Did OpenAI Cheat on Its Massive Math Take a look at? – Decrypt

How clever is a mannequin that memorizes the solutions earlier than an examination? That’s the query going through OpenAI after it unveiled o3 in December, and touted its mannequin’s spectacular benchmarks. On the time, some pundits hailed it as being virtually as highly effective as AGI, the extent at which synthetic intelligence is able to attaining the identical efficiency as a human on any activity required by the person.

However cash adjustments all the things—even math assessments, apparently.

OpenAI’s victory lap over its o3 mannequin’s gorgeous 25.2% rating on FrontierMath, a difficult mathematical benchmark developed by Epoch AI, hit a snag when it turned out the corporate wasn’t simply acing the check—OpenAI helped write it, too.

“We gratefully acknowledge OpenAI for his or her help in creating the benchmark,” Epoch AI wrote in an up to date footnote on the FrontierMath whitepaper—and this was sufficient to boost some purple flags amongst lovers.

Worse, OpenAI had not solely funded FrontierMath’s growth but in addition had entry to its issues and options to make use of because it noticed match. Epoch AI later revealed that OpenAI employed the corporate to supply 300 math issues, in addition to their options.

“As is typical of commissioned work, OpenAI retains possession of those questions and has entry to the issues and options,” Epoch mentioned Thursday.

Neither OpenAI nor Epoch replied to a request for remark from Decrypt. Epoch has nonetheless mentioned that OpenAI signed a contract prematurely indicating it might not use the questions and solutions in its database to coach its o3 mannequin.

The Info first broke the story.

Whereas an OpenAI spokesperson maintains OpenAI did not immediately practice o3 on the benchmark, and the issues had been “strongly held out” (which means OpenAI didn’t have entry to a few of the issues), consultants word that entry to the check supplies might nonetheless permit efficiency optimization by way of iterative changes.

Tamay Besiroglu, affiliate director at Epoch AI, mentioned that OpenAI had initially demanded that its monetary relationship with Epoch not be revealed.

“We had been restricted from disclosing the partnership till across the time o3 launched, and in hindsight we should always have negotiated more durable for the flexibility to be clear to the benchmark contributors as quickly as potential,” he wrote in a publish. “Our contract particularly prevented us from disclosing details about the funding supply and the truth that OpenAI has information entry to a lot, however not all the dataset.”

Tamay mentioned that OpenAI mentioned it wouldn’t use Epoch AI’s issues and options—however didn’t signal any authorized contract to ensure that can be enforced. “We acknowledge that OpenAI does have entry to a big fraction of FrontierMath issues and options,” he wrote. “Nonetheless, we’ve got a verbal settlement that these supplies is not going to be utilized in mannequin coaching.”

Fishy because it sounds, Elliot Glazer, Epoch AI’s lead mathematician, mentioned he believes OpenAI was true to its phrase: “My private opinion is that OAI’s rating is legit (i.e., they did not practice on the dataset), and that they haven’t any incentive to lie about inside benchmarking performances,” he posted on Reddit.

The researcher additionally took to Twitter to handle the state of affairs, sharing a hyperlink to a web-based debate concerning the difficulty within the on-line discussion board Much less Unsuitable.

Not the primary, not the final

The controversy extends past OpenAI, pointing to systemic points in how the AI trade validates progress. A current investigation by AI researcher Louis Hunt revealed that different prime performing fashions together with Mistral 7b, Google’s Gemma, Microsoft’s Phi-3, Meta’s Llama-3 and Alibaba’s Qwen 2.5 had been capable of reproduce verbatim 6,882 pages of the MMLU and GSM8K benchmarks.

MMLU is an artificial benchmark, identical to FrontierMath, that was created to measure how good fashions are at multitasking. GSM8K is a set of math issues used to benchmark how proficient LLMs are at math.

That makes it inconceivable to correctly assess how highly effective or correct their fashions really are. It’s like giving a scholar with a photographic reminiscence an inventory of the issues and options that might be on their subsequent examination; did they purpose their option to an answer, or just spit again the memorized reply? Since these assessments are supposed to display that AI fashions are able to reasoning, you’ll be able to see what the fuss is about.

“It is really A VERY BIG ISSUE,” RemBrain founder Vasily Morzhakov warned. “The fashions are examined of their instruction variations on MMLU and GSM8K assessments. However the truth that base fashions can regenerate assessments—it means these assessments are already in pre-training.”

Going ahead, Epoch mentioned it plans to implement a “maintain out set” of fifty randomly chosen issues that might be withheld from OpenAI to make sure real testing capabilities.

However the problem of making really impartial evaluations stays important. Laptop scientist Dirk Roeckmann argued that superb testing would require “a impartial sandbox which isn’t simple to understand,” including that even then, there is a danger of “leaking of check information by adversarial people.”

Edited by Andrew Hayward