives, where the judge lets a hallucination slip by.

> We measure reproducibility through knowledge distillation. By fine-tuning a weaker model on the generated CoT traces, we use the downstream performance gain of the student as a proxy.

And my problem here, as a member of the GPU proletariat, is that this just seems incredibly inefficient. In other words, you’re going to generate a bunch of rollouts from your model then wait for the student to train? I guess if you have the compute to train a trillion params then maybe you don’t care.

> LLM-as-a-Judge

Empirically, many problems look like they're easier to check than they are to solve. This seems like a reasonable way to bootstrap a little extra performance, with prior art in well-known DeepMind experiments. It's unclear if it works recursively (I imagine not), but the core idea is solid.

I have my LLM agents fact check each other as a matter of course. They each regularly find things that the other missed. They are typically the same model (Opus).