analiza aiqbench
This research introduces SKIN-DEEP, a geometric diagnostic tool that analyzes residual-stream activations to predict how benign fine-tuning might weaken safety refusals in large language models. By identifying low-rank separation patterns between aligned and base checkpoints, the study provides an activation-based method to assess alignment fragility before behavioral interventions occur. These findings offer a critical technical advance for understanding the stability of model safety representations under standard training updates.
Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations
arXiv:2606.22676v2 Announce Type: replace Abstract: Refusal on a safety benchmark does not reveal how stable that behavior will remain after model updates. Benign downstream fine-tuning can weaken refusal, yet behavioral evaluations typically expose this fragility only after an intervention. We introduce SKIN-DEEP, a geometric diagnostic that examines the unmodified model's residual-stream activations. It compares aligned and base checkpoints to identify safety-separating directions, tests their behavioral relevance through ablation, and summarizes the layer-wise pattern in the Geometric Fragility Score (GFS). Across twenty-one instruction-tuned models, harmful requests and benign instructions exhibit a recurring low-rank separation pattern. Selected direction ablations weaken refusal, with the effective direction varying across models. In benign low-rank fine-tuning experiments, the initially safe model with the lowest score before fine-tuning has the lowest harmful-compliance rate when trained on the largest tested set of harmless examples. These findings connect representation geometry to subsequent behavioral susceptibility and support activation-based diagnostics as a complement to refusal tests. Our code is available at https://github.com/js-lee-AI/skin-deep.
Abstract:Refusal on a safety benchmark does not reveal how stable that behavior will remain after model updates. Benign downstream fine-tuning can weaken refusal, yet behavioral evaluations typically expose this fragility only after an intervention. We introduce SKIN-DEEP, a geometric diagnostic that examines the unmodified model's residual-stream activations. It compares aligned and base checkpoints to identify safety-separating directions, tests their behavioral relevance through ablation, and summarizes the layer-wise pattern in the Geometric Fragility Score (GFS). Across twenty-one instruction-tuned models, harmful requests and benign instructions exhibit a recurring low-rank separation pattern. Selected direction ablations weaken refusal, with the effective direction varying across models. In benign low-rank fine-tuning experiments, the initially safe model with the lowest score before fine-tuning has the lowest harmful-compliance rate when trained on the largest tested set of harmless examples. These findings connect representation geometry to subsequent behavioral susceptibility and support activation-based diagnostics as a complement to refusal tests. Our code is available at this https URL.
Accepted to Findings of AACL-IJCNLP 2026. 14 pages, 4 figures, 10 tables. The first two authors contributed equally. Code: this https URL
Artificial Intelligence (cs.AI)
arXiv:2606.22676 [cs.AI]
(or arXiv:2606.22676v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2606.22676 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.
Text integral sub licența deschisă a publicației · 272 cuvinte
Subiecte conexe
Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting
arXiv · 31 iul. 2026
When Partial Credit Opposes Completion
Snorkel AI Blog · 15 oct. 2026
What can language models teach us about understanding?
LessWrong · 9 oct. 2026
Amazon is investing in the Lean Focused Research Organization
Amazon Science · 26 iul. 2026