Generative AI is usually described as a system that learns by imitation. A model studies examples, identifies patterns, and produces outputs that resemble the behavior found in its training data. New research described by researchers from Harvard University, UC Santa Barbara, and Princeton University tests a sharper question: can an AI model trained on limited human performance sometimes go beyond it?
The researchers say the answer can be yes, at least in the controlled setting of chess. They call the effect "transcendence" and demonstrate it with a transformer model trained on chess games from players with capped skill levels.
What the chess experiment tested
The model at the center of the work is called "ChessFormer." It was trained on games played by players with limited skill levels, rather than on games from the strongest possible players. That setup matters because it makes the result more specific: the model was not simply copying elite examples from its dataset.
In some cases, ChessFormer was able to play better than all the players included in the training dataset. The study therefore challenges a simple view of AI imitation, where a model can only reproduce the ceiling of the examples it was given.
The researchers trained several ChessFormer models using games from players with maximum ELO ratings of 1000, 1300, and 1500, respectively. They then tested how those models performed under different sampling conditions.
Why low-temperature sampling mattered
The key mechanism identified by the researchers is low-temperature sampling. In the source description, this means that the model always chooses the token with the highest probability. In a chess model, that choice affects which move the system selects from the possibilities it has learned to represent.
At low temperatures, the researchers say a kind of majority decision takes place. Individual players in the training data may make mistakes, but those mistakes do not always point in the same direction. If the model consistently favors the most probable choice, errors made by individual experts can be compensated for in the aggregate.
That is the central logic behind the claimed performance gain. The model is not presented as discovering chess from scratch, and the authors do not claim that it has found a new form of intelligence. Instead, the system appears to benefit from combining many imperfect examples in a way that reduces some of their noise.
Where ChessFormer exceeded its data
The empirical result was clearest for the ChessFormer 1000 and ChessFormer 1300 models. At low temperatures, both could achieve ELO ratings of up to 1500. That is significantly higher than the maximum rating of the training data used for those models.
The improvement was not described as a uniform advantage across every position. The researchers found that the performance increase came mainly from significantly better moves in a few key game situations. Those moments were presumably crucial positions that could decide the outcome of a game.
This detail makes the result more grounded. Chess performance can turn on a limited number of decisions, so a model does not need to improve every move equally to produce a better overall result. If it avoids important mistakes at the right moments, its measured strength can rise above the level of the players it learned from.
Why more skilled data did not automatically help
The study also found an important limit. The model trained only on players up to 1500 rating points could not surpass its human trainers. The researchers attribute that failure to a lack of diversity in the dataset.
Their explanation is that players with higher ELO ratings make fewer mistakes that can be corrected through the majority-decision effect. In other words, if the data contains less variation in errors, there may be less noise for the model to filter out. A stronger dataset is not automatically better for this specific kind of transcendence if it does not provide the diversity needed for the effect to appear.
The study therefore points to two conditions working together:
- The model needs enough varied examples to support a useful majority decision.
- The sampling method needs to favor the most probable choice strongly enough to reduce individual errors.
Without those conditions, imitation may remain closer to the performance boundaries of the training data.
What the result does and does not prove
The researchers frame the result carefully. The study shows that an AI model can sometimes outperform the people represented in its training data in a specific domain. It does not show that the model has developed novel abstract thinking.
The authors state the point directly: "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.
That distinction is important for interpreting the finding. The word "transcendence" may sound broad, but the mechanism described in the study is narrower and more technical. ChessFormer appears to surpass its training data by selecting from learned patterns in a way that filters mistakes, not by demonstrating a new kind of reasoning process.
For generative AI, the implication is still significant. A model trained to imitate human behavior may, under certain conditions, produce better outcomes than any individual source in the dataset. But the chess example also shows why the details matter: data diversity, sampling temperature, and the structure of the task all shape whether that improvement appears.