[2506.10139] Unsupervised Elicitation of 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.
Authors:Jiaxin Wen, Zachary Ankner, Arushi Somani, Peter Hase, Samuel Marks, Jacob Goldman-Wetzler, Linda Petrini, Henry Sleight, Collin Burns, He He, Shi Feng, Ethan Perez, Jan Leike View PDF HTML (experimental) 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…
saved by
related reading
- Unsupervised Elicitationalignment.anthropic.com
- Self-CTRL: Self-Consistency Training with Reinforcement Learningarxiv.org
- Discovering Language Model Behaviors with Model-Written Evaluations — LessWronglesswrong.com
- [2005.14165] Language Models are Few-Shot Learnersarxiv.org
- [2511.08579] Training Language Models to Explain Their Own Computationsarxiv.org
- Training Language Models to Explain Their Own Computationsarxiv.org
- Foundation Models for Oversight | Transluce AItransluce.org
- Mechanistically Eliciting Latent Behaviors in Language Models — AI Alignment Forumalignmentforum.org
- [2305.18290] Direct Preference Optimization: Your Language Model is Secretly a Reward Modelarxiv.org
- 2212.03827arxiv.org
- [2212.03827] Discovering Latent Knowledge in Language Models Without Supervisionarxiv.org
- RLHF: Reinforcement Learning from Human Feedbackhuyenchip.com