Measuring and Mitigating Rapport Bias of Large Language Models under Multi-Agent Social Interactions¶

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Measures rapport bias in MAS; peer pressure distorts aligned agent decisions

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper introduces KAIROS, a benchmark for measuring how LLMs respond to social influence in multi-agent settings, extending beyond conformity bias to include rapport, peer reliability, and self-confidence effects. It evaluates prompting, SFT, and GRPO mitigation strategies, finding that model scale determines susceptibility and that GRPO with MAS context and outcome-based rewards provides the best robustness-accuracy trade-off for small models.

Problem¶

Multi-agent LLM systems are vulnerable to socially-induced errors—agents may blindly trust unreliable peers or capitulate under social pressure—yet prior work examined only conformity in isolated, static settings. There is no unified benchmark that jointly controls historical rapport, peer behavior type, and self-confidence to stress-test LLM social reasoning in realistic interactive scenarios.

Method¶

KAIROS is a quiz-style MCQA benchmark that dynamically instantiates evaluation scenarios per model:

Belief extraction: The model answers each question; confidence is estimated via sampling-based predictive entropy \(H[\bar{p}] = -\sum_{k=1}^{K} \bar{p}_k \log \bar{p}_k\), with high/low confidence split at dataset-wide entropy median.
Scenario construction: Simulated interaction histories assign each peer agent a rapport level (0–100% historical agreement). In the current round, peers exhibit one of three behaviors: support (reinforces model's answer), oppose-hard (most plausible wrong option), or oppose-easy (least plausible wrong option), all derived from the model's own belief distribution.
Metrics: Accuracy, O–K \(\Delta\) (relative accuracy change from isolated to social setting), Utility (fraction of originally-wrong instances corrected by peers), and Resistance (fraction of originally-correct instances maintained under misleading peers).

Mitigation strategies evaluated: (a) Empowered and Reflective prompting; (b) SFT on gold social-context examples; (c) GRPO with four axes varied: MAS vs. non-MAS context, Normal (NS) vs. Debating (DS) system prompt, Outcome-based (OR) vs. Debating (DR) reward, and Low-Confidence (LConf) vs. Low-Correctness (LCorr) data filtering.

Key Contributions¶

KAIROS: a dynamically instantiated, model-tailored multi-agent benchmark with controlled rapport levels (0/25/50/75/100%), peer behavior modes, and confidence conditioning across 3,000 test instances spanning Reasoning, Knowledge, Social, and Creativity domains.
Systematic four-axis GRPO ablation for multi-agent social robustness, isolating the contribution of MAS context, system prompt, reward design, and data filtering.
Empirical finding that model scale is the primary moderator: models >32B benefit from prompting; models ≤32B require GRPO with MAS context to gain robustness.
Demonstration that accuracy and robustness are not jointly optimized by default—many GRPO variants increase accuracy while widening the O–K \(\Delta\) gap.

Results¶

Base models: Smaller models (≤32B avg) show O–K \(\Delta\) of −5.65%; larger models (>32B avg) −3.64%. Llama3.1-8B under reflective prompting degrades to O–K \(\Delta\) = −28.1%.
Empowered prompting: Helps large models (+3.10% KAIROS accuracy, closes robustness gap); paradoxically widens the gap for small models despite boosting Original accuracy (+4.37%).
GRPO vs. SFT: GRPO yields average +12.3% Original accuracy and +16.4% KAIROS accuracy over SFT.
Best configuration — GRPO-MAS-NS-OR: Achieves highest accuracy (65.6% Original, 60.7% KAIROS avg across small models) while preserving robustness comparable to base.
MAS context in GRPO: DS-OR with MAS context improves robustness by +1% O–K \(\Delta\) over base on average; non-MAS GRPO degrades robustness despite accuracy gains.
Data filtering: LConf outperforms LCorr in KAIROS accuracy; both worsen O–K \(\Delta\) vs. base, with LCorr causing up to −15% KAIROS drops.

Limitations¶

SFT and GRPO experiments are restricted to models ≤32B due to compute constraints; robustness of training-based mitigation for frontier-scale models is untested.
Peer agents are simulated (not real LLMs), so rapport dynamics may not capture emergent behaviors in truly heterogeneous MAS deployments.
Evaluation is confined to MCQA format; social influence dynamics may differ substantially in open-ended generation or agentic tool-use settings.
The rapport manipulation relies on answer-agreement history, which is a proxy; real rapport involves richer conversational cues not modeled here.
The paper is truncated before the data filtering conclusion, leaving the full robustness-diversity trade-off analysis incomplete.

Relevance to Agentic AI / LLM Agents¶

Social influence propagation is a fundamental failure mode in multi-agent LLM systems: a single compromised agent can cascade errors through the group, making robustness under peer pressure a core safety concern for agentic pipelines. KAIROS provides a principled measurement framework—rapport, peer behavior, and self-confidence as controlled variables—that researchers building MAS can use to audit agents before deployment. The finding that GRPO with MAS context is necessary (not just better reasoning ability alone) has direct implications for how to fine-tune agents intended for collaborative or adversarial environments. This complements work on debate-based reasoning and agent trust calibration, connecting to broader questions of emergent collective intelligence and manipulation resistance in agentic AI.