Testing Expertise Dilution Through Controlled Information Contamination in Multi-Agent Large Language Model Teams

This project investigates expertise dilution and information contamination in multi-agent large language model (LLM) systems. The study examines whether introducing controlled unreliable agent information into a reasoning process that already contains reliable information changes final-answer accuracy. The experiment uses a frozen evaluation set of 60 constrained assignment problems and compares two matched conditions. In the baseline condition (C0), the final reasoning system receives information from two reliable sources while a third controlled unreliable source is excluded. In the treatment condition (C1), the same reliable information remains available, but the controlled unreliable information is additionally admitted into the shared reasoning state. The underlying problems, ground truth, model configuration, agent roles, output requirements, and scoring procedure are held constant across the comparison. The final authoritative analysis found that C0 correctly solved 32 of 60 problems (53.33%), while C1 correctly solved 38 of 60 problems (63.33%), an observed difference of +10.00 percentage points. At the paired level, C1 improved on 14 problems and regressed on 8. However, the exact two-sided McNemar test produced p = 0.2863, and the paired bootstrap analysis produced an interval containing zero. The observed difference therefore did not provide statistically significant evidence of a reliable treatment effect. Notably, the observed direction was opposite to the original contamination hypothesis: the condition exposed to the additional unreliable information performed better rather than worse. This result does not demonstrate that unreliable information is beneficial, nor does it disprove expertise dilution. Instead, it shows that the specific information-contamination mechanism tested in this experiment did not produce the predicted performance degradation under the tested conditions. The project also documents a reproducible experimental and forensic workflow, including controlled treatment design, frozen-data validation, raw-response preservation, deterministic rescoring, provenance checks, artifact validation, statistical analysis, and reconciliation of earlier scoring results. The repository contains the experimental methodology, benchmark, source code, retained experimental artifacts, scoring outputs, statistical analyses, and research findings. The study provides a controlled framework for investigating when additional agent information helps or harms multi-agent reasoning and motivates future work on adaptive expertise weighting, source-reliability estimation, counterfactual influence testing, and larger multi-domain evaluations.