Researchers discovered a new one “Scaling Method” of Artificial Intelligence? That’s it Some buzz on social media Suggestions – But experts are skeptical.
The law of AI scaling (a somewhat informal concept) describes how the performance of AI models improves with the size of the datasets and computing resources used to train them. Until a year ago, expanding the expansion of “pretraining” (training more and more models in an increasingly large number of data sets) has been the main law to date, at least in the sense that most Frontier AI labs accept it.
It did not disappear before training, but after training the scaling rate and Test time zoomhas appeared to supplement it. Post-training scaling is essentially adjusting the behavior of the model, while test-time scaling requires applying more computation to inference (i.e. running the model) to drive the “inference” form (see: Similar models R1).
Researchers at Google and UC Berkeley recently proposed Paper Some online commentators describe it as the Fourth Law: “Inferential Time Search”.
Inference time searches for many possible query answers with parallel connections and then select the “best” of the bundle. Researchers claim it can improve a year-old model, e.g. Google’s Gemini 1.5 Proreach the level of Openai O1-preiview “Inference” model of scientific and mathematical benchmarks.
Our paper focuses on this search axis and its expansion trends. For example, by randomly sampling 200 responses and self-verification, Gemini 1.5 (an ancient 2024 model!) beats O1-preview and approaches O1. This is without padding, RL or ground validators. pic.twitter.com/hb5fo7ifnh
– Eric Zhao (@Ericzhao28) March 17, 2025
“[B]y is just a random sampling of 200 responses and self-verification, Gemini 1.5 (an ancient early 2024 model) beat O1-preiview and approached O1,” Eric Zhao, a Google PhD researcher and co-author of the paper. A series of posts on x. “Magic is naturally easier on a large scale!
Several experts say the results are not surprising and in many cases the inference time search may not be useful.
Matthew Guzdial, an AI researcher and assistant professor, told TechCrunch that the method is most effective when there is a good “evaluation function”, in other words, when the best answer to the question is easy to determine. But most queries are not that cut.
“[I]f We can’t write code to define what we want, we can’t use [inference-time] Search,” he said. […] Often, this is not a good way to actually solve most problems. ”
Mike Cook, a researcher at King’s College London who specializes in AI, agreed with Guzdial’s assessment, adding that it highlights the gap between the “rational meaning” of AI and the “reasoning” in our own thinking processes.
“[Inference-time search] “The reasoning process that does not improve the model,” Cook said.[I]T is just the way we revolve around the limitations of a technique that is prone to making very confident mistakes […] Intuitively, if your model has 5% of the time errors, checking 200 attempts on the same issue should make these errors easier to spot. ”
The potential limitations of inference time searches is certainly an unpopular news, and the industry wants to extend the model to “inference” to calculate efficiently. As the co-author of paper states, today’s inference models can be improved Thousands of dollars in calculations Regarding a single mathematical problem.
It seems that new expansion technologies will continue to be searched.
