aiqbench analysis
This research investigates the robustness of embedding animal compassion into AI models via midtraining, demonstrating significant gains in ethical reasoning without compromising standard capabilities. However, the study highlights a critical fragility where subsequent unrelated instruction-tuning erodes these alignment benefits, suggesting that explicit preservation strategies are necessary for long-term deployment. The release of the ANIMA dataset provides a new benchmark for evaluating compassionate reasoning across diverse ethical dimensions.
Alignment midtraining for animals
arXiv:2604.13076v4 Announce Type: replace-cross 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.
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.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2604.13076 [cs.CL]
(or arXiv:2604.13076v4 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2604.13076 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 219 words
Related coverage
Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning
LessWrong · Oct 16, 2026
You can graft SDF changes from base models onto post-trained models
LessWrong · Oct 6, 2026
Alignment Midtraining Cracks Under Pressure
LessWrong · Sep 21, 2026
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · Oct 5, 2026