Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

TL;DR

This paper investigates whether weak supervision from less capable models can effectively train stronger models, finding that naive finetuning often improves performance but additional techniques are needed for full capabilities.

Contribution

It introduces the concept of weak-to-strong generalization and demonstrates that simple methods can significantly enhance model capabilities under weak supervision.

Findings

Naive finetuning on weak model labels improves strong model performance.

Simple auxiliary methods can recover near state-of-the-art performance.

Weak supervision can be a feasible approach for aligning superhuman models.

Abstract

Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior - for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language models in the GPT-4 family on natural language processing (NLP), chess, and reward modeling tasks. We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call weak-to-strong generalization. However, we are still far from recovering the full capabilities of strong models with naive finetuning alone, suggesting that techniques like RLHF may scale poorly to superhuman models without further work. We find that simple methods can often significantly improve weak-to-strong generalization: for example, when finetuning GPT-4 with a GPT-2-level supervisor and an auxiliary confidence loss, we can recover close to GPT-3.5-level performance on NLP tasks. Our results suggest that it is feasible to make empirical progress today on a fundamental challenge of aligning superhuman models.