Gonality as a Predictor of Dissent Survival in Networked Delphi Panels

When several language-model agents deliberate over repeated rounds, a correct minority often gets talked out of its answer. The same failure motivated the Delphi method, the structured survey technique in which anonymous experts rate items, see an aggregate summary, and revise over two or three rounds. Recent multi-agent work shows that who can talk to whom matters: a sparse ring of agents beats a fully connected panel. So far, though, "sparse versus dense" is the only structural variable anyone has tested. This note asks whether a finer graph invariant, divisorial gonality, predicts how long a correct dissenting position survives in a Delphi-style panel of agents, beyond what edge density and ordinary connectivity measures already explain.

Gonality comes from chip-firing on graphs. Place chips on vertices. "Firing" a set of vertices pushes one chip across every edge leaving the set. The gonality of a graph is the fewest chips from which, by legal firings, you can put a chip on any vertex an opponent names. Informally, a graph has low gonality when it can be swept end to end by a moving front whose boundary never has many edges. The hypothesis is that a wrong majority's momentum behaves like such a front. On graphs with low gonality, the front must pass through narrow boundaries, so pockets of correct dissent last longer. The contribution is a controlled test of that hypothesis. We hold agent count, degree, and edge count fixed, vary gonality using graph families whose gonality is known exactly, and ask whether gonality adds anything once spectral gap and diameter are accounted for. If it adds nothing, the divisor lens is decorative and we will say so.

What prior work establishes and what it leaves open

Three results motivate the study. Improving Multi-Agent Debate with Sparse Communication Topology models debate as a graph of agents and dialogue channels. It finds that neighbor-connected rings avoid "premature consensus," with about a 2% accuracy gain on MATH and over 40% fewer input tokens than fully connected debate, and it finds that putting a strong model at a high-centrality node helps by up to 3%. It compares only a clique with a ring, so density and structure are never separated. Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate defines the Disagreement Collapse Rate, the fraction of correct dissenters who abandon their answer. It reports roughly 81% collapse in decentralized debate on CommonsenseQA and 41% with a central judge, and finds that gains stop after two to three rounds. Voting or Consensus? Decision-Making in Multi-Agent Debate shows that agreement-threshold protocols work as filters on knowledge tasks (MMLU 54.0% versus 51.7% for voting) and fail on multistep reasoning (MuSR 28.4% versus 54.8%). It also documents "drift" as rounds accumulate. Together these papers say structure matters and collapse is the main failure mode. None of them names the structural quantity responsible.

One result constrains the study. Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? proves that unguided debate is a martingale: expected correctness does not rise across rounds. Empirically, majority voting over independent answers matches or beats debate, for example 94.0% versus 90.3% for the best debate variant on GSM8K with Qwen2.5-7B. Topology alone therefore cannot be expected to raise expected accuracy. What topology can change is the variance and timing of belief change, meaning whether correct minorities are erased before the round cap. For this reason the primary outcome is dissent survival and collapse rate, with accuracy secondary. A second, exploratory arm injects an external correction signal, which that paper identifies as the only way to break the martingale.

The method comes from divisor theory. Treewidth and gonality of glued grid graphs gives exact gonality for two families. For stacked prisms $Y_{m,n}$ (a cycle of length $m$ stacked $n$ layers deep), gonality is $\min\{m,2n\}$ when $m\neq 2n$. For toroidal grids $T_{m,n}$, it is $2\min\{m,n\}$ when $|m-n|\ge 2$. The same paper shows that treewidth coincides with gonality on these families, and that treewidth is always a lower bound on gonality. Gonality Sequences of Graphs extends the idea to the sequence $\text{gon}_r(G)$, the fewest chips that can cover any $r$ opponent-chosen vertices, and supplies a modified burning algorithm for checking it. No work in the file links any of these invariants to deliberation. The closest bridge, chip-firing dynamics such as The devil's staircase for chip-firing on random graphs and on graphons, studies abstract tokens rather than agent beliefs.

Core Claim

At a fixed number of agents and fixed degree, a Delphi-style agent panel whose communication graph has lower gonality keeps correct minority positions alive for more rounds and shows lower disagreement collapse. Gonality, or its computable lower bound treewidth, explains variance in these outcomes that edge count, algebraic connectivity (the spectral gap of the graph Laplacian), and diameter do not. The claim is falsifiable. It fails if, in a regression across graphs, gonality's coefficient vanishes once spectral gap is included.

Research Program

The panel protocol follows Delphi rather than free-form debate. In round one, each agent answers independently with a rationale, which matches the All-Agents Drafting idea from Kaesberg et al. In each later round, an agent sees only an anonymized summary from its graph neighbors: the distribution of their answers plus up to two neighbor rationales per distinct answer. An agent whose answer disagrees with its neighborhood majority must either revise or write an explicit justification. This mirrors the controlled-feedback-and-outlier rule used in Delphi practice, for example in An international, Delphi consensus study to identify priorities for methodological research in behavioral trials in health research. The panel runs four rounds, one beyond the two-to-three-round window Yao et al. identify, so that collapse has room to show up. The final answer is a plurality vote over the last round.

The graph suite uses 24 agents. The known-gonality families are the following. Stacked prisms $Y_{3,8}$, $Y_{4,6}$, $Y_{6,4}$, $Y_{8,3}$, and $Y_{12,2}$ have gonality 3, 4, 6, 6, and 4, with 45, 44, 42, 40, and 36 edges respectively. The 4-regular toroidal grids $T_{3,8}$ and $T_{4,6}$ both have 48 edges and gonality 6 and 8. The cycle $C_{24}$ (gonality 2) and the complete graph $K_{24}$ serve as the Li et al. anchors. For all product graphs, the spectral gap and diameter follow exactly from the factor graphs.

