Preprint Explores Predicting LLM Alignment Generalization from Value Representations
arXiv preprint from 2026-10-08 introduces alignment generalization prediction task. Activation-based representations of values in context outperform textual descriptions…
- Activation representations when applying values in context correlate 0.45 with observed generalization, versus 0.05 for description-based baselines.
- Value similarity measured via these representations correlates with improved model robustness on multi-value alignment targets.
- Initial evidence points to a shared, model-independent value space supporting an empirical taxonomy of LLM values.
New Task for Predicting Value Generalization
This arXiv preprint, available as abstract only, defines alignment generalization prediction. The task involves forecasting how fine-tuning a model on one specific value from an alignment target will affect its behavior across many held-out values and unseen contexts.
Developers post-train LLMs for prosocial traits, yet narrow training often produces unexpected influences elsewhere. The work addresses this by focusing on empirical prediction rather than benchmark scores alone.
Large-Scale Analysis Across 66 Values
Researchers conducted analysis using 66 values common in modern alignment targets. They built a generalization matrix capturing real effects of training on one value and measuring impacts on others.
Multiple representational techniques were benchmarked for their ability to predict matrix entries. Evidence is limited to the supplied abstract.
Activations Outperform Textual Descriptions
Representations based on model activations, captured while applying values in context, significantly outperformed methods using textual value descriptions.
Best activation-based methods reached 0.45 correlation with the generalization matrix. Description-based baselines managed only 0.05. This indicates internal states carry richer information about value interactions than surface language.
Applications to Robustness and Taxonomy
The representations were applied to measure similarity among values in multi-value alignment targets. Higher similarity correlated significantly with greater model robustness.
Preliminary findings suggest a shared, model-independent value space. This enables the first taxonomy of LLM values grounded in observed generalization dynamics instead of intuition.
Implications for Alignment Research
The preprint underscores the importance of studying value generalization in LLMs for more empirical design of alignment targets and training methods.
While high scores on narrow evaluations are common, better prediction tools could reduce unexpected cross-context effects. Further full-paper review would be needed to assess detailed methods and results.
