Amortizing intractable inference in large language models
Abstract
Autoregressive large language models (LLMs) compress knowledge from their training data through next-token conditional distributions. This limits tractable querying of this knowledge to start-to-end autoregressive sampling. However, many tasks of interest -- including sequence continuation, infilling, and other forms of constrained generation -- involve sampling from intractable posterior distributions. We address this limitation by using amortized Bayesian inference to sample from these intractable posteriors. Such amortization is algorithmically achieved by fine-tuning LLMs via diversity-seeking reinforcement learning algorithms: generative flow networks (GFlowNets). We empirically demonstrate that this distribution-matching paradigm of LLM fine-tuning can serve as an effective alternative to maximum-likelihood training and reward-maximizing policy optimization. As an important application, we interpret chain-of-thought reasoning as a latent variable modeling problem and demonstrate that our approach enables data-efficient adaptation of LLMs to tasks that require multi-step rationalization and tool use.
Cite
@article{arxiv.2310.04363,
title = {Amortizing intractable inference in large language models},
author = {Edward J. Hu and Moksh Jain and Eric Elmoznino and Younesse Kaddar and Guillaume Lajoie and Yoshua Bengio and Nikolay Malkin},
journal= {arXiv preprint arXiv:2310.04363},
year = {2024}
}
Comments
ICLR 2024; 23 pages; code: https://github.com/GFNOrg/gfn-lm-tuning