Imagine that you and a friend are playing a game in which your goal is to communicate secret messages using only cryptic phrases. Your friend's job is to guess the secret message behind your sentences. Sometimes you give clues directly and other times your friend has to guess the message by asking yes or no questions about the clues you've given them. The challenge is that you both want to make sure you understand each other correctly and agree on the secret message.
Researchers at MIT's Computer Science and artificial intelligence Laboratory (CSAIL) have created a similar “game” to help improve the way ai understands and generates text. It is known as a “consensus game” and involves two parts of an artificial intelligence system: one part tries to generate sentences (such as giving clues) and the other part tries to understand and evaluate those sentences (such as guessing the secret message).
The researchers found that by treating this interaction like a game, where both parts of the ai work together under specific rules to agree on the correct message, they could significantly improve the ai's ability to give correct and coherent answers to questions. They tested this new game-like approach on a variety of tasks, such as reading comprehension, math problem solving, and conversation, and found that it helped the ai perform better across the board.
Traditionally, large language models respond in two ways: by generating responses directly from the model (generative query) or by using the model to score a set of predefined responses (discriminative query), which can lead to different results and sometimes incompatible. Using the generative approach, “Who is the president of the United States?” you might give a simple answer like “Joe Biden.” However, a discriminatory query could incorrectly question this fact by evaluating the same answer, such as “Barack Obama.”
So how do we reconcile mutually incompatible scoring procedures to achieve consistent and efficient predictions?
“Imagine a new way to help language models understand and generate text, like a game. We have developed a training-free game theory method that treats the entire process as a complex game of clues and signals, where a generator tries send the right message to a discriminator using natural language, instead of chess pieces, they are using words and sentences,” says Athul Jacob, an MIT doctoral student in electrical and computer engineering and a CSAIL affiliate. “Our way of navigating this game is to find the 'approximate equilibria,' which leads to a new decoding algorithm called 'equilibrium sort.' It's a pretty exciting demonstration of how bringing game theory strategies into the mix can tackle big problems.” challenges to making language models more reliable and consistent.”
When tested on many tasks, such as reading comprehension, common sense reasoning, math problem solving, and dialogue, the team's algorithm consistently improved the performance of these models. Using the ER algorithm with the LLaMA-7B model even dwarfed the results of much larger models. “Given that they are already competitive, people have been working on it for a while, but the level of improvements we saw in being able to outperform a model that is 10 times larger was a pleasant surprise,” says Jacob.
Game on
“Diplomacy”, a strategic board game set in pre-World War I Europe, where players negotiate alliances, betray friends, and conquer territory without the use of dice, relying solely on skill, strategy, and interpersonal manipulation , recently had a second coming. . In November 2022, computer scientists including Jacob developed “Cicero,” an ai agent that achieves human-level capabilities in the seven-player mixed-motive game, requiring the same skills mentioned above, but with natural language. The mathematics behind this partially inspired the Consensus Game.
While the history of ai agents long predates the entry of OpenAI software into chat in November 2022, it is well documented that they can still disguise themselves as your well-intentioned but pathological friend.
The consensus game system achieves equilibrium as an agreement, ensuring accuracy and fidelity to the model's original ideas. To achieve this, the method iteratively adjusts the interactions between the generative and discriminative components until you reach a consensus on a response that accurately reflects reality and aligns with your initial beliefs. This approach effectively bridges the gap between the two query methods.
In practice, implementing the consensus game approach for querying language models, especially for question answering tasks, involves significant computational challenges. For example, when using data sets like MMLU, which have thousands of multiple-choice questions and answers, the model must apply the mechanism to each query. Next, you must reach a consensus between the generative and discriminative components of each question and its possible answers.
The system ran into trouble with one primary school right of passage: the math problems posed. It could not generate incorrect answers, which is a critical component of understanding the process of finding the correct answer.
“The last few years have seen really impressive progress in both strategic decision-making and language generation from ai systems, but we're just starting to figure out how to bring the two together. Equilibrium classification is a first step in this direction, but I think there is a lot we can do to extend this to more complex problems,” says Jacob.
One avenue of future work involves improving the base model by integrating the results of the current method. This is particularly promising as it can produce more objective and consistent responses across various tasks, including factuality and overt generation. The potential for such a method to significantly improve the performance of the base model is high, which could result in more reliable and factual results from ChatGPT and similar language models that people use every day.
“Although modern language models, such as ChatGPT and Gemini, have led to the resolution of various tasks through chat interfaces, the statistical decoding process that generates a response from such models has not changed for decades,” he says. Google research scientist Ahmad Beirami, who was not involved in the work. “The MIT researchers' proposal is an innovative game theory framework for decoding language models by solving the equilibrium of a consensus game. The significant performance improvements reported in the research paper are promising and open the door to a possible paradigm shift in language. decoding models that may drive an avalanche of new applications.”
Jacob wrote the paper with MIT-IBM Watson Laboratory researcher Yikang Shen and MIT Department of Electrical Engineering and Computer Science assistant professors Gabriele Farina and Jacob Andreas, who is also a CSAIL member. They presented their work at the International Conference on Learning Representations (ICLR) earlier this month, where it was highlighted as a “featured paper.” The research also received a “best paper award” at the NeurIPS R0-FoMo workshop in December 2023.
