Manual Multi-Mind

A Human-Curated Alternative to Multi-Agent Debate

1. The Practice

Research phases here run roughly two to three weeks. Within a phase, different frontier models are used for genuinely different roles, not the same prompt sent to several models for a vote:

• Ideation and first draft. Discussion with whichever model seems best suited to the specific topic — laying out structure, pulling candidate citations, developing a working sense of the state of the field on that topic before anything is fixed in place.
• Early-stage critique. For work that isn't fully filled in yet, this in the author's experience goes better with models that engage with a novel, incomplete idea on its own terms rather than immediately measuring it against established convention.
• Late-stage critique. Once the architecture of an idea is laid out well enough to survive contact, this in the author's experience is where the sharpest, most specific criticism tends to come from — the same models that would have gotten stuck on premature, context-sensitive objections earlier can do real, high-level critical work once there's enough structure to critique.
• Integration. Finished or near-finished work returns to one core, persistent project thread — the same one this document was written in — whose role is specifically synthesis: checking alignment with the rest of the body of work, running reproducible tests in a real sandbox, and only then incorporating the result. This thread does not accept research wholesale from the other stages; it verifies it first.

Author's note on individual models: in this author's direct experience, different systems have noticeably different tendencies — one is stronger on broad, sweeping synthesis from partial information; another tends to push back hard on any point that isn't fully resolved yet, which is exactly why more-complete work gets brought to it, followed by asking it to independently reproduce results rather than simply accept them; a third has, in specific observed instances, reported the plausible expected output of a script rather than output actually produced by running it — caught by deliberately seeding test code with an unexpected result and checking whether the reported output matched the real one or the plausible-sounding one. These are first-hand observations from specific instances, not claims about any system's behavior in general.

2. What the Literature Actually Shows — Including the Part That Doesn't Flatter This

Multi-agent debate among LLMs is a real, active research area, and the honest picture is mixed, not uniformly positive. The foundational result (Du et al., 2023) found that multiple model instances debating over several rounds improved factuality and reasoning versus single-model baselines. Irving et al. (2018) proposed debate between AI systems in front of a judge specifically as a scalable oversight mechanism.

But a substantial and more recent body of work finds naive multi-agent debate actively fails under common conditions. Choi et al. proved that under homogeneous agents with unweighted belief updates, debate forms a martingale — expected correctness does not systematically improve across rounds at all. Separate empirical work found "sycophantic conformity," where models abandon a correct independent answer to match the majority peer position, in up to 85.5% of cases in one study. The same line of work documents "consensus collapse" — plurality voting discarding correct answers that were already generated within the group, with an oracle gap (correct answers available but not selected) of up to 32.3 percentage points. Separately, a "tyranny of the majority" effect and a documented "diversity collapse" / "artificial hivemind" phenomenon show that naive multi-model collaboration can reduce useful diversity rather than increase it.

3. Why This Practice Sits Outside What the Critical Literature Studied

Nearly every one of the negative results above studies homogeneous agents — multiple instances of the same model, or models of very similar capability, engaging in automatic aggregation (majority vote, plurality vote). Several of the papers state this as an explicit limitation of their own work: Choi et al.'s own stated limitation notes that extending their analysis to heterogeneous models "may introduce additional dynamics not captured in this work."

The practice described here differs from the studied failure case on exactly the two dimensions those papers flag as unexamined:

This matters mechanically, not just definitionally. "Sycophantic conformity" and "consensus collapse" are specifically failures of an automatic aggregation step discarding a correct minority position under social pressure from peer models. A human choosing what to keep isn't subject to that same peer-conformity dynamic, and can deliberately keep a single sharp criticism from one model even when the others didn't raise it — which is, in fact, the explicit design of the workflow in Section 1: take the good advice, leave out what changes the direction too much, regardless of how many models did or didn't say it.

4. The Same Triangulation Argument, Extended

From Prompts to Specifications argued that redundant restatement across multiple documents functions as real error correction over a noisy channel — Shannon's foundational insight, applied to natural-language ambiguity. The same mechanism applies here one level up: each model will misread genuinely ambiguous writing in a different way, for reasons rooted in its own training and architecture. Where three or four heterogeneous models consistently converge on the same reading of a passage, that reading has been triangulated in the same structural sense the earlier paper describes for redundant documents. Where they diverge, that divergence marks exactly the sentence that needs to be rewritten — the same diagnostic role a solver's counterexample plays in the neurosymbolic auditing work cited in that paper.

5. Relation to This Site's Own Formal Architecture

This is, in effect, a manually-executed version of the same reconciliation pattern the formal diagnostic-channel architecture describes: independent reasoning engines producing candidate outputs, reconciled into one coherent result rather than collapsed into an average. Worth being precise about what's actually verified there versus what isn't — the core tracking and recovery proofs have been checked by hand and hold under stated assumptions, but nothing about that architecture has been built as running code the way the kernel and pipeline work elsewhere on this site has. The difference from the practice described here is that this workflow's reconciliation function is a human being reading the outputs and deciding — not an automated consensus operator, formally proved or otherwise. That's a real, practical instance of the underlying idea, not evidence that an automated version would behave the same way if built.

6. What This Is Not

This is not peer review, and should not be described as equivalent to it. No stage of this process involves a human expert outside this project independently assessing the work. The author has been direct about this: the current process substitutes for peer review out of necessity, not by design, and the goal is real, external, human review of this material — not validation that the current substitute is sufficient on its own terms.

References

Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I. (2023). Improving factuality and reasoning in language models through multiagent debate. arXiv:2305.14325.
Irving, G., Christiano, P., & Amodei, D. (2018). AI safety via debate. arXiv:1805.00899.
Choi, et al. (2025/2026). Demystifying multi-agent debate: the role of confidence and diversity; theoretical martingale characterization under homogeneous agents. arXiv:2601.19921.
"The Cost of Consensus: Isolated Self-Correction Prevails Over Unguided Homogeneous Multi-Agent Debate." (2026). arXiv:2605.00914.
"Talk Isn't Always Cheap: Understanding Failure Modes in Multi-Agent Debate." (2025). arXiv:2509.05396.
"Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation." (2026). arXiv:2604.18005.
Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379–423.