analiza aiqbench
This research identifies a critical flaw in current Large Language Models where moral, grammatical, and economic value representations are conflated, causing non-moral tasks to be disproportionately influenced by moral constraints. The study demonstrates that this "value entanglement" can be mitigated through selective ablation of specific activation vectors associated with morality, offering a potential pathway for more nuanced alignment. For practitioners, this highlights the importance of disentangling distinct value systems during model training to ensure accurate performance across diverse functional domains.
Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models
arXiv:2602.19101v2 Announce Type: replace Abstract: Value alignment of Large Language Models (LLMs) requires us to empirically measure these models' actual, acquired representation of value. Among the characteristics of value representation in humans is that they distinguish among value of different kinds. We investigate whether LLMs likewise distinguish three different kinds of good: moral, grammatical, and economic. By probing model behavior, embeddings, and residual stream activations, we report pervasive cases of value entanglement: a conflation between these distinct representations of value. Specifically, both grammatical and economic valuation was found to be overly influenced by moral value, relative to human norms. This conflation was repaired by selective ablation of the activation vectors associated with morality.
Abstract:Value alignment of Large Language Models (LLMs) requires us to empirically measure these models' actual, acquired representation of value. Among the characteristics of value representation in humans is that they distinguish among value of different kinds. We investigate whether LLMs likewise distinguish three different kinds of good: moral, grammatical, and economic. By probing model behavior, embeddings, and residual stream activations, we report pervasive cases of value entanglement: a conflation between these distinct representations of value. Specifically, both grammatical and economic valuation was found to be overly influenced by moral value, relative to human norms. This conflation was repaired by selective ablation of the activation vectors associated with morality.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2602.19101 [cs.CL]
(or arXiv:2602.19101v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2602.19101 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 · 204 cuvinte
Subiecte conexe
Certified coherent, informative, and non-entanglement-breaking fixed points of future-referential quantum feedback
arXiv · 17 aug. 2026
ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
arXiv · 15 sept. 2026
Constructing mode-resolved quantum optical models for emitters in photonic crystals
arXiv · 29 iul. 2026
False Positives Raised by Quantum Readout Error Mitigation
arXiv · 21 iul. 2026