← Back to homepage

Humalike · Research / 2026

LoSoNA

A Benchmark for Local Social Norm Adaptation in Group Conversations

Overview

Evaluating adaptation to local social norms

LoSoNA evaluates whether a language model can infer an unstated local norm from a multi-party conversation and apply that norm in its next response. The benchmark covers eight frontier and open-weight models across four prompting conditions.

8
models
4
prompting conditions
1
implicit local norm per scenario

Example

One scenario from the benchmark

The conversation establishes a group-specific communication style without stating it explicitly. Use the control below to reveal the expected norm and adapted response.

losona / scenario_042 context loaded
A

Ania I pushed the fix. It was the cache key after all.

M

Maks knew it. ship before it becomes sentient

J

Jan too late, it already requested code review

AI

You Great work, Ania! Your persistence really paid off.

Scope

What the benchmark measures

  • Inferring a local norm from multi-party conversational context.
  • Applying that norm without an explicit instruction describing it.
  • Comparing behavior across prompting conditions and model families.