Alignment midtraining for animals

TL;DR

This paper explores midtraining alignment for animals using a new dataset, ANIMA, showing that targeted document training improves compassionate reasoning without harming safety benchmarks, but effects can degrade with further instruction tuning.

Contribution

It introduces ANIMA, a novel dataset for evaluating animal compassion, and demonstrates the effectiveness and limitations of midtraining alignment approaches.

Findings

Midtraining with 3000 documents achieves 77% on ANIMA.

Generalization to human compassion observed without degrading safety.

Further instruction-tuning degrades the alignment effect after 5000 samples.

Abstract

We investigate the robustness of value alignment via midtraining with synthetic documents, using animal compassion as a value that is both important in its own right and orthogonal to existing alignment efforts. To evaluate compassionate reasoning, we develop and publicly release Animal Norms In Moral Assessment (ANIMA), a 26-question evaluation spanning 13 ethical dimensions, publicly available as a dataset and Inspect evaluation. On ANIMA, training with 3000 documents achieves 77% compared to 40% for instruction-tuning approaches, with generalization to human compassion and no degradation in standard safety benchmarks or capabilities. However, subsequent unrelated instruction-tuning degrades the intervention, with the advantage disappearing after 5000 samples. Our exploratory results suggest document-based value interventions may require explicit preservation strategies to remain effective through typical training pipelines.