Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics

Auto-regressive large language models (LLMs) show impressive capacities to solve many complex reasoning tasks while struggling with some simple logical reasoning tasks such as inverse search: when trained on ''A is B'', LLM fails to directly conclude ''B is A'' during inference, which is known as the ''reversal curse'' (Berglund et al., 2023). In this paper, we theoretically analyze the reversal curse via the training dynamics of (stochastic) gradient descent for two auto-regressive models: (1) a bilinear model that can be viewed as a simplification of a one-layer transformer; (2) one-layer transformers using the framework of Tian et al. (2023a). Our analysis reveals a core reason why the reversal curse happens: the (effective) weights of both auto-regressive models show asymmetry, i.e., the increase of weights from a token $A$ to token $B$ during training does not necessarily cause the increase of the weights from $B$ to $A$. Moreover, our analysis can be naturally applied to other logical reasoning tasks such as chain-of-thought (COT) (Wei et al., 2022b). We show the necessity of COT, i.e., a model trained on ''$A \to B$'' and ''$B \to C$'' fails to directly conclude ''$A \to C$'' without COT (also empirically observed by Allen-Zhu and Li (2023)), for one-layer transformers via training dynamics, which provides a new perspective different from previous work (Feng et al., 2024) that focuses on expressivity. Finally, we also conduct experiments to validate our theory on multi-layer transformers under different settings.
Comment: 40 pages, 15 figures