News for hackers. Stuff that matters.

rodhruv

I don’t think it’s nearly as impressive as it is hyped to be:

They only ever evaluate on 100 prompts, and the error bars on their headline metrics are +-4-5. The headline result (65%) only applies to Qwen3-0.6B chat; for GPT-4o, it’s 11%, and for Llama-2, 17%. It’s fine-tuned on 400 ShareGPT prompts, and I suspect most of the gains are due to learning the ShareGPT prompt format. Indeed, if fine-tuned on a different dataset, it only achieves 11%.

The paper also has several misleading claims, like calling it “black-box” (needs the tokenizer) or “data-free” (only during pretraining, but see their own appendix).