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 referred to as the Sunday Puzzle. Whereas written to be solvable with out too a lot foreknowledge, the brainteasers are often difficult even for expert contestants.
That’s why some specialists assume they’re a promising method to check the boundaries of AI’s problem-solving talents.
In a current research, a crew of researchers hailing from Wellesley School, Oberlin School, the College of Texas at Austin, Northeastern College, Charles College, and startup Cursor created an AI benchmark utilizing riddles from Sunday Puzzle episodes. The crew says their check uncovered shocking insights, like that reasoning fashions — OpenAI’s o1, amongst others — typically “surrender” and supply solutions they know aren’t right.
“We wished to develop a benchmark with issues that people can perceive with solely common data,” Arjun Guha, a pc science school member at Northeastern and one of many co-authors on the research, informed TechCrunch.
The AI trade is in a little bit of a benchmarking quandary for the time being. Many of the exams generally used to guage 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 not too long ago — are rapidly approaching the saturation level.
Some great 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 resolve them, defined Guha.
“I believe what makes these issues arduous is that it’s actually troublesome to make significant progress on an issue till you clear up it — that’s when every thing clicks collectively abruptly,” Guha mentioned. “That requires a mix of perception and a means of elimination.”
No benchmark is ideal, in fact. The Sunday Puzzle is U.S. centric and English solely. And since the quizzes are publicly obtainable, it’s potential that fashions skilled 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 are able to anticipate the most recent inquiries to be actually unseen,” he added. “We intend to maintain the benchmark recent and monitor how mannequin efficiency modifications 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 totally fact-check themselves earlier than giving out outcomes, which helps them keep away from among the pitfalls that usually journey up AI fashions. The trade-off is that reasoning fashions take slightly longer to reach at options — usually seconds to minutes longer.
Not less than one mannequin, DeepSeek’s R1, provides options it is aware of to be flawed for among the Sunday Puzzle questions. R1 will state verbatim “I surrender,” adopted by an incorrect reply chosen seemingly at random — habits this human can definitely relate to.
The fashions make different weird selections, like giving a flawed reply solely to right away retract it, try to tease out a greater one, and fail once more. In addition they get caught “pondering” eternally and provides nonsensical explanations for solutions, or they arrive at an accurate reply instantly however then go on to think about various solutions for no apparent purpose.
“On arduous issues, R1 actually says that it’s getting ‘annoyed,’” Guha mentioned. “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 not too long ago launched o3-mini set to excessive “reasoning effort” (47%). (R1 scored 35%.) As a subsequent step, the researchers plan to broaden their testing to further reasoning fashions, which they hope will assist to determine areas the place these fashions is perhaps enhanced.
“You don’t want a PhD to be good at reasoning, so it ought to be potential to design reasoning benchmarks that don’t require PhD-level data,” Guha mentioned. “A benchmark with broader entry permits a wider set of researchers to understand 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 consider everybody ought to be capable to intuit what these fashions are — and aren’t — able to.”
