These researchers used NPR Sunday Puzzle inquiries to benchmark AI ‘reasoning’ fashions
Each Sunday, NPR host Will Shortz, The New York Instances’ crossword puzzle guru, will get to quiz 1000’s of listeners in a long-running section known as the Sunday Puzzle. Whereas written to be solvable with out too a lot foreknowledge, the brainteasers are normally difficult even for expert contestants.
That’s why some consultants suppose they’re a promising strategy to check the boundaries of AI’s problem-solving skills.
In a current examine, a group of researchers hailing from Wellesley Faculty, Oberlin Faculty, the College of Texas at Austin, Northeastern College, Charles College, and startup Cursor created an AI benchmark utilizing riddles from Sunday Puzzle episodes. The group says their check uncovered shocking insights, like that reasoning fashions — OpenAI’s o1, amongst others — generally “hand over” and supply solutions they know aren’t appropriate.
“We needed to develop a benchmark with issues that people can perceive with solely common data,” Arjun Guha, a pc science college member at Northeastern and one of many co-authors on the examine, informed TechCrunch.
The AI business is in a little bit of a benchmarking quandary in the intervening time. A lot of the checks generally used to judge AI fashions probe for abilities, like competency on PhD-level math and science questions, that aren’t related to the common person. In the meantime, many benchmarks — even benchmarks launched comparatively lately — are shortly approaching the saturation level.
The benefits of a public radio quiz recreation just like the Sunday Puzzle is that it doesn’t check for esoteric data, and the challenges are phrased such that fashions can’t draw on “rote reminiscence” to unravel them, defined Guha.
“I believe what makes these issues onerous is that it’s actually troublesome to make significant progress on an issue till you remedy it — that’s when every part clicks collectively all of sudden,” Guha stated. “That requires a mix of perception and a technique of elimination.”
No benchmark is ideal, after all. The Sunday Puzzle is U.S. centric and English solely. And since the quizzes are publicly accessible, it’s doable that fashions educated on them can “cheat” in a way, though Guha says he hasn’t seen proof of this.
“New questions are launched each week, and we will count on the most recent inquiries to be actually unseen,” he added. “We intend to maintain the benchmark recent and observe how mannequin efficiency adjustments over time.”
On the researchers’ benchmark, which consists of round 600 Sunday Puzzle riddles, reasoning fashions akin to o1 and DeepSeek’s R1 far outperform the remaining. Reasoning fashions completely fact-check themselves earlier than giving out outcomes, which helps them keep away from a number of the pitfalls that usually journey up AI fashions. The trade-off is that reasoning fashions take slightly longer to reach at options — sometimes seconds to minutes longer.
At the least one mannequin, DeepSeek’s R1, provides options it is aware of to be fallacious for a number of the Sunday Puzzle questions. R1 will state verbatim “I hand over,” adopted by an incorrect reply chosen seemingly at random — conduct this human can actually relate to.
The fashions make different weird selections, like giving a fallacious reply solely to instantly retract it, try to tease out a greater one, and fail once more. Additionally they get caught “considering” without end and provides nonsensical explanations for solutions, or they arrive at an accurate reply immediately however then go on to contemplate various solutions for no apparent purpose.
“On onerous issues, R1 actually says that it’s getting ‘annoyed,’” Guha stated. “It was humorous to see how a mannequin emulates what a human may say. It stays to be seen how ‘frustration’ in reasoning can have an effect on the standard of mannequin outcomes.”
The present best-performing mannequin on the benchmark is o1 with a rating of 59%, adopted by the lately launched o3-mini set to excessive “reasoning effort” (47%). (R1 scored 35%.) As a subsequent step, the researchers plan to broaden their testing to extra reasoning fashions, which they hope will assist to determine areas the place these fashions is likely to be enhanced.
“You don’t want a PhD to be good at reasoning, so it ought to be doable to design reasoning benchmarks that don’t require PhD-level data,” Guha stated. “A benchmark with broader entry permits a wider set of researchers to grasp and analyze the outcomes, which can in flip result in higher options sooner or later. Moreover, as state-of-the-art fashions are more and more deployed in settings that have an effect on everybody, we imagine everybody ought to be capable to intuit what these fashions are — and aren’t — able to.”
