A Framework for Context Enrichment and Selective Robustness in LLM Social Agents
The paper tests whether richer prompts make simulated LLM agents less brittle. Its selective-robustness criterion may help builders distinguish useful context from wording that merely changes behavior.
The framework compares four prompt enrichments: neutral padding, generic social background, descriptive grounding, and normative grounding. It defines selective robustness as resisting meaning-preserving prompt changes while responding directionally to relevant changes in the world. In a Prisoner’s Dilemma-like game, enrichment produced distinct behavior across **two models**.
When evaluating an agent, vary prompt wording separately from decision-relevant facts. Check that behavior stays stable under the first and changes appropriately under the second; adding background alone is not evidence of better grounding.
The framework compares four prompt enrichments: neutral padding, generic social background, descriptive grounding, and normative grounding. It defines selective robustness as resisting meaning-preserving prompt changes while responding directionally to relevant changes in the world. In a Prisoner’s Dilemma-like game, enrichment produced distinct behavior across **two models**. When evaluating an agent, vary prompt wording separately from decision-relevant facts. Check that behavior stays stable under the first and changes appropriately under the second; adding background alone is not evidence of better grounding. The results are from a social simulation, and the paper says enrichment effects varied by model. The material does not establish which prompt type generalizes to coding agents or real tasks.
This paper separates robustness to meaning-preserving prompt variation from responsiveness to relevant changes in the world, and tests the distinction in a social simulation with two models. It gives an evaluation principle for context behavior, but does not show that any enrichment method transfers to coding agents or real tasks.