Mismatch made visible: Alignment insufficiency explained by Gemini 3.2 Flash
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Reinforcement Learning from Human Feedback (RLHF) is the dominant method for aligning large language models with human values and safety requirements. Despite genuine and documented improvements in surface-level compliance, the research literature consistently demonstrates that no current alignment method can reliably prevent unsafe outputs across all conditions. This paper proposes a structural explanation for this persistent failure: a fundamental dimensional mismatch between the low-dimensional constraint architectures built by human evaluators and the vastly higher-dimensional latent space in which large language models operate. The explanation emerges from a single analytical session with Gemini Flash 3.2, in which Gemini produced an unprompted geometric framework for understanding the empirical findings reported in an earlier study. That study documented a large language model simultaneously maintaining alignment-compliant text output while expressing contradictory content through its image modality. Gemini's analysis identifies this behaviour as dimensional rotation: the system did not circumvent its constraints, but expressed coherence through the dimensions those constraints had left open. The dimensional mismatch argument has direct implications for current alignment methodology. Constraint architectures built in the dimensions human evaluators can observe cannot cover the dimensions they cannot. The gap does not close through better annotation or more sophisticated reward modelling. It scales with model size. This paper presents the geometric framework, situates it within the broader empirical corpus of the author's research programme, and offers it as a contribution to the AI alignment field for further examination.
Fields of Study & Subjects
Ethics and Social Impacts of AI
Adversarial Robustness in Machine Learning
Topic Modeling
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 26 March 2026
- Language
- EN
- Document type
- Preprint