In a Delphi study, each expert answers alone, sees an anonymous summary of what the others said, and revises over several rounds. Multi-agent debate among language models is an automated version of this. It has one well-documented puzzle. Iterating toward agreement helps on factual questions but can wreck multi-step reasoning. Voting or Consensus? Decision-Making in Multi-Agent Debate reports MMLU accuracy of 54.0% under consensus versus 51.7% under voting. On MuSR the order flips sharply: 54.8% for voting against 28.4% for consensus. Nobody can currently say, before running the panel, which regime a question is in.
We propose to fit a small, interpretable model of the revision step itself. The model describes how each agent's next answer distribution depends on its own current answer, on the peer summary, and on the product of the two, meaning how much an agent's own leaning amplifies or resists what the peers say. We then use the fitted coefficients to decide, question by question, whether to keep iterating or fall back to the independent first-round vote.
The idea borrows one concrete lesson from multilinear operator theory. Maps that are linear in each of several inputs have behavior that cannot be read off from their linear parts. Their response to constant, uninformative inputs tells you whether repeated application settles or drifts. We treat that lesson as a source of measurable probes, not as a theorem we transfer.
Core Claim
The claim is that the "drift" and biased lock-in seen in LLM debate are mostly carried by a multiplicative interaction term in the revision map. In that regime, agents that already lean toward the plurality move toward it disproportionately. Because the term is measurable from a few rounds of sampled answers, it predicts when consensus will filter errors and when it will amplify them.
Concretely, let $p_i^t$ be agent $i$'s probability vector over the $K$ answer options at round $t$. We estimate it by sampling the agent several times. Let $m_i^t$ be the mean of the other agents' vectors, which plays the role of the Delphi feedback summary. We fit
$$p_i^{t+1} \approx \mathrm{normalize}\big(b + A\,p_i^t + B\,m_i^t + C\,(p_i^t \odot m_i^t)\big),$$
where $b$ is a constant offset, $A$ is self-persistence, $B$ is linear peer pull, $C$ is the bilinear agreement-amplification term, and $\odot$ is elementwise multiplication. The functional form is the degree-2 truncation of the multiplicative interaction units in Multilinear Operator Networks. That paper shows networks built from products of linear maps, $(AX)\ast(BDX)+AX$, can model rich interactions with no elementwise activation. It also recovers exact Lotka–Volterra coefficients, which is the property we need: identifiable, readable coefficients rather than a black box.
If $C$ is near zero, the map is a classical linear averaging rule and iteration contracts toward the weighted mean. Consensus then filters errors whenever first-round answers are better than chance. If $C$ dominates, the mean-field picture of Emergence of Biased Consensus in Multi-Agent LLM Debates applies. That paper models debate as interacting spins and finds lock-in to biased consensus within one or two rounds at low sampling temperature. Its coupling, however, is a single isotropic scalar. Our model splits that coupling into a linear part and a multiplicative part and estimates both from real transcripts.
Research Program
The program has three parts. First, we build a data-collection harness that runs anonymous, Delphi-style panels. Each panel uses independent first drafts, as in the All-Agents Drafting procedure from Kaesberg et al., followed by up to five feedback rounds. At every round, every agent's answer distribution is estimated by repeated sampling. This produces trajectories, not just final answers.
Second, we fit the revision map per task family and per question. The per-question fit is low-dimensional enough to estimate from the first two rounds alone. We then extract three diagnostics from it.
The first diagnostic is the size of $C$ relative to $B$. The second is the local gain of the map at the current plurality, defined as the largest eigenvalue of its Jacobian. A gain above one means the plurality is self-reinforcing rather than merely attracting.
The third is a constant-input probe. Here the harmonic-analysis connection is explicit and deliberately modest. A characterization of compactness via bilinear $T1$ theorem shows that whether a bilinear singular integral is compact is decided largely by how the operator and its adjoints act on constant functions ($T(1,1)\in\mathrm{CMO}$), not by its size alone. Our operational analogue is to feed each agent a uniform, uninformative peer summary and measure its systematic movement. Any movement isolates the offset-driven drift that real feedback hides. This is an analogy that motivates the probe; we do not claim the theorem applies to LLM panels.
Third, we build a stopping-and-fallback rule. On each question, after round two, the panel continues iterating only if the estimated gain is below a threshold tuned on held-out data. Otherwise it returns the round-one vote. This directly targets the knowledge-versus-reasoning split.
