The Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025), was held from December 2-7 in San Diego. The conference brought together leading researchers in artificial intelligence and machine learning from around the world. As one of the top-tier conferences in AI research, NeurIPS 2025 featured a wide range of advanced work, such as generative models, security, reinforcement learning, AI for science and more.

alignAI Doctoral Candidate Cen Lu attended and presented his poster “Chain-of-Model Learning for Language Model” at the poster session. The work proposes a novel Chain-of-Model learning paradigm that introduces causal relationships into hidden states at each layer, allowing progressive model scaling and elastic inference with multiple sub-models of varying sizes. This collaborative research was conducted with Microsoft Research and other leading institutions, and demonstrates how this new paradigm offers greater flexibility in training and deployment efficiency. The poster session provided a fantastic opportunity to engage directly with fellow researchers, receive feedback and discuss the implications of this research within the community.

Beyond presenting his work on Chain-of-Model, the Human-AI Alignment Tutorial at NeurIPS 2025 also left a profound impression on Cen. Current alignment research is undergoing a paradigm shift from traditional unidirectional constraints toward “Bidirectional Human-AI Alignment”. In the future, research will not only be about aligning AI with human values but also about supporting humans in cognitively and behaviourally adapting to AI’s evolution. Particularly insightful was the exploration of “Pluralistic Alignment”, where research employs interactive architectures like “Jury Learning” to integrate diverse societal perspectives, thereby overcoming the limitations of a single “fundamental truth”. Additionally, bridging the “value-action gap” between AI’s declared values and actual behaviour, alongside Yoshua Bengio’s proposal for safety oversight through non-agentic “scientist AI”, represent challenging yet compelling research directions.

Attending NeurIPS 2025 provided Cen with the opportunity to exchange information about the latest advancements in artificial intelligence and made him realise that addressing the challenges facing AI today requires interdisciplinary collaboration. Sharing his research findings and exchanging ideas with researchers worldwide provided insights for future projects, particularly in areas such as AI alignment, model interpretability and the responsible use of AI in real-world scenarios. With the extremely rapid development of AI, conferences like NeurIPS play an important role in knowledge-sharing and building an academic community focused on technological progress and ethical responsibility.

References:

Shen, H., Kalai, A., Gordon, M., & Bengio, Y. (2025, December 2). Human-AI Alignment: Foundations, Methods, Practice, and Challenges [Tutorial]. Conference on Neural Information Processing Systems (NeurIPS 2025), San Diego, CA. https://neurips.cc/virtual/2025/loc/san-diego/109592.

Song, K., Wang, X., Tan, X., Jiang, H., Zhang, C., Shen, Y., Lu, C., Li, Z., Song, Z., Shan, C., Wang, Y., Ren, K., Zheng, X., Qin, T., Yang, Y., Li, D., & Qiu, L. (2025, December 2–7). Chain-of-Model Learning for Language Model [Poster presentation]. Conference on Neural Information Processing Systems (NeurIPS 2025), San Diego, CA, United States. https://neurips.cc/virtual/2025/loc/san-diego/poster/115677.