AI TrackerFrontier desk · AEST
Back to news
normalaitech

Predicting Alignment Generalization with Value Representations

arXiv cs.AI · 9 Oct 2026, 03:47 AEST

The paper discusses the task of predicting alignment generalization in language models. It highlights the effectiveness of model activations over textual descriptions in understanding how fine-tuning affects behavior across various values.

Key facts

Read aloud

The paper discusses the task of predicting alignment generalization in language models. It highlights the effectiveness of model activations over textual descriptions in understanding how fine-tuning affects behavior across various values. L L M developers post-train their models to exhibit prosocial values and behavioral traits, which are enumerated in an alignment target. We conduct a large-scale analysis of alignment generalization effects across sixty six values found in modern alignment targets, and benchmark representational techniques on the alignment generalization prediction task. We find that representations based on model activations when applying values in context significantly outperform methods based on textual descriptions of the values. Specifically, the best activations-based methods achieve correlations of zero point four five with our generalization matrix, compared with zero point zero five from description-based baselines. Our work demonstrates the importance of studying value generalization in L L M s and its application toward the more empirical design and training of model behavior.

Sources in this cluster