5 papers
Universal Transformers for Circuit Computations: Perfect Length Generalization in Tiny Transformers
Takuya Ito, Ruchir Puri, Murray Campbell +1
Learning generalizable algorithmic computations remains a challenge for neural networks, as reflected in persistent failures on compositional and length generalization benchmarks.…
The Need for Verification in AI-Driven Scientific Discovery
Cristina Cornelio, Takuya Ito, Ryan Cory-Wright +2
Artificial intelligence (AI) is transforming the practice of science. Machine learning and large language models (LLMs) can generate hypotheses at a scale and speed far exceeding t…
Transformer Circuits Can Realize Clustering Algorithms
Kenneth L. Clarkson, Lior Horesh, Takuya Ito +2
Although transformers are most commonly optimized as statistical sequence models, it is unclear to what extent they can implement and learn exact algorithmic computations. Here, we…
Quantifying artificial intelligence through algorithmic generalization
Takuya Ito, Murray Campbell, Lior Horesh +2
The rapid development of artificial intelligence (AI) systems has created an urgent need for their scientific quantification. While their fluency across a variety of domains is imp…
Learning interpretable positional encodings in transformers depends on initialization
Takuya Ito, Luca Cocchi, Tim Klinger +3
In transformers, the positional encoding (PE) provides essential information that distinguishes the position and order amongst tokens in a sequence. Most prior investigations of PE…