Imagine you and a friend are playing a game where your goal is to speak secret messages to one another using only cryptic sentences. Your friend’s job is to guess the key 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 on the clues you have given. The challenge is that each of you ought to make certain you are understanding one another appropriately and agreeing on the key message.
MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers have created an identical “game” to assist improve how AI understands and generates text. It’s often known as a “consensus game” and it involves two parts of an AI system — one part tries to generate sentences (like giving clues), and the opposite part tries to grasp and evaluate those sentences (like guessing the key message).
The researchers discovered that by treating this interaction as a game, where each parts of the AI work together under specific rules to agree on the proper message, they may significantly improve the AI’s ability to present correct and coherent answers to questions. They tested this latest game-like approach on quite a lot of tasks, reminiscent of reading comprehension, solving math problems, and carrying on conversations, and located that it helped the AI perform higher across the board.
Traditionally, large language models answer one among two ways: generating answers directly from the model (generative querying) or using the model to attain a set of predefined answers (discriminative querying), which might result in differing and sometimes incompatible results. With the generative approach, “Who’s the president of the USA?” might yield an easy answer like “Joe Biden.” Nevertheless, a discriminative query could incorrectly dispute this fact when evaluating the identical answer, reminiscent of “Barack Obama.”
So, how can we reconcile mutually incompatible scoring procedures to attain coherent, efficient predictions?
“Imagine a brand new solution to help language models understand and generate text, like a game. We have developed a training-free, game-theoretic method that treats the entire process as a posh game of clues and signals, where a generator tries to send the proper message to a discriminator using natural language. As an alternative of chess pieces, they’re using words and sentences,” says Athul Jacob, an MIT PhD student in electrical engineering and computer science and CSAIL affiliate. “Our solution to navigate this game is finding the ‘approximate equilibria,’ resulting in a brand new decoding algorithm called ‘equilibrium rating.’ It’s a reasonably exciting demonstration of how bringing game-theoretic strategies into the combination can tackle some big challenges in making language models more reliable and consistent.”
When tested across many tasks, like reading comprehension, commonsense reasoning, math problem-solving, and dialogue, the team’s algorithm consistently improved how well these models performed. Using the ER algorithm with the LLaMA-7B model even outshone the outcomes from much larger models. “On condition that they’re already competitive, that individuals have been working on it for some time, but the extent of improvements we saw with the ability to outperform a model that is 10 times the dimensions was a pleasing 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 territories without the usage of dice — relying purely 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 within the mixed-motive seven-player game, which requires the identical aforementioned skills, but with natural language. The mathematics behind this partially inspired the Consensus Game.
While the history of AI agents long predates when OpenAI’s software entered the chat in November 2022, it’s well documented that they’ll still cosplay as your well-meaning, yet pathological friend.
The consensus game system reaches equilibrium as an agreement, ensuring accuracy and fidelity to the model’s original insights. To realize this, the strategy iteratively adjusts the interactions between the generative and discriminative components until they reach a consensus on a solution that accurately reflects reality and aligns with their initial beliefs. This approach effectively bridges the gap between the 2 querying methods.
In practice, implementing the consensus game approach to language model querying, especially for question-answering tasks, does involve significant computational challenges. For instance, when using datasets like MMLU, which have hundreds of questions and multiple-choice answers, the model must apply the mechanism to every query. Then, it must reach a consensus between the generative and discriminative components for each query and its possible answers.
The system did struggle with a grade school right of passage: math word problems. It couldn’t generate mistaken answers, which is a critical component of understanding the means of coming up with the proper one.
“The previous few years have seen really impressive progress in each strategic decision-making and language generation from AI systems, but we’re just beginning to work out how you can put the 2 together. Equilibrium rating is a primary step on this direction, but I believe there’s lots we’ll give you the chance to do to scale this as much as more complex problems,” says Jacob.
An avenue of future work involves enhancing the bottom model by integrating the outputs of the present method. This is especially promising since it could yield more factual and consistent answers across various tasks, including factuality and open-ended generation. The potential for such a technique to significantly improve the bottom model’s performance is high, which could end in more reliable and factual outputs from ChatGPT and similar language models that individuals use each day.
“Despite the fact that modern language models, reminiscent of ChatGPT and Gemini, have led to solving various tasks through chat interfaces, the statistical decoding process that generates a response from such models has remained unchanged for many years,” says Google Research Scientist Ahmad Beirami, who was not involved within the work. “The proposal by the MIT researchers is an modern game-theoretic framework for decoding from language models through solving the equilibrium of a consensus game. The numerous performance gains reported within the research paper are promising, opening the door to a possible paradigm shift in language model decoding which will fuel a flurry of latest applications.”
Jacob wrote the paper with MIT-IBM Watson Lab researcher Yikang Shen and MIT Department of Electrical Engineering and Computer Science assistant professors Gabriele Farina and Jacob Andreas, who can be a CSAIL member. They presented their work on the International Conference on Learning Representations (ICLR) earlier this month, where it was highlighted as a “highlight paper.” The research also received a “best paper award” on the NeurIPS R0-FoMo Workshop in December 2023.
