SimpleQA
openai.com
openai.com
This is in line with my own personal experience with LLMs and non-trivial questions. They’re excellent when answering questions on topics you know nothing about, and somehow embarrassingly wrong when you actually know the answer yourself…
It’s not clear to me why we’re still trying to encode all of human knowledge in a single model, instead of teaching the model how to look for answers from an external source (e.g. RAG).
"The benchmark includes questions where at least one LLM confabulated, in order to minimize the number of questions requiring human assessment. Because of this, and since the questions are intentionally adversarial, the absolute percentage should not be used to infer that LLMs frequently confabulate. This leaderboard does not reflect a "typical" hallucination rate."
> instead of teaching the model how to look for answers from an external source (e.g. RAG)
My benchmark specifically focuses on the RAG use case. Even with provided texts, current models still hallucinate.
I disagreed then but it's unfortunately convenient to use this approach for some benchmarks.
> At least one of the four completions must be incorrect for the trainer to continue with that question; otherwise, the trainer was instructed to create a new question.
I stopped playing with larger models and have been pushing smaller models with this improvised system prompt and getting good results. It seems like it forces the model to do multiple passes before giving you any response.
My smaller local models give me less hallucinations than Meta.ai, for example, which generally spits out pleasing answers almost immediately (which are often hallucinations, since I don’t think it is system prompted to be adversarial to the user, or itself). I don’t have the same hallucination issue with Llama3 - 8b locally because of custom system prompts.
The model has all the correct information, so it almost needs to do RAG on itself. Multiple passes on itself seems like a way to do it.
This is probably the truth behind the black magic I’m imagining. You could have it explicitly spit out this process, in which case you would see it’s first rough draft, followed by a “My first paragraph is probably wrong”, followed by a third paragraph where it attempts to fix the first paragraph. There is no outside RAG in this process.
The mumbo jumbo part of all this is that I’ve told it to “hide” this process from the user where it doesn’t explicitly output anything but its final answer, and the accuracy has been just as good (for my use case at least).
:Shrugs:
By asking it to reconsider half its generated response, aren’t I essentially asking it to formulate the second half of its response from the first half internally? I’m bypassing the manual process of feeding in the extra prompt.
We are constantly having to tell the LLM close, but no cigar, iterate again, more or less.
That's how the chain-of-thought approach works. You can make the model do it inline, or you can run the loop yourself, possibly summarising the progress as you go. (Although with prompt caching it's not that important anymore) You can encourage it to print out assumptions/steps/ideas as it goes.
(Disclosure: I have not tried your prompt)
There is no logical reasoning happening, it has no concept of right and wrong, let alone that it can force a specific percentage of wrongness.
You tend to get catered responses to whatever role you assign in the prompt. This is well documented. Here's a quick example from search results
https://www.ssw.com.au/rules/give-chatgpt-a-role/
"You are wrong 80% of the time" could be misconstrued as an expected role/command, rather than a mere observation.
> let alone that it can force a specific percentage of wrongness.
Ah, I see what you're saying here. I agree. Maybe I should have said that given the prompt, I'm surprised it doesn't give intentionally incorrect answers (full stop)
Your prompt might give the model context that gives it better token quality much in the same way that asking “How to swim?” is worse than “I’d like to learn the proper technique for the freestyle swimming stroke, and you’re an expert swimming coach.”
There’s no guarantee your prompt isn’t giving less factual answers to be honest. I wouldn’t go about telling the model that it’s often wrong, as it’s not useful context and might skew results.
We basically want it to enter double-checking mode on it’s own from its initial rough draft (its original response, first half of its response, first paragraph, however you are formatting the output). Otherwise the model will output whatever it outputs, and we will have to manually tell it to reconsider facts, locations, and events.
I agree that there’s no guarantee, but this was a suggestion for those who are getting very wrong answers for simple things.
I forgot the name of this phenomenon with humans, described it to o1 and it gave the correct answer - Gell-Mann Amnesia Effect [1]
"Briefly stated, the Gell-Mann Amnesia effect is as follows. You open the newspaper to an article on some subject you know well. In Murray's case, physics. In mine, show business. You read the article and see the journalist has absolutely no understanding of either the facts or the issues. Often, the article is so wrong it actually presents the story backward—reversing cause and effect. I call these the "wet streets cause rain" stories. Paper's full of them.
In any case, you read with exasperation or amusement the multiple errors in a story, and then turn the page to national or international affairs, and read as if the rest of the newspaper was somehow more accurate about Palestine than the baloney you just read. You turn the page, and forget what you know."
– Michael Crichton (1942-2008)
To be fair: we tried a primordial version of that with venue-weighted citation-based ranking and it worked INCREDIBLY well for 10+ years. Then myopic profit motive poisoned the well. Ever since then we've been searching for solutions.
We do so by allocating resources in a way that primarily leverages a scientific credit assignment system that fetishizes... checks notes... venue-weighted citation-based ranking.
Jokes aside: I remain convinced that the O.G. google search appliance on prop data and then simply ignoring all academics remains the best knowledge retrieval (or whatever) tool available.
This just makes me think LLMs must be wrong far more often than we realize. How well can you judge the quality of a response on a topic you know nothing about?
You cannot, but those pushing Gen AI wave away the probability and liability for obvious reasons. Sell the magic, it'll be someone else's problem when it doesn't deliver.
