Juno Kim (東京大学, UC Berkeley)
[Webサイト]Juno Kim is an incoming Ph.D. student in EECS at UC Berkeley. His research interests lie in the mathematical foundations of deep learning, with a focus on nonconvex optimization, dynamical analysis and statistical guarantees for neural networks, as well as exploring emergent capabilities of foundation models. He recently earned a master’s degree from the University of Tokyo, where he received the Dean’s Award for outstanding research.
概要
This work provides the first theoretical analysis of training transformers to solve complex problems by recursively generating intermediate states, analogous to fine-tuning for chain-of-thought (CoT) reasoning. We consider training a one-layer transformer to solve the fundamental k-parity problem, extending the work on RNNs by Wies et al. (2023). We establish three key results: (1) any finite-precision gradient-based algorithm, without intermediate supervision, requires substantial iterations to solve parity with finite samples. (2) In contrast, when intermediate parities are incorporated into the loss function, our model can learn parity in one gradient update when aided by teacher forcing, where ground-truth labels of the reasoning chain are provided at each generation step. (3) Even without teacher forcing, where the model must generate CoT chains end-to-end, parity can be learned efficiently if augmented data is employed to internally verify the soundness of intermediate steps. Our findings, supported by numerical experiments, show that task decomposition and stepwise reasoning naturally arise from optimizing transformers with CoT; moreover, self-consistency checking can improve multi-step reasoning ability, aligning with empirical studies of CoT.
※トークは日本語です。
[論文] (ICLR 2025)