The science behind Cabinet.

Four systematic literature reviews examining the foundational questions behind Cabinet's architecture. Each study proposes falsifiable hypotheses and experimental designs.

1. 1

   Does prompt-level persona assignment produce reliably distinct reasoning, or do agents converge on a single model?

   Patel (2026) found cosine similarity of 0.888 between hidden representations of same-model agents with different personas. Understanding the depth of this "representational collapse" would determine whether Cabinet needs to enforce model heterogeneity to deliver genuine diversity.

2. 2

   Is the Umpire's synthesis strategy a stronger determinant of output quality than round count or agent diversity?

   Study III's evidence synthesis points toward yes: the D3 ablation shows aggregation mechanism accounts for +8.8pp while persona diversity adds only +3.8pp. But direct experimental validation on Cabinet's architecture and task distribution remains an open gap.

3. 3

   Can proxy confidence signals close the gap with architecture-level uncertainty methods?

   Study II found DebUnc's Oracle gap leaves 22% relative improvement available through better confidence weighting alone. But text-based proxies captured only 0.01 of that gap. The question is whether consistency-based and semantic-dispersion signals can close more of it without open-weight model access.

4. 4

   Can a lightweight classifier trained on query features predict at submission time whether multi-agent debate will help or hurt for a given query?

   Study IV projects that a DeBERTa-v3-base router can achieve AUROC >0.70 with <50ms latency, reducing inference cost by 40%+ while preserving quality. But no published system has trained a classifier for the debate-vs-single-agent routing decision specifically. The existence proof from HybridLLM and RouteLLM (model-strength routing) does not guarantee transfer to the debate routing problem.

5. 5

   For open-ended, high-stakes tasks, does multi-agent debate reduce answer variance more than it improves accuracy, and does that variance reduction matter to users?

   Most academic benchmarks use structured tasks with correct answers. Cabinet's users ask open-ended strategic questions where noise reduction, not accuracy improvement, may be the primary value.

1. 1.Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I. (2023). "Improving Factuality and Reasoning in Language Models through Multiagent Debate." ICML 2024. arXiv
2. 2.Yoffe, L., Amayuelas, A., & Wang, W. Y. (2024). "DebUnc: Improving Large Language Model Agent Communication With Uncertainty Metrics." EMNLP 2025. arXiv
3. 3.Motwani, S. R., et al. (2024). "MALT: Improving Reasoning with Multi-Agent LLM Training." arXiv
4. 4.Hegazy, M. (2024). "Diversity of Thought Elicits Stronger Reasoning in Multi-Agent Debate Frameworks." arXiv
5. 5.Chen, J. C., Saha, S., & Bansal, M. (2023). "ReConcile: Round-table Conference Improves Reasoning via Consensus Among Diverse LLMs." arXiv
6. 6.Smit, A., et al. (2023). "Should We Be Going MAD? A Look at Multi-Agent Debate Strategies for LLMs." arXiv
7. 7.Zhang, H., et al. (2025). "If Multi-Agent Debate Is the Answer, What Is the Question?" arXiv
8. 8.Harrasse, A., et al. (2026). "Debate, Deliberate, Decide (D3): A Cost-Aware Adversarial Framework for Reliable and Interpretable LLM Evaluation." EACL 2026. arXiv
9. 9.Wang, J., et al. (2024). "Mixture-of-Agents Enhances Large Language Model Capabilities." arXiv
10. 10.Nguyen, P., & Pham, T. (2026). "Toward Reliable Evaluation of LLM-Based Financial Multi-Agent Systems: Taxonomy, Coordination Primacy, and Cost Awareness." arXiv
11. 11.Li, L., et al. (2025). "PartnerMAS: An LLM Hierarchical Multi-Agent Framework for Business Partner Selection on High-Dimensional Features." arXiv
12. 12.Sentosa, A. D., & Widianto, J. (2025). "Multi-Agent Consensus System (MACS) for Bias Mitigation."
13. 13.Patel, D. (2026). "Representational Collapse in Multi-Agent LLM Committees."
14. 14.Malmqvist, L. (2024). "Sycophancy in Large Language Models: Causes and Mitigations." arXiv
15. 15.Wan, D., et al. (2025). "MAMM-Refine: A Recipe for Improving Faithfulness in Generation with Multi-Agent Collaboration." arXiv
16. 16.Wynn, A., Satija, H., & Hadfield, G. K. (2025). "Talk Isn't Always Cheap: Understanding Failure Modes in Multi-Agent Debate." arXiv
17. 17.Wu, H., Li, Z., & Li, L. (2025). "Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning." arXiv
18. 18.Yang, Y., et al. (2025). "Revisiting Multi-Agent Debate as Test-Time Scaling: A Systematic Study of Conditional Effectiveness." ICML 2025. arXiv
19. 19.Kim, Y. & Liu, X. (2026). "Towards a Science of Scaling Agent Systems." Google Research. arXiv
20. 20.Ong, I., et al. (2024). "RouteLLM: Learning to Route LLMs with Preference Data." arXiv
21. 21.Ding, D., et al. (2024). "HybridLLM: Cost-Efficient and Quality-Aware Query Routing." arXiv
22. 22.Gao, C., et al. (2025). "Are Hybrid LLM-Agent Systems Effective for Complex Reasoning? An Empirical Analysis." arXiv
23. 23.Minsky, M. (1986). The Society of Mind. Simon & Schuster.
24. 24.Tetlock, P. E., & Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown.
25. 25.Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. Little, Brown.