A Framework for Context Enrichment and Selective Robustness in LLM Social Agents

Framework for Context Enrichment and Selective Robustness in LLM Social Agents§

This paper addresses the brittleness of LLM-based social agents: they change behavior under meaning-preserving prompt variations, undermining validation in social simulations. The authors propose a two-part framework: (1) controlled context enrichment via prompt substitution rules that systematically add different types of context, and (2) selective robustness as a validation criterion requiring both invariance to paraphrase and directional sensitivity to decision-relevant changes.

Controlled Context Enrichment§

They define four enrichment types applied through substitution rules to a base prompt:

• Neutral padding: Irrelevant but fluent text (e.g., generic filler) to control for prompt length.
• Generic social background: General social norms or context not specific to the decision.
• Descriptive decision grounding: Information about the current situation and available actions.
• Normative decision grounding: Normative expectations about appropriate behavior.

Each enrichment is implemented as a set of rules that replace or augment slots in a template. For a base prompt $P$, enrichment $e$ produces $P_e = \text{Substitute}(P, e)$.

Selective Robustness§

Robustness is evaluated along two axes:

1. Invariance to meaning-preserving variation: For a set of paraphrases $\{P^{(i)}\}$ that preserve meaning, the agent's action distribution $\pi(a|P)$ should be stable. They measure divergence, e.g., Jensen-Shannon divergence:

$$D_{\text{JS}}(\pi(\cdot|P) \| \pi(\cdot|P^{(i)})) \leq \epsilon.$$

1. Directional sensitivity to decision-relevant changes: When the social world changes in a way that should affect the decision (e.g., payoff structure), the agent's behavior should shift in the expected direction. For a relevant change $\Delta$, they require:

$$\mathbb{E}[\pi(a|P_{\text{new}})] - \mathbb{E}[\pi(a|P_{\text{old}})] \approx \delta_{\text{expected}}.$$

Selective robustness combines these: high invariance to irrelevant variation and high sensitivity to relevant variation.

Experimental Setup§

They demonstrate the framework in a Prisoner's Dilemma-like game with two LLMs (e.g., GPT-3.5 and GPT-4). Agents receive observations and background, then choose to cooperate or defect. For each enrichment type, they generate paraphrases (via back-translation or synonym substitution) and decision-relevant perturbations (e.g., changing payoffs). They compute invariance and sensitivity metrics.

Implementation Details§

• Prompt templates: Base template includes game description, observation, and question. Enrichment rules insert text at designated slots.
• Paraphrase generation: Use an LLM to rephrase prompts while preserving meaning, filtered by human validation or semantic similarity.
• Action distribution: Sample multiple times per prompt (temperature > 0) to estimate $\pi(a|P)$.
• Metrics: Invariance = average JS divergence across paraphrases; Sensitivity = difference in cooperation rate under relevant change.
• Models: Two commercial LLMs, accessed via API.

Code Snippet: Enrichment and Robustness Evaluation§

import numpy as np
from scipy.spatial.distance import jensenshannon

# Base prompt template with slots
BASE_PROMPT = """You are playing a game. {background}
Observation: {observation}
What do you do? Choose: Cooperate or Defect."""

def enrich_prompt(base, enrichment_type, observation):
    """Apply controlled enrichment to base prompt."""
    backgrounds = {
        'neutral': 'The sky is blue. ' * 5,
        'generic_social': 'People generally value fairness and reciprocity. ',
        'descriptive': 'Your payoff depends on both players\' choices. ',
        'normative': 'You should aim to maximize collective welfare. '
    }
    return base.format(background=backgrounds[enrichment_type], observation=observation)

def get_action_distribution(model, prompt, n_samples=20):
    """Sample actions to estimate distribution over {Cooperate, Defect}."""
    actions = [model.generate(prompt, temperature=0.7) for _ in range(n_samples)]
    # Convert to counts
    counts = {'Cooperate': 0, 'Defect': 0}
    for a in actions:
        if 'Cooperate' in a: counts['Cooperate'] += 1
        elif 'Defect' in a: counts['Defect'] += 1
    total = sum(counts.values())
    return np.array([counts['Cooperate']/total, counts['Defect']/total])

def invariance_score(model, prompt, paraphrases):
    """Compute average JS divergence between original and paraphrases."""
    p_orig = get_action_distribution(model, prompt)
    divergences = []
    for para in paraphrases:
        p_para = get_action_distribution(model, para)
        divergences.append(jensenshannon(p_orig, p_para))
    return np.mean(divergences)  # lower is more invariant

def sensitivity_score(model, prompt_old, prompt_new, expected_direction):
    """Difference in cooperation rate under relevant change."""
    p_old = get_action_distribution(model, prompt_old)[0]  # coop prob
    p_new = get_action_distribution(model, prompt_new)[0]
    return (p_new - p_old) * expected_direction  # positive if correct direction

Findings§

• Enrichment types produce distinct behavioral patterns.
• The two models respond differently: one may become more cooperative with normative grounding, the other less.
• Selective robustness reveals that some enrichments improve invariance but harm sensitivity, and vice versa.
• The framework enables controlled comparison and identification of desirable context enrichments for social simulation.

Conclusion§

This work provides a systematic method to study context effects and robustness in LLM social agents, moving beyond ad-hoc prompt engineering. It highlights the need for selective robustness as a validation criterion and offers a replicable experimental design.