Following Each Voter: Trajectory-Based Aggregation for Multi-Round Deliberation

When several AI agents, or several human experts in a Delphi panel, debate a question over multiple rounds, the usual practice is to take a snapshot of the group at each round. A Delphi panel is a structured expert survey in which people answer anonymously, see the group's summary statistics and reasons, and revise. The snapshot might be the median, the spread, or the vote count. Atmospheric scientists call this the Eulerian view: you watch fixed points and never follow any individual. This note borrows the opposite, Lagrangian, view from fluid dynamics. You ride along with each participant and record what it was exposed to on the way. In cloud physics this change of view decides what can be explained. A cloud's state depends on how long its air parcel spent over sea ice or open water, and a snapshot of local conditions does not reveal that. The same may hold for opinions. Two agents who end on the same answer may have arrived by very different routes. One was moved by a new argument. The other simply slid toward the majority.

The proposal is to follow each participant's answer across rounds and split every change into two parts. One part is explained by the content of the arguments that participant read. The other is explained by pull toward the visible group position. The final verdict is then built from the argument-driven part of each path, and the conformity-driven part is thrown away. If this works, it addresses two documented failures of multi-round LLM debate. The first is accuracy loss as rounds increase ("drift" and sycophancy, meaning agents telling peers what they want to hear). The second is deliberation that spreads opinions apart instead of bringing them together. The same bookkeeping would also give Delphi facilitators a per-panelist audit of why each position moved.

Core Claim

The claim is that conformity and persuasion leave different signatures in individual trajectories, and that these signatures are measurable when the answer space is discrete. Let $p_i^t$ be agent $i$'s probability distribution over answer options at round $t$. Let $m^t$ be the peer distribution it was shown. Let $q_i^t$ be a counterfactual "tracer" state: what a frozen reader concludes from the same peer messages after vote labels, stated answers, and the group summary have been removed. The round-to-round displacement is decomposed as

$$p_i^{t+1} - p_i^t = \alpha_i^t\,(q_i^t - p_i^t) + \beta_i^t\,(m^t - p_i^t) + r_i^t,$$

where $\alpha_i^t$ measures movement toward what the arguments alone support, $\beta_i^t$ measures movement toward the crowd beyond that, and $r_i^t$ is a residual. The coefficients are fit by constrained least squares per agent and round. Accumulating $\sum_t \beta_i^t$ along a trajectory gives a conformity exposure for that agent, playing the same role as cumulative time over ice for an air parcel. Accumulating the $\alpha$ terms gives an evidence exposure. The "de-conformed" final state of each agent is its round-zero answer plus the argument-driven increments only. The verdict aggregates these de-conformed states.

The prediction is falsifiable. On tasks where debate is known to lose accuracy after several rounds, de-conformed aggregation should recover most of that loss while keeping the gains from genuine correction. Rounds where $\beta$ dominates $\alpha$ should predict errors better than any group-level spread statistic.

Why This Is Not Already Done

Several lines of prior work motivate the gap. Voting or Consensus? Decision-Making in Multi-Agent Debate compares seven decision rules for Llama 3 agent debates. It finds that consensus helps on knowledge tasks, that voting helps on reasoning tasks, and that extra rounds cause drift. Every one of its rules still aggregates round-level answers; none looks at why an individual agent moved. Can LLMs deliberate? Benchmarking Collective Reasoning for Democratic AI Applications reports that LLM deliberation increases opinion dispersion where humans decrease it, and that reasoning gains are only 29% of human gains. That paper also measures only group-level spread. The Human-AI Hybrid Delphi Model compresses Delphi into a single round, using a model as an evidence floor and a human facilitator for synthesis. That choice discards the multi-round path this proposal needs.

Two mathematical cousins constrain the design rather than supply it. An interacting particle consensus method for constrained global optimization deliberately drives particles toward the swarm centroid. For deliberation, that centroid pull is exactly the force we want to detect and remove. Predicting deliberative outcomes models actors as best-responding to a weighted sum of peer actions and learns who influences whom. It has no channel for argument content, so it cannot separate persuasion from pressure.

The physical inspiration is specific. Air mass history linked to the development of Arctic mixed-phase clouds and Investigating the development of clouds within marine cold-air outbreaks explain cloud transitions through integrated exposure along parcel paths. Cloud fraction response to aerosol driven by nighttime processes shows that earlier segments of a trajectory set the response observed later. These papers supply the method: integrate exposures along the path and attribute the outcome to them. The loose "opinions as particles" metaphor is not what is borrowed.

Research Program

The first component is the instrumented debate harness. It follows the setup of Kaesberg et al. (Llama 3 8B and 70B agents, multiple-choice answers) with one change: every agent reports a full distribution over options each round. The harness also logs the exact peer text and peer statistics each agent saw.

