Simplifying the Modeling of Arbitrary Conditionals in Natural Language
For NLP practitioners needing flexible conditional generation, this provides a simple fine-tuning method for existing LLMs to handle arbitrary conditionals.
AC-GPT modifies standard causal Transformers to enable tractable sampling and evaluation of arbitrary conditionals (e.g., text conditioned on past and future tokens) in a single forward pass, outperforming baselines without degrading left-to-right performance.
Causal Transformers model sequences through an autoregressive factorization of the joint distribution, which enables efficient left-to-right decoding and conditional likelihood computation. However, they cannot tractably sample from or evaluate arbitrary conditionals -- e.g., a block of text conditioned on past and future tokens. Recent work aims to solve this problem through novel architectures, but they often lead to sub-optimal modeling of such conditionals and degraded generations. We propose Arbitrary Conditionals GPT (AC-GPT) which introduces a simple modification to standard causal Transformers to enable evaluating and sampling from arbitrary conditionals -- including past, future, and mixed contexts -- within a single forward pass. Unlike prior approaches, our method preserves the standard left-to-right ordering and next-token prediction objective essential for both strong performance and efficient training on natural language. Crucially, this compatibility allows existing LLMs to be fine-tuned for arbitrary conditioning. Our empirical results indicate that our method outperforms baselines on modeling arbitrary conditionals, without degrading standard left-to-right performance.