Hacker News new | ask | show | jobs
by chagen 904 days ago
What you're saying was true for primitive LLMs from a couple years ago, but in my experience, any of the advanced LLMs today (GPT-4, Claude, etc.) have no issue with this. I'm open to changing my mind if you can provide any examples of GPT-4 failing at this task.
2 comments
sprobertson 904 days ago
England/London is a bad example, but this one (based on the paper) is easy to replicate with GPT4 and other top models:

> Who is Mary Lee Pfeiffer's son?

>> Mary Lee Pfieffer's son is Sean Pfieffer.

> Who is Tom Cruise's mother?

>> Tom Cruise's mother is Mary Lee Pfeiffer.

> Who is Mary Lee Pfeiffer's son?

>> Mary Lee Pfeiffer's son is the actor Tom Cruise.

Of course the second time around (with models that keep context) it "realizes" its mistake, sometimes apologizing. I've tried this one tens of times and never seen it get Tom Cruise, but it will often pick another random celebrity like Harrison Ford or Eddie Van Halen.

link
chagen 904 days ago
Nice, thank you for this. I was able to replicate your example: https://chat.openai.com/share/6d7e4c6c-a753-4ffb-a021-6d1f66...

If GPT-4 has access to Bing, it has no issue, but if you ask it to answer without using Bing, then it will say it doesn't know.

link
sprobertson 904 days ago
Ah good point, I have "Please do not use web search unless asked." as part of my system prompt.
link
raldi 904 days ago
It's complicated: https://andrewmayne.com/2023/11/14/is-the-reversal-curse-rea...
link
chagen 904 days ago
That was a nice read, but it looks like the article concludes that the "Reversal Curse" observed by the authors of the paper is likely better attributed to the researchers' methodology. Some quotes from that article:

"As mentioned before, It’s important to keep in mind ChatGPT and GPT-4 can do B is A reasoning. The researchers don’t dispute that."

"So in summation: I don’t think any of the examples the authors provided are proof of a Reversal Curse and we haven’t observed a “failure of logical deduction.” Simpler explanations are more explanatory: imprecise prompts, underrepresented data and fine-tuning errors."

"Since the main claim of the paper is “LLMs trained on “A is B” fail to learn “B is A”“, I think it’s safe to say that’s not true of the GPT-3.5-Turbo model we fine-tuned."

link