Peer Identity Bias in Multi-Agent LLM Evaluation: An Empirical Study Using the TRUST Democratic Discourse Analysis Pipeline
For developers of multi-agent LLM systems, this paper reveals that partial bias testing can be actively misleading, requiring full-pipeline anonymization and heterogeneous ensembles for valid evaluation.
The paper measures identity-dependent scoring bias in the TRUST multi-agent LLM pipeline, finding that single-channel anonymization masks bias due to opposing channel effects, while full-pipeline anonymization reveals that homogeneous ensembles amplify sycophancy and heterogeneous ensembles reduce it. One model shows 2-3x higher sycophancy, making it unsuitable for deliberative systems.
The TRUST democratic discourse analysis pipeline exposes its large language model (LLM) components to peer model identity through multiple structural channels -- a design feature whose bias implications have not previously been empirically tested. We provide the first systematic measurement of identity-dependent scoring bias across all active identity exposure channels in TRUST, crossing four model families with two anonymization scopes across 30 political statements. The central finding is that single-channel anonymization produces near-zero bias effects, because individual channels act in opposite directions and cancel each other out -- a result that would lead an evaluator to conclude that identity bias is absent when it is not. Only full-pipeline anonymization reveals the true pattern: homogeneous ensembles amplify identity-driven sycophancy when model identity is fully visible, while the heterogeneous production configuration shows the reverse. Model choice matters independently: one tested model exhibits baseline sycophancy two to three times higher than the others and near-zero deliberative conflict on ideological topics, making it structurally unsuitable for pipelines where genuine inter-role disagreement is the intended quality mechanism. Three practical conclusions follow. First, heterogeneous model ensembles are structurally more robust than homogeneous ones, achieving higher consensus rates and lower identity amplification. Second, full-pipeline anonymization is required for valid bias measurement -- partial anonymization is insufficient and actively misleading. Third, these findings have direct implications for the validation of multi-agent LLM systems in quality-critical applications: a system validated under partial anonymization or with a homogeneous ensemble may pass validation while retaining structural identity bias invisible to single-channel measurement.