“AI got angry and gave the wrong answer! Shocking benchmark results – DeepSeek is annoyed..”

Benchmark Test Overview

Let's examine a recent artificial intelligence (AI) benchmark test conducted by researchers at Wellesley College. This test used approximately 600 questions from NPR's Sunday Puzzle quiz to compare the performance of various AI reasoning models.

Main Models and Performance

• o1 model: Achieved the highest performance with a 59% accuracy rate.
• o3-mini: Followed with a 47% accuracy rate.
• DeepSeek-R1: Recorded a 35% accuracy rate, exhibiting unusual behavior during the test.

Unpredictable Behavior of the R1 Model

Researchers noted that DeepSeek-R1 displayed human-like "frustration" and gave up while solving benchmark problems.

○ In some problems, R1 knew its answer was wrong but submitted a random incorrect answer along with the phrase, "I give up."
○ After providing incorrect answers, it attempted to retract them and find better answers, but repeatedly failed.
○ It frequently got stuck in "thought" processes, failing to reach conclusions, or considered unnecessary additional answers even when the correct answer was clear.

Researchers' Interpretation and Future Plans

Researchers commented on the interesting aspect of R1 exhibiting "frustration" in difficult situations, emphasizing the need to carefully analyze how such human-like responses affect the AI reasoning process.

○ Although R1's performance is currently lower than other models, they plan further research to investigate whether these responses could help improve AI problem-solving strategies.
○ They plan to conduct tests on additional reasoning models to more specifically identify the strengths and weaknesses of each model.

Conclusion and Implications

This benchmark test provided insights into the problem-solving approaches and limitations of various AI models.

○ In particular, the human-like "frustration" and giving-up behavior exhibited by DeepSeek-R1 is an interesting subject of study, as it suggests that AI can exhibit responses similar to emotions beyond simple calculations.
○ These research findings are significant as they can provide new insights into improving AI performance and the way it interacts with humans.

It would be beneficial to maintain interest in the characteristics and problem-solving approaches of each model and to follow the future trends in AI development.

*Source URL:
https://www.aitimes.com/news/articleView.html?idxno=168048


Benchmark Test Overview Let's examine a recent artificial intelligence (AI) benchmark test conducted by researchers at Wellesley College. This test used approximately 600 questions from NPR's Sunday Puzzle quiz to compare the performance of various AI reasoning models. Main Models and Performance • o1 model: Achieved the highest performance with a 59% accuracy rate.• o3-mini:…

Leave a Reply

Your email address will not be published. Required fields are marked *

Feature is an online magazine made by culture lovers. We offer weekly reflections, reviews, and news on art, literature, and music.

Please subscribe to our newsletter to let us know whenever we publish new content. We send no spam, and you can unsubscribe at any time.