New research shows that generative AI models can surpass their human trainers. The researchers call this phenomenon "transcendence" and demonstrate it using the example of chess.
AI models are typically trained to imitate human behavior. However, is it possible for these models to outperform their human "trainers" in certain areas? Researchers from Harvard University, UC Santa Barbara, and Princeton University show in a new study that this is possible through what they call "transcendence."
The researchers trained a transformer on chess games played by players with limited skill levels. The resulting model, called "ChessFormer," was able to play better than all the players in the training dataset in some cases.
According to the team, this transcendence is made possible by low-temperature sampling, where only the token with the highest probability is always chosen. With a low temperature, a kind of majority decision effectively takes place, compensating for any errors made by individual experts and thus raising the model's performance on average above even the best expert's performance.
ChessFormer surpasses the performance of the training data
In their experiments, the researchers then empirically demonstrate this effect. They trained several ChessFormer models on games played by players with maximum ELO ratings of 1000, 1300, and 1500, respectively. They found that the ChessFormer 1000 and ChessFormer 1300 models could achieve ELO ratings of up to 1500 at low temperatures, significantly higher than the maximum of the training data. The researchers show that the performance improvement at low temperatures is primarily due to significantly better moves in a few key game situations - presumably crucial moments that decide the outcome of a game.
Furthermore, the scientists found that data diversity is a necessary condition for an effective majority decision in practice. The model trained only on players up to 1500 rating points could not surpass its human trainers. The researchers attribute this to a lack of diversity in this dataset: players with higher ELO ratings make fewer mistakes that could be compensated for by a majority decision.
Low-temperature sampling leads to "de-noising of errors"
The study shows that it is possible to develop AI models that not only imitate human experts but can even surpass their abilities in certain areas.
However, the authors point out that their results do not provide evidence of novel abstract thinking processes in AI, but rather can be attributed to a denoising effect. "We want to emphasize that we do not provide evidence that low-temperature sampling leads to novel abstract reasoning, but rather to a denoising of errors," they explain.
