TeleAI-t1-preview - A complex inference model launched by China Telecom

What is TeleAI-t1-preview?

TeleAI-t1-preview, a "complex reasoning model" released by the China Telecom Artificial Intelligence Research Institute, possesses powerful logical reasoning and mathematical derivation capabilities. Through reinforcement learning training methods and the introduction of exploratory and reflective thinking paradigms, it improves the accuracy of solving complex problems. In the 2024 AIME and MATH500 evaluations of the American Mathematical Olympiad, the model scored 60 and 93.8 points respectively, surpassing benchmark models such as OpenAI's o1-preview and GPT-4o. It can accurately handle classical Chinese problems from the *Nine Chapters on the Mathematical Art*, converting them into modern Chinese and providing detailed derivations. TeleAI-t1-preview will soon be available on the China Telecom AI Open Platform and will play an important role in education, scientific research, and other fields in the future.

Main functions of TeleAI-t1-preview

• Mathematical and logical reasoning ability
  • Solutions to difficult math problemsTeleAI-t1-preview excels in mathematical reasoning and can handle complex mathematical problems. In the 2024 American Mathematics Competitions AIME and MATH500 evaluations, it achieved high scores of 60 and 93.8 respectively, significantly surpassing benchmark models such as OpenAI's o1-preview and GPT-4o.
  • Graduate-level Q&A TestIn the graduate-level question-answering test GPQA Diamond, TeleAI-t1-preview scored higher than GPT-4o and matched the performance level of Claude 3.5 Sonnet.
  • Classical Mathematical Text UnderstandingIt can handle classic mathematical texts such as "Nine Chapters on the Mathematical Art". It first understands and simplifies the classical Chinese questions, converts them into modern Chinese, and then performs mathematical derivation and solution.
• Thinking and reasoning ability
  • Combining figurative and abstract thinkingTeleAI-t1-preview can combine visual and abstract thinking, allowing users to visualize the scenarios involved in complex problems and aiding in understanding the questions.
  • Complex Strategy ReasoningWhen faced with extremely challenging strategic reasoning problems, one can quickly understand the game rules and solve the problem, outlining one's understanding of the game rules, analysis of the scene and props, analysis of advantages and disadvantages, and providing a solution strategy.
  • Unit conversion between ancient and modern timesTeleAI-t1-preview demonstrates rigor in handling conversions between ancient and modern units, ensuring the accuracy of the answers.

The technical principles of TeleAI-t1-preview

• Strengthen learning and thinking paradigmsThe model employs reinforcement learning training methods and introduces thinking paradigms such as exploration and reflection. It can optimize reasoning ability through trial and error, and significantly improve the accuracy of complex problems such as mathematical derivation and logical reasoning.
• Data preparationThe research institute collected and constructed a high-quality reasoning dataset with mathematics as its core and multiple disciplines as supplements, to ensure that the model can adapt to different types of reasoning tasks.
• Judge ModelA specialized evaluation model was trained to analyze and evaluate the correctness of the model's long-term thought process, providing precise guidance for model reflection and error correction.
• Supervised fine-tuning (SFT) phaseHigh-quality long inference data is constructed based on Monte Carlo Tree Search (MCTS), and the optimal path is selected by combining the accuracy and solution length of each step. Low-accuracy paths are analyzed and corrected through the Judge Model, and high-quality thought chain data is constructed for SFT training.
• reinforcement learning phaseAn additional rule-based reward model was constructed to provide sufficiently accurate reward signals, and the model's logical reasoning ability was further improved through online reinforcement learning algorithms.

Application scenarios of TeleAI-t1-preview

• Mathematics learning and competition coachingTeleAI-t1-preview can handle complex mathematical problems, including high school math competition problems and graduate-level math problems.
• Analysis of Ancient Mathematical ProblemsThe model can understand and simplify ancient Chinese mathematical problems such as those in "Nine Chapters on the Mathematical Art," convert them into modern Chinese, and perform mathematical derivations, providing strong support for learning ancient mathematics.
• Logical Reasoning and Strategy AnalysisTeleAI-t1-preview performs exceptionally well when handling complex strategy reasoning problems. It can quickly understand the rules and solve the problem, listing its understanding of the game rules, analysis of the scene and props, analysis of advantages and disadvantages, and providing problem-solving strategies.
• Interdisciplinary research supportStrong logical reasoning ability can assist researchers, help solve complex logical problems, and improve research efficiency.