Microsoft's Well-Thought-Out Plan for Addressing LLM Issues

As irrational language models continue to have a growing impact on various aspects of our lives, Microsoft has released a method called "Everything of Thought (XOT)" that improves artificial intelligence reasoning. This method draws inspiration from Google DeepMind's AlphaZero, which uses small neural networks to achieve better performance than larger ones.

The new XOT method was developed in collaboration with the Georgia Institute of Technology and East China Normal University. They combined reinforcement learning and Monte Carlo tree search (MCTS), both known for their effectiveness in complex decision-making.

Researchers say that combining these techniques allows language models to effectively generalize unknown problems. Impressive results have been achieved in solving challenging tasks such as the 24-point game, the 8-puzzle problem, and the Rubik's Cube, surpassing other methods in solving problems that were previously unsolvable. However, this advantage is not without limitations. While the system has made progress, it has not yet reached 100% reliability.

Nevertheless, the research team believes that this framework is an effective way to incorporate external knowledge into language model reasoning. They are confident that it can simultaneously improve performance, efficiency, and flexibility, which cannot be achieved through other methods.

Attempting Reasoning

The reason researchers focused on integrating game next-step into language models is that these models can generate impressive sentences with accuracy, but they lack the ability for a crucial aspect of human-like thinking: logical reasoning.

Researchers have been studying this issue for a long time. Over the years, the academic and technology communities have been deeply involved in researching this challenge. However, despite their efforts to enhance artificial intelligence with more layers, parameters, and attention mechanisms, a solution has been lacking. They have also been exploring multimodal approaches, but advanced and reliable results have not been achieved.

Earlier this year, a collaboration between Virginia Tech and Microsoft released a method called "Algorithm of Thoughts (AoT)" to improve algorithmic reasoning in artificial intelligence, as we all know how bad ChatGPT was in mathematics. Furthermore, it demonstrated that with this training method, large language models can integrate their intuitions into optimized searches for better results.

Additionally, Microsoft recently scrutinized the ethical reasoning of these models. Therefore, the team proposed a new framework to evaluate their ethical decision-making skills. In the results, the 700 billion-parameter LlamaChat model outperformed larger models. This challenges the long-held belief that bigger is better and the community's excessive reliance on large parameters.

As large tech companies continue to face the consequences of their irrational language models, Microsoft's strategy seems to be a cautious step forward. They are not rushing to increase the complexity of the models but rather addressing one problem at a time.

More Possibilities

Microsoft has not yet revealed its plans to implement the XOT method in its products. Meanwhile, Google DeepMind, led by CEO Demis Hassabis, is considering integrating concepts inspired by AlphaGo into its Gemini project, as mentioned in an interview.

Meta's CICERO project, named after the famous Roman orator, also entered the competition a year ago, astonishing the entire AI community with its exceptional skills in the complex board game "Diplomacy." This game not only requires strategic awareness but also the art of negotiation, long considered a challenge for artificial intelligence. However, CICERO excelled in these areas, demonstrating subtle, human-like conversational abilities.

This discovery has not gone unnoticed, especially considering the benchmarks set by DeepMind. Over the years, the UK-based research lab has used games to develop and refine neural networks.

They set a high standard with AlphaGo, and Meta adopted DeepMind's strategy by combining strategic reasoning algorithms like AlphaGo with natural language processing models like GPT-3.

What sets Meta's model apart is that, for an AI playing Diplomacy, it must not only understand the game rules but also accurately assess the likelihood of betrayal by other human players.

The integration of CICERO's capabilities with Meta's broader AI initiatives may mark the beginning of true conversational artificial intelligence.