The second component is the tracer. A frozen copy of the base model reads each agent's incoming peer messages with answer letters, vote tallies, and agreement phrases ("I agree with Agent 2") removed. It outputs $q_i^t$. Removing these cues is imperfect, so the stripping is validated by a leakage check: the tracer should not be able to predict the peer majority from stripped text above chance on items where arguments and majority disagree.

The third component is the decomposition and the aggregation rules. The first rule is de-conformed voting. The second is a stopping rule that halts debate for an agent once its argument-driven displacement falls to near zero while its conformity term stays positive. This is the point where further rounds can only add pressure. It also gives a principled bridge between the "explore" and "filter" regimes that Kaesberg et al. found in conflict.

The fourth component moves to open-ended deliberation on the DelibSim setting from Flechtner et al. There, the answer space is the agents' stated position scales, and the question is whether conformity terms or argument terms account for the reversed convergence.

Baselines

The naive reference is first-round independent voting: each agent answers once without seeing peers, and the majority wins. This is the "do nothing" point that tells us whether debate adds anything at all. Given the reported drift, it is a serious competitor rather than a strawman. It uses the same agents and the same prompt budget for round zero.

The strongest existing method is the best protocol from Voting or Consensus? Decision-Making in Multi-Agent Debate, which is voting after debate combined with All-Agents Drafting. For MuSR, that paper reports voting at 54.8% accuracy versus 28.4% for consensus. It reports a 3.3% average accuracy gain from All-Agents Drafting, and on MMLU 54.0% for consensus versus 51.7% for voting. We will first reproduce these numbers with the same models and prompts, and then compare on our reproduction. Beating this baseline means a statistically reliable accuracy gain on MuSR and StrategyQA without losing more than noise on MMLU, MMLU-Pro, and GPQA. The tracer adds model calls, so the baselines receive matched extra compute. One form is an extra agent per panel with the same total token budget. The other is extra rounds. We report whichever helps them more.

Two ablations isolate the key ingredient. The first, an "Eulerian" ablation, computes the same conformity and argument quantities from group-average movement instead of per-agent paths. It then applies the same discounting uniformly. If this matches the full method, following individuals is unnecessary. The second, a "tracer-only" ablation, uses the stripped-argument reader's own final answer as the verdict. If this matches, the gain comes from the reader and the trajectory bookkeeping is not doing the work.

Why not just…?

Why not just stop after round one? If drift is the main effect, independent answers may be the best available. The first-round baseline tests this directly. The proposal only earns its keep if some later-round moves are argument-driven corrections worth keeping. We expect this mainly on MuSR's multi-step narratives, where one agent spotting a clue should persuade others. If first-round voting ties the method everywhere, we will report that debate adds nothing on these tasks, and the decomposition then serves only as a diagnostic.

Why not just hide peer answers and show arguments only, as a stricter Delphi anonymity? This is the most serious threat. It removes the conformity force at its source instead of subtracting it afterwards, and it costs nothing extra. We run it as a full protocol. Two outcomes are plausible. Hiding answers may fail because agents still infer the majority from the wording, which our leakage check measures. Or it may work, in which case the decomposition becomes an explanation of why it works rather than a better aggregator. Either result is informative for Delphi practice, which by default feeds back medians.

Why not just use a group-level dispersion threshold to decide when to stop? Classical Delphi stability checks and the spread measures in Flechtner et al. are exactly this. A shrinking spread cannot tell persuasion from capitulation, and in LLM groups the spread even grows. The Eulerian ablation settles whether per-agent paths add predictive power for errors beyond spread. It does this by comparing error-prediction area under the curve for $\beta$-dominated rounds against a spread-change signal.

Why not just prompt agents to follow deliberative norms? Flechtner et al. tested normative prompting and obtained ΔDRI = 0.029, only 29% of the human gain, with a diversity ceiling of 6.5 versus 18.8 for humans. Prompting changes how agents talk more than how they move. We include a normative-prompt condition crossed with our aggregator, so the two can be additive.

Proposed Experiments

The main experiment runs five-agent, five-round debates on MuSR, StrategyQA, MMLU, MMLU-Pro, and GPQA with Llama 3 8B and 70B. It compares first-round voting, the reproduced Kaesberg protocols, answer-hidden debate, the two ablations, and de-conformed aggregation, with at least three seeds. The primary metric is accuracy.

The decisive diagnostic check has two parts. First, among agents whose answer changed, conformity-dominated changes should turn right answers into wrong ones more often than argument-dominated changes do. Second, injecting one confidently wrong agent should produce a spike in $\beta$ for the others, not in $\alpha$. If neither holds, the decomposition is not measuring what we claim, and the aggregation results should not be trusted.

The deliberation experiment reuses DelibSim's topics. It tests whether removing conformity increments reverses the dispersion growth Flechtner et al. observed, and whether ΔDRI moves at all. No reported result was found on Lacuna for any trajectory-based method on DelibSim, so the comparison is against their prompted baselines. Human Delphi validation would require per-panelist round-by-round data with rationales. The file identifies no public dataset of that kind, so it is left as a follow-up with a prospective panel rather than claimed here.