A New Solution in Step Learning Artificial Intelligence Agents: Trajectory Graphs Method

Group-based strengthening learning (RL) methods used to train large language models with their skills of reasoning and decision-making are taking importance quickly in the field of artificial intelligence. However, there is a critical weakness of these methods:**The level of step is creating systematic deviations in learning.

New research published in ArXiv solved this problem using trajektori charts (trajectory graphs):

• GRPO and similar group-based methods are able to make reliable advantage predictions at response level
• But because the overall trajektori advantages fail to measure the contribution of individual steps, cannot properly reflect the unsuccessful steps
• Trajectory graphs provide more precise assessment by showing the actual effect of each step on the map
• This method is possible to learn better than LLMs with logic and agent property

Although the fundamental bulgus of the research works for all levels of response to traditional group normalization, it is necessary to understand the quality of each step in complex reasoning tasks. Trajectory graphs reveals where the model is error and how to improve.

This development is an important step in the development of AI agents, especially with the ability to solve the problem in the way. The first research results show that the method significantly increases the performance of group-based methods.