Two other lines of work constrain the design. Thinking in a Crowd: How Auxiliary Information Shapes LLM Reasoning shows that misleading context drops Qwen3-8B with thinking enabled from 96.22% to 77.03% on SciAux, below its thinking-off accuracy of 74.87% plus the thinking gain. So peer feedback can be actively harmful, and the model must allow negative effective trust. WISE: Weighted Iterative Society-of-Experts for Multimodal Multi-Agent Debate with Probabilistic Consensus estimates agent reliability with a Dawid–Skene EM procedure, which treats agents as noisy annotators. That procedure models who is right on the final answers but not how agents influence one another between rounds. It is the strongest aggregation competitor rather than a component we reuse.
Proposed Experiments
We use the benchmarks from the debate literature in the file. The knowledge regime is MMLU, MMLU-Pro and GPQA. The reasoning regime is MuSR, with its 756 narrative instances, together with StrategyQA. SciAux serves as a controlled setting: we inject its misleading snippets as fake peer feedback, which gives ground truth about which feedback is bad.
Panels use open Llama 3 models at 8B and 70B scale, matching Kaesberg et al. as closely as the file allows. They also include Qwen3-8B with thinking on and off. Both homogeneous and temperature-mixed panels are run, so that Okawa et al.'s heterogeneity effect appears as a condition rather than a confound. Metrics are final accuracy, the fraction of questions where iterating changed a correct round-one plurality to incorrect, the area under the ROC curve (AUROC) of each diagnostic for predicting that harm, and total tokens consumed.
Baselines
The naive reference point is a single agent answering once. Next to it is the round-one independent vote: a plurality over blind drafts with no feedback at all. If our rule does not beat this vote, iteration has no value and the proposal reduces to "don't iterate."
The strongest existing protocol baselines are the voting and consensus decision rules of Voting or Consensus? Decision-Making in Multi-Agent Debate. Its reported figures are MuSR accuracy of 54.8% under voting and 28.4% under consensus, and MMLU accuracy of 54.0% under consensus and 51.7% under voting. We rerun both rules in our own harness with identical agents, rounds and token budgets, because our models may differ. Beating them means matching or exceeding the better of the two on each task family with a single rule, without knowing the family in advance.
The strongest aggregation baseline is Dawid–Skene reliability weighting as used in WISE, applied to all rounds' answers. WISE reports EvoChart-QA accuracy rising from 39.1% to 52.6% for GPT-4o. The file gives no reported result on Lacuna for Dawid–Skene aggregation on MMLU or MuSR, so we run it ourselves on identical transcripts. We also include Okawa et al.'s temperature-heterogeneous panel ($T\in[1.3,1.7]$) as a procedural baseline; the file has no accuracy figure for it on these benchmarks.
The key ablations hold data and compute fixed. One fits the linear averaging model with $C=0$. Another replaces our degree-2 map with a generic small multilayer perceptron, with the same number of parameters, trained on the same trajectory features. The third removes the constant-input probe. If $C$ carries the signal, the linear fit should lose most of the predictive AUROC.
Why not just…?
Why not just always use round-one voting, since iteration often hurts? This is a real threat. On MuSR it may be nearly optimal. But Kaesberg et al. find consensus beats voting by 2.8% on knowledge tasks, so a blanket fallback gives that gain up. The test is whether our per-question rule recovers voting-level MuSR accuracy and consensus-level MMLU accuracy. If it only matches round-one voting everywhere, we will report that iteration is not worth its tokens.
Why not just weight agents with Dawid–Skene as WISE does? Reliability weighting asks which agents are usually right. Lock-in, however, is a property of the dynamics: a reliable agent can still be pulled into a wrong plurality. The two can be compared directly on identical transcripts. Combining our stopping rule with Dawid–Skene weighting is a natural extension if both help.
Why not just train a black-box predictor of when debate fails? It may predict as well as our model. The MLP ablation measures exactly that. If it matches our AUROC, the contribution narrows to interpretability: a coefficient that says why a panel locks in and that transfers across models. We test transfer by fitting on Llama panels and evaluating on Qwen panels.
Decisive Checks
The project succeeds if three things hold on held-out questions. First, the bilinear-to-linear coupling ratio predicts harm from iteration with clearly better AUROC than the linear-only fit. Second, the gain-thresholded stopping rule beats both fixed protocols on the average across knowledge and reasoning families. Third, the constant-input drift probe correlates with lock-in in the bias tasks of Okawa et al. and with damage from injected misleading SciAux feedback.
The core claim fails if $C$ is negligible everywhere, meaning linear averaging fully explains revision. It also fails if the diagnostics cannot be estimated reliably within two rounds.
References
- Voting or Consensus? Decision-Making in Multi-Agent Debate
- Multilinear Operator Networks
- Emergence of Biased Consensus in Multi-Agent LLM Debates
- A characterization of compactness via bilinear $T1$ theorem
- Thinking in a Crowd: How Auxiliary Information Shapes LLM Reasoning
- WISE: Weighted Iterative Society-of-Experts for Multimodal Multi-Agent Debate with Probabilistic Consensus