Discourse-Driven Alignment: Automated Curation of Alignment-Aware Training Corpora via Explainable AI

TL;DR: ELI5: What if you could use an AI to pick the best stories about AI for training another AI to behave well? Experiment: Build a pipeline that uses explainable AI models to score and curate pretraining corpora for alignment-relevant features (e.g., trust, ethics, risk framing), then compare LLMs pretrained on these “alignment-aware” corpora to those using random or manual curation. Hypothesis: Alignment-aware automated curation yields more robustly aligned models than random or naive upsampling.

Research Question: Can automated, explainable-AI-based curation of pretraining corpora for alignment-relevant discourse features improve LLM alignment outcomes compared to manual or random data selection?

Hypothesis: Alignment-aware, explainable AI-driven curation of training data leads to stronger, more robust alignment priors and reduces the risk of hidden misalignment due to overlooked discourse cues.

Experiment Plan: - Develop/finetune explainable AI models to classify and score pretraining texts for alignment-relevant content (e.g., via ethical reasoning, risk framing, trust markers).

• Curate large pretraining datasets using these models, prioritizing high-scoring alignment discourse and minimizing ambiguous/misalignment cues.
• Pretrain LLMs on the curated vs. control corpora and evaluate alignment using behavioral tests and adversarial scenarios.
• Analyze which discourse features most strongly predict robust downstream alignment.

References:

• Tice, C., Radmard, P., Ratnam, S., Kim, A., Africa, D., & O'Brien, K. (2026). Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment.
• Lochab, A., & Zhang, R. (2025). Energy-Based Reward Models for Robust Language Model Alignment. arXiv.org.
• Wang, J., Xue, S., Li, J., & Wang, X. (2025). Diverse Human Value Alignment for Large Language Models via Ethical Reasoning. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society.
• Shu, D., Zhao, H., Hu, J., Liu, W., Cheng, L., & Du, M. (2025). Large Vision-Language Model Alignment and Misalignment: A Survey Through the Lens of Explainability. Conference on Empirical Methods in Natural Language Processing.