Implicit Intermediate Supervision for Learning Complex Functions
Primary Area: learning theory
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Large Language Models, Learning Theory, Theory of Deep Learning, Multi-Task Learning, Chain-of-Thought
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Large language model often require intermediate supervision to solve complex task, but we find that this can be done implicitly with multi-task or multi-label training.
Abstract: Large Language models often rely on explicit intermediate step-by-step supervision, such as chain-of-thought, to solve complex tasks. However, this approach necessitates highly curated data and incurs increased inference time costs. In this study, we investigate the potential of implicit intermediate supervision as an alternative, focusing on multi-task and multi-label learning settings. We demonstrate that training on a dataset with a mixture of tasks allows the learner to utilize the solutions of simpler tasks as intermediate steps for solving more complex ones, reducing the reliance on curated data and explicit supervision. In the multi-label setting, the learner can leverage the signal propagated from easily inferred labels to learn targets that require more subtle computations. We present both theoretical and empirical evidence supporting the notion that neural networks can effectively harness such implicit supervision to tackle complex tasks. Our findings suggest that implicit supervision can shed light on how large language models learn complex tasks while potentially offering valuable insights into developing new versatile methods for solving intricate tasks in language modeling.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5472
Loading