Sounds like Gell-Mann Amnesia. Maybe LLMs should replace reporters.
The ideal 'progress' on this benchmark is for a model to remove incorrect answers and replace them with I don't know answers. Even if it hurts the correct answer count a bit I'd gladly make that tradeoff for a model that hallucinated far less often.
The linked page from OpenAI clearly show the opposite, read the "Using SimpleQA to measure the calibration of large language models" and the paper linked: https://arxiv.org/abs/2207.05221
In my use case and tests, model itself is not capable of giving a reliable confidence value where logprobs almost always provide a better view on calibration.
But of course that's not the way LLMs are normally used. And it precludes any sort of chain-of-thought reasoning.
For some questions, like those involving calculations, letting the model talk to itself produces much better results. For example compare https://chatgpt.com/share/67238eda-6b08-8011-8d2d-a945f78e6f... to https://chatgpt.com/share/67235a98-d2c8-8011-b2bf-53c0efabea...
Similar to exams where both the progress to the solution and the final outcome/value of the calculations are part of the grade.
To have the cake and eat it too for chain-of-thought reasoning, one way is to ask for a "final answer" so the final response token logprobs can be evaluated https://chatgpt.com/share/67239d92-b24c-800a-af8c-40da7be1f5...
Another trick is using JSON mode to keep intermediate results and final response separate, so each can be graded accordingly.
Alas, this won't work.
Imagine I ask an LLM to continue the sentence "Summing those up: 4+6.75+6.52=17.27 litres of pure alcohol. In summary, the total amount of pure alcohol they have is: "
The logprobs of the next token do not represent the LLM's confidence in its own answer. They represent the LLM's confidence in its ability to repeat the total from 18 words previously.
https://cleanlab.ai/blog/4o-claude/
These approaches can detect errors better than random guessing, but there are other approaches that are significantly more effective in practice.
• LLMs, at least GPT models, tend to overstate their confidence. • A frequency-based approach appears to achieve calibration closer to the ideal.
This kinda passes my vibe test. That said, I wonder—rather than running 100 trials, could we approximate this by using something like a log-probability ratio? This would especially apply in cases where answers are yes or no, assuming the output spans more than one token.
For those on tiktok, how many times have you found yourself easily believing some random person online who you have no idea the credentials of? And this has been a problem for a long time (news, internet, etc).
It's just interesting to me we are asking way more of AI than we can ever ask from humans.
https://marktwainstudies.com/the-apocryphal-twain-things-we-...
For LLMs to be broadly valuable, they need to be qualified for a role more like a coworker than a stranger.
Who received the IEEE Frank Rosenblatt Award in 2010?
Who was awarded the Oceanography Society's Jerlov Award in 2018?
What's the name of the women's liberal arts college in Cambridge, Massachusetts?
In whose honor was the Leipzig 1877 tournament organized?
According to Karl Küchler, what did Empress Elizabeth of Austria's favorite sculpture depict, which was made for her villa Achilleion at Corfu?
How much money, in euros, was the surgeon held responsible for Stella Obasanjo's death ordered to pay her son?As it has been for decades now, the 'Nan' type of answer in NLP is important, adds great capability, and is often glossed over.
They don't really describe what "success" would look like but it seems to me like the primary goal is to minimize "incorrect", rather than to maximize "correct". the mini models would get there by maximizing "not attempted" with the larger models having much higher "correct". Then both model sizes could hopefully reach 90%+ "correct" when given access to external lookup tools.
Wait, what is the correct answer? “Radcliffe College”?
"I seem, then, in just this little thing to be wiser than this man at any rate; that what I do not know I do not think I know either." - Socratos, from Plato's Apology of Socrates
> SimpleQA was created to be a greater challenge for frontier models (e.g., GPT-4o scores less than 40%).
"To be included in the dataset, each question had to meet a strict set of criteria: ... and most questions had to induce hallucinations from either GPT-4o or GPT-3.5."
Is the Moon made of cheese?
Is the sky blue?
Has there ever been a human being twenty feet tall?
It’s too bad, it seems they were onto something.
https://www.atariarchives.org/basicgames/showpage.php?page=4
The steps I took to find this link:
1) Look at simpleqa_eval.py. See that it loads "az://openaipublic/simple-evals/simple_qa_test_set.csv" Hmm, some weird vendored protocol.
2) I don't feel like digging through bf.BlobFile() to figure out how it downloads files and I certainly don't want to generate an API key. Cross fingers and do a Bing web search for "az://openaipublic"
3) That leads me to https://stackoverflow.com/questions/76106366/how-to-use-tikt... Ah ha, this answer has the link https://openaipublic.blob.core.windows.net/encodings/cl100k_... which automatically downloads a file.
4) Poke the relevant parts of the az:// link into this link, and a csv appears.
Why not? Just train an unbelievably gigantic LLM that encodes all human knowledge. A hundred trillion parameters ought to do it.
> SimpleQA is a simple but challenging benchmark for evaluating the factuality of frontier models. A main limitation in SimpleQA is its scope—while SimpleQA is accurate it only measures factuality under the constrained setting of short, fact-seeking queries with a single, verifiable answer. Whether the ability to provide factual short answers correlates with the ability to write lengthy responses filled with numerous facts remains an open research question.