An arXiv paper proposes substitution rules for prompt context and a validation criterion called selective robustness for social simulation agents.
An arXiv paper proposes substitution rules for prompt context and a validation criterion called selective robustness for social simulation agents.
A paper on arXiv cs.LG addresses why large language models used as agents in social simulations change their behavior substantially under meaning-preserving prompt variations. The authors argue this brittleness creates a validation problem for social simulation and may reflect a lack of grounding. They introduce controlled context enrichment, a set of prompt substitution rules that systematically adds distinct types of context: neutral padding, generic social background, descriptive decision grounding, and normative decision grounding.
The paper also defines selective robustness as a validation criterion: an agent should be insensitive to meaning-preserving prompt variation while showing a directional response to decision-relevant changes in the social world. The authors note that context richness and robustness are rarely defined in ways that support controlled comparisons, which the framework is meant to fix. They demonstrate the approach in a Prisoner's Dilemma setting, according to the excerpt.
The excerpt cuts off before reporting results, so it does not say whether richer context improved robustness in the demonstration, how large any effect was, or which context types mattered. The authors themselves state that whether richer context improves robustness remains unclear. Readers should look for the full paper's experimental findings and any replication before treating the framework as validated.
Source: A Framework for Context Enrichment and Selective Robustness in LLM Social Agents