arXiv:2405.17399cs.LG
Transformers Can Do Arithmetic with the Right Embeddings
Abstract
The poor performance of transformers on arithmetic tasks seems to stem in large part from their inability to keep track of the exact position of each digit inside of a large span of digits. We mend this problem by adding an embedding to each digit that encodes its position relative to the start of the number. In addition to the boost these embeddings provide on their own, we show that this fix enables architectural modifications such as input injection and recurrent layers to improve performance even further. With positions resolved, we can study the logical extrapolation ability of transformers. Can they solve arithmetic problems that are larger and more complex than those in their training data? We find that training on only 20 digit numbers with a single GPU for one day, we can reach state-of-the-art performance, achieving up to 99% accuracy on 100 digit addition problems. Finally, we show that these gains in numeracy also unlock improvements on other multi-step reasoning tasks including sorting and multiplication.
> there are some benchmarks which show fundamental inability of LLM perform certain tasks which human can, for example add 100 digits numbers. fundamental inability ? No. Current Sota LLM (4o, claude, gemini) woes with arithmetic is not a transformer weakness never mind a large…
famouswafflesView on HNParent context
> https://arxiv.org/abs/2405.17399 they built specialized model which is after bunch of trickery still has 99% accuracy(naive model had very low accuracy) on very simple deterministic algo. I also think most of the accuracy came from memorization of training set(model didn't…
riku_ikiView on HNParent context