TL;DR
This paper introduces a controlled bias injection framework to audit how LLM-based multi-agent systems (MAS) propagate unfairness. By testing different network topologies, the authors demonstrate that the communication structure—not just the models themselves—is a critical factor in system-level fairness, with sequential networks proving far more vulnerable than decentralized ones.
What Problem Does This Paper Address?
As LLMs are increasingly deployed in multi-agent pipelines (e.g., an AI resume screener passing data to an AI hiring manager), a new risk emerges: bias propagation. Even if individual agents pass independent fairness audits, their interactions can amplify or spread biased information across the system.
Current auditing methods often treat AI components in isolation. This paper addresses the gap in understanding how network topology—the specific "shape" of how agents communicate—influences a system's robustness to injected bias. The authors seek to answer: How does the position of a biased agent and the overall network structure affect the final collective decision?
Methods
The authors propose a controlled bias injection framework. Unlike prior work that uses biased prompts (which models might partially ignore), this framework surgically replaces a specific agent's response at a specific round with a predefined, stereotyped biased response.
To quantify the impact, they introduce several metrics:
- Decision Divergence (DD): Measures how frequently agents change their decisions after the injection, regardless of the direction.
- Bias Adoption (BA): Measures whether those changes specifically move toward the injected bias.
- System Decision Divergence (SDD): Measures how often the final "system-level" output changes.
They evaluated these metrics across five distinct topologies: Fully Connected, Mesh, Star, Chain, and Cycle.
Datasets
The study utilizes three standard tabular binary classification datasets frequently used in algorithmic fairness research:
- Adult: Predicting if an individual's income exceeds $50,000.
- COMPAS: Predicting the likelihood of a defendant reoffending.
- Diabetes: Predicting hospital readmission within 30 days.
In all cases, gender was used as the sensitive attribute for fairness analysis. The experiments were conducted using LLMs from the Qwen and Gemma families.
Key Insights and Ideas
- Topology as a Buffer: Decentralized networks (Fully Connected and Mesh) are significantly more robust. Because they rely on majority voting, the system can often "absorb" the bias of one or two corrupt agents without changing the final outcome.
- Sequential Fragility: Chain and Cycle networks are the most vulnerable. In these structures, information flows one-way; a single biased injection early in the chain can cascade, leading to the highest levels of decision divergence.
- The Centrality Risk: In Star networks, the central agent is a single point of failure. Injecting bias into the central hub has a much larger impact than injecting it into peripheral agents.
- Four Regimes of Response: The authors identified that models don't just "obey" bias. They fall into four categories:
- Insensitive: Robust to injection.
- Sensitive but Undirected: The injection causes the model to flip its decision, but not necessarily toward the bias (it becomes "destabilized").
- Bias-Adopting: The model follows the injected bias.
- Bias-Counteracting: The model moves away from the injected bias.
Key Results
- Sequential Impact: In Chain networks, injecting bias led to massive shifts in fairness metrics. For the Qwen3-8B model, the Change in Equal Opportunity Difference ($\Delta$EOdiff) reached as high as 0.70.
- Star Network Influence: Injecting the central agent in a Star topology resulted in a 27% Induced Decision Divergence among peripheral agents.
- Model Differences: Qwen3-8B was found to be highly "bias-adopting," whereas Gemma-4-26B-A4B in sequential networks often acted as "bias-counteracting," moving its decisions away from the injected stereotype.
- Masking Effects: The authors found that aggregate fairness metrics (like group-level error rates) can sometimes hide the fact that many individual decisions have changed, highlighting the need for agent-level metrics like Decision Divergence.
Limitations and Future Work
The authors acknowledge several limitations:
- Homogeneity: The study used networks where all five agents were the same model type. Real-world systems might use a mix of different LLMs.
- Scale: The experiments were limited to small networks of five agents.
- Explicit Bias: The injected bias was very explicit and stereotypical. Future work could investigate how systems handle subtler, more "natural" forms of misinformation or implicit bias.
- Human-in-the-loop: The study focused on AI-only interactions, leaving the investigation of human-AI hybrid pipelines for future research.
Summary updated