These structured families move gonality, spectral gap, and diameter largely together. $Y_{4,6}$ and $Y_{12,2}$, for instance, share gonality 4, spectral gap, and diameter. The families alone therefore cannot separate the invariants. To break that collinearity, we add about 40 random 4-regular graphs on 24 vertices. For each we compute exact treewidth, and we bound gonality using SageMath chip-firing search seeded by the burning algorithm, relying on treewidth ≤ gonality for the lower bound. We keep graphs spanning a range of treewidth at matched spectral gap, and matched spectral gap at differing treewidth. Where gonality cannot be pinned down exactly, we report results with treewidth and with the gonality upper bound separately.

The outcome measures are four. Dissent survival is the round at which the last correct agent flips, among questions where the correct answer starts as a minority. Disagreement Collapse Rate and Negative Agreement Rate follow Yao et al. We also record final accuracy and input tokens.

The exploratory anchor arm tests the gonality sequence. We designate $k$ "anchor" agents that receive an external correction signal: a stronger model's answer and rationale, following Li et al.'s strong-model-at-a-node setup. Anchors are placed in one of three ways: on the support of a minimum-degree rank-1 divisor, at the top-$k$ nodes by centrality, or at random. The divisor placement puts anchors on a narrow sweep front, and the question is whether that front carries correction to arbitrary agents better than hubs do. We treat this as exploratory because LLM persuasion is not literal chip-firing.

Baselines

The naive reference is independent majority voting. The same 24 agents answer once with no communication, and the plurality answer is taken. This is the strongest competitor on accuracy. On GSM8K with Qwen2.5-7B, Debate or Vote reports 94.0% for majority voting against 90.3% for the best debate variant. We run it with the same model, sampling temperature, and question subsets. Since the martingale result predicts that no topology beats it in expectation, "beating" it on accuracy is not the bar. The bar is matching its accuracy within about one point while preserving dissent that can be flagged, for example by abstaining on questions where a surviving minority persists, as consensus protocols do on SQuAD 2.0 in Kaesberg et al.

The strongest structural competitor is the sparse ring of Improving Multi-Agent Debate with Sparse Communication Topology, run as $C_{24}$ inside our Delphi protocol and also with its original debate prompting. That paper reports a 2% gain over fully connected debate on MATH but no absolute accuracy figure in the file. A gonality-selected graph beats it if it achieves lower collapse at equal or higher accuracy, since the ring is itself the minimum-gonality case at degree 2.

The centralized-judge baseline follows Yao et al., who report collapse falling from about 81% to 41% on CommonsenseQA. We implement one judge reading all 24 round-one answers under the same token budget. A decentralized graph beats it if it reaches comparable collapse without the single point of failure.

The key ablation is density matching. Every structured graph is compared with random graphs of identical degree sequence and edge count. Any effect that survives this comparison is attributable to structure rather than sparsity.

All conditions use Qwen2.5-7B, for comparability with the GSM8K figures above, and identical prompts apart from the neighbor list. Each benchmark uses 500 sampled questions, with five seeds per graph.

Why not just…?

Why not just use majority voting and skip deliberation? On accuracy this is a real threat, and the martingale proof in Debate or Vote predicts it. Voting, however, discards the structured minority signal that Delphi exists to surface, and it fails at detecting unanswerable questions according to Voting or Consensus?. If our best graph offers no calibration or abstention benefit over voting at equal cost, the practical case for networked panels collapses. We will report that outcome as such.

Why not just add a central judge? Yao et al. show a judge halves collapse. A judge is also a single agent whose own sycophancy or error becomes the panel's. The experiment compares judge-led collapse with the best decentralized graph directly. If the judge wins outright, gonality is at best a tool for choosing among decentralized designs, which is a narrower claim.

Why not just use density, or spectral gap, which is the standard measure of how fast things mix on a graph? This is the alternative most likely to absorb the effect, because on product graphs gonality and spectral gap move nearly together. The random-regular arm exists to settle it. If spectral gap alone predicts dissent survival as well as gonality does, the note's conclusion is that the divisor invariant adds nothing. That would still be a useful negative result, because it tells designers which single number to compute.

Why not just cap rounds at two or three? Yao et al.'s round limit does reduce collapse, but it truncates both collapse and correction. We run all graphs at rounds two, three, and four and ask whether the gonality ordering persists at the recommended cap. If it vanishes at round two, topology is irrelevant under good round hygiene.

Proposed Experiments

We use CommonsenseQA for collapse, where Yao et al.'s 81% and 41% figures give reference points, and MMLU-Pro for knowledge filtering. MuSR serves as the reasoning case where consensus fails, with Kaesberg et al. reporting 28.4% for consensus versus 54.8% for voting. GSM8K checks against the voting ceiling.

The first decisive check is a mixed-effects regression of dissent-survival round on gonality or treewidth, spectral gap, diameter, and edge count, pooled over all 24-node graphs with a question-level random effect. The hypothesis survives only if the gonality coefficient is negative and its confidence interval excludes zero with spectral gap included.

The second check compares collapse rates between $T_{3,8}$ and $T_{4,6}$, and between each and its degree-matched random controls. This asks whether the within-family ordering matches the gonality ordering.

The third check tests whether the effect reverses on MuSR. There, slower consensus might help because voting-like exploration wins. On MMLU-Pro, filtering might instead favor faster agreement. A sign flip across task types would tie gonality to Kaesberg et al.'s filter-versus-explorer dichotomy.

The exploratory anchor test asks whether divisor-support anchor placement lifts accuracy above centrality placement at equal $k$. The only reference number here is Li et al.'s "up to 3%" gain from centrality placement. A null result on this arm would not affect the main claim.