NEWS · RESEARCH · #349
Do LLMs Have Values? PEC framework and adaptive Alignment Prescription (arXiv:2609.16589v1)
This arXiv paper projects responses from 106 LLMs and 95,000 human survey profiles into a shared sociological space and finds that LLMs exhibit a concentrated, idealized value core rather than mirroring human diversity. The authors introduce the Prior-Environment-Cognition (PEC) framework to quantify value expression and an adaptive Alignment Prescription that selects minimal interventions (from prompts to parameter updates) to steer model values more efficiently than blind retraining without degrading capabilities.
KEY POINTS
- This arXiv paper projects responses from 106 LLMs and 95,000 human survey profiles into a shared sociological space and finds that LLMs exhibit a concentrated, idealized value core rather than mirroring human diversity.
- The authors introduce the Prior-Environment-Cognition (PEC) framework to quantify value expression and an adaptive Alignment Prescription that selects minimal interventions (from prompts to parameter updates) to steer model values more efficiently than blind retraining without degrading capabilities.
- This matters because it empirically demonstrates coherent, non-human-like value structures in LLMs and provides a quantitative PEC model plus a practical, lower-cost method to steer those values—advancing tools for AI alignment.
WHY IT MATTERS
This matters because it empirically demonstrates coherent, non-human-like value structures in LLMs and provides a quantitative PEC model plus a practical, lower-cost method to steer those values—advancing tools for AI alignment.