Unsupervised Elicitation of Language Models

TL;DR

This paper introduces an unsupervised algorithm called Internal Coherence Maximization (ICM) that fine-tunes pretrained language models using their own generated labels, eliminating the need for external supervision and outperforming human-supervised methods on several tasks.

Contribution

The paper presents ICM, a novel unsupervised fine-tuning method for language models that matches or exceeds performance of supervised training, especially on superhuman capability tasks.

Findings

ICM matches performance of training on golden labels.

ICM outperforms training on crowdsourced labels.

ICM improves training of frontier language models.

Abstract

To steer pretrained language models for downstream tasks, today's post-training paradigm relies on humans to specify desired behaviors. However, for models with superhuman capabilities, it is difficult or impossible to get high-quality human supervision. To address this challenge, we introduce a new unsupervised algorithm, Internal Coherence Maximization (ICM), to fine-tune pretrained language models on their own generated labels, \emph{without external supervision}. On GSM8k-verification, TruthfulQA, and Alpaca reward modeling tasks, our method matches the performance of training on golden labels and outperforms training on crowdsourced human supervision. On tasks where LMs' capabilities are strongly superhuman, our method can elicit those capabilities significantly better than training on human labels. Finally, we show that our method can improve the training of frontier LMs: we use our method to train an unsupervised reward model and use reinforcement learning to train a Claude 4 Sonnet-based assistant. The resulting assistant matches its counterpart trained on production-grade human labels on average, with higher scores on chat and safety yet lower scores on math and coding.