Speaker unverified: uncertainty

Transcript context

…I still disagree with the premise because there’s a lot of weird complexities that you can’t prove. One of the things that points to weirdness is that if you take the Qwen 3 so-called base model—you could Google “math dataset Hugging Face” and take a problem—if you put it into Qwen 3 base… all these math problems have words, so it would be like, “Alice has five apples and gives three to whoever,” and there are these word problems. With these Qwen-based models, why people are suspicious of them is if you change the numbers but keep the words— —Qwen will produce, without tools, a very high accuracy decimal representation— —of the answer, which means at some point it was shown problems that were almost identical to the test set, and it was using tools to get a very high precision answer. But a language model without tools will never actually have this. So it’s been this big debate in the research community: how much of these reinforcement learning papers that are training on Qwen and measuring specifically on this math benchmark—where there’s been multiple papers talking about contamination—how much can you believe them? I think this is what caused the reputation of RLVR being about formatting, because you can get these gains so quickly and therefore it must already be in the model. But there’s a lot of complexity here. It’s not really like controlled experimentation— —so we don’t really know. But if it weren’t true, I would say distillation wouldn’t work, right? Distillation can work to some extent, but the biggest problem—and I’m researching this contamination—is we don’t know what’s in the data. Unless you have a new dataset, it is really impossible. Even something simpler like MMLU, which is a multiple-choice benchmark—if you just change the format slightly, like using a dot instead of a parenthesis, the model accuracy will vastly differ. I think that that could be like a model issue rather than a general issue.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.