Have we forgotten how bad LLMs were at citing sources when they first came out? So, we had to build a lot of structure (harness engineering) and frontier labs had to do specific training to try to compensate for this.
So, LLMs are inherently bad at citing sources. A lot of effort has been put in to improve this behavior, but it's compensating for an inherent flaw.
Huh? Oh! Were they still treating the LLM as an "oracle box"/online chatbot at the time? (as opposed to a more agentic workflow?)
If they weren't, ignore I said the following, and please tell me what else was going wrong (and with what models and harnesses!).
Models weights are like Wikipedia. A nice starting point, but should never be referenced directly. You need to have your agent actually go out onto the internet and do the research. Now the actual references will be in your agent's actual Context (memory), so then it'd at least be rather more surprising if they don't cite correctly.
I do realize there's still corner cases even in the best setups though; So a final crosscheck sweep is never not a good idea.
So, LLMs are inherently bad at citing sources. A lot of effort has been put in to improve this behavior, but it's compensating for an inherent flaw.