Small LLMs Match 27B Model via Multi-Agent Cultural Debate

A quiet result from ACL 2025 is worth pausing on if you care about getting language models to behave sensibly outside the culture they were mostly trained in. In a paper published in the ACL Anthology, Dayeon Ki, Rachel Rudinger, Tianyi Zhou and Marine Carpuat report that having several small open-weight LLMs debate each other can match a much larger single model on cultural questions.

The setup is a Multi-Agent Debate framework with two variants. In the first, the agents exclusively debate over cultural scenarios. In the second, each agent dynamically chooses between self-reflection and debate on a given turn. The authors evaluate 7 open-weight LLMs and 21 LLM combinations on NormAd-ETI, a benchmark of social etiquette norms across 75 countries. The headline finding, in their words, is that multi-agent debate enables relatively small LLMs in the 7 to 9 billion parameter range to reach accuracies comparable to a 27 billion parameter model, while also improving cultural group parity over single-model baselines.

Why that is interesting if you are not writing papers: cultural adaptation has usually been argued as a scale problem, so bigger model, better etiquette. A result like this says at least part of the gap can be closed by running two or three smaller open-weight models against each other and letting them disagree. For a team localising a product to markets where the paid frontier APIs are pricey or unavailable, that changes the shortlist of viable options.

The honest caveat is that a benchmark win is not the same as production robustness. NormAd-ETI is a curated set of etiquette scenarios, and the paper as summarised does not give a per-country breakdown, real inference cost for the debate loop, or evidence on open-ended cultural generation. Parity across 75 countries can still hide worse performance on specific low-resource cultures the benchmark under-samples, so take the specifics as reported rather than settled.

The direction, though, is the part worth watching. If small ensembles of open-weight models can credibly stand in for a larger single model on a fairness-flavoured task, the interesting engineering question shifts away from parameter count and toward how you orchestrate disagreement between models cheaply enough to actually ship.