NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 26·5 min read
Whose Alignment? Comparing LLM Process Alignment Across Diverse Organizational Decision Contexts
TL;DR
Organizations often struggle to ensure that large language models (LLMs) reflect their specific decision-making policies. A decision-policy capturing method was developed to measure how well LLMs align with these policies.
✦ Why It Matters
Engineers can use process alignment methods to ensure LLMs reflect organizational policies while addressing ethical concerns.
Key Takeaways
- Implement decision-policy capturing to measure LLM alignment.
- Evaluate alignment impact on output accuracy in decision-making.
- Consider ethical implications of aligning LLMs with historical policies.
- Adapt LLMs to reflect organizational values and avoid biases.
Full Summary
Organizations increasingly use large language models (LLMs) to make decisions that align with their policies, but existing research typically focuses on individual or demographic perspectives. To address this, a decision-policy capturing method was introduced to evaluate process alignment, which measures how well an LLM reproduces an organization's decision-making policy.
The study found that alignment varied significantly across different models and organizations. For instance, in European Court of Human Rights (ECHR) Article 6 decisions, process alignment correlated strongly with output accuracy (r = 0.85, p < .001).
In contrast, consumer credit decisions exhibited low overall alignment, with models resisting the organization's weighting of protected attributes. This highlights the complexity of aligning LLMs with organizational policies, especially when historical decisions may encode biases.
The findings suggest that organizations must carefully consider which policies to align with and the implications of achieving higher alignment.