Training-Free Cultural Alignment of Large Language Models via Persona Disagreement

The framing as a 'steering problem' rather than a data collection problem is the buried lede here. Most cultural alignment work assumes you need more representative training data; DISCA instead argues you already have the signal and just need to surface it differently at inference time, which is a meaningful reorientation of where the bottleneck actually sits.

This sits in a cluster of inference-time and alignment work we covered on the same day. DGPO ('Beyond Pairwise Preferences') is the most direct neighbor: both papers are trying to extract better alignment signal from existing information rather than collecting more of it, and both treat the alignment problem as one of structured reasoning about disagreement rather than simple preference labeling. The difference is that DGPO targets logical consistency across related queries while DISCA targets value pluralism across geographies. Together they suggest a broader move away from treating alignment as a labeling pipeline toward treating it as a signal-design problem.

The real test is whether DISCA's gains hold when evaluated against native speakers in the target cultures rather than proxy metrics derived from the World Values Survey itself. If independent replication using held-out cultural benchmarks like CULTURALBENCH or similar shows comparable shifts, the method is credible; if results only hold on WVS-adjacent evaluations, the approach may be overfitting to its own grounding data.