When several language models argue toward one answer, the argument often does not settle. Agents flip positions, return to answers they had dropped, or drift away from the question as more rounds are added. Kaesberg et al. report that extra rounds can lower accuracy (Voting or Consensus?), and Pitre et al. describe "wild, non-progressive oscillation" (A Diagnostic Study). Pricing theory has studied a similar failure for decades. When sellers compete over bundles of items for a buyer whose value for one item depends on which other items are in the bundle, there may be no stable set of deterministic choices. In game-theory terms, a pure strategy Nash equilibrium (a profile of choices where no one gains by changing alone) can fail to exist, and the market cycles instead of settling. Lev et al. prove this for vendors that sell several items at once. They also prove the fix: when items fall into categories and the buyer wants at most one item per category, a stable and efficient outcome is guaranteed (The Pricing War Continues).
This note proposes that debate with an LLM judge has the same structure. The judge acts as the buyer. The claims each agent puts forward act as the items. An agent that defends a whole solution is a multi-item vendor, and the judge's value for any one claim depends on which other claims are present. This predicts two things. Whole-answer debate under a judge will cycle, and the cycling will be worst when candidates are close in quality. Re-organizing the task so that the solution is split into slots, with each slot holding exactly one claim and agents competing slot by slot, should remove the cycling at no extra cost in tokens. The split into slots is task decomposition, used here as a way to make the debate stable rather than as a way to divide labor. The prediction is falsifiable, and it comes with a controlled test of the mechanism.
Core Claim
The claim is that oscillation in judged debate is caused by bundled ownership of interacting claims, not by the number of rounds or by judge weakness alone. We model one revision round as a best-response step. Each agent $i$ holds a set of sub-claims $S_i$, and the judge assigns a value $v(S)$ to the assembled answer $S$. Each agent revises to maximize the judge-approved share of its own claims. Suppose $v$ is non-additive, meaning the value of a claim depends on which other claims are present, and each agent controls several interacting claims. Then the Lev et al. argument carries over: a move on one claim changes the marginal value of claims the other agents control, and best responses can cycle. Now suppose the task is split into category-divided slots. The final answer takes one claim per slot (for a travel plan: one flight, one hotel per night, one restaurant per meal), and agents compete within a slot rather than across bundles. The CDSP result then says a stable profile exists and matches the best achievable outcome.
A second, independent contributor is the judge's own inconsistency. Pairwise LLM judges show rock-paper-scissors cycles, and these peak when candidates are close in quality, which the authors call the "Similarity Trap" (Investigating Non-Transitivity in LLM-as-a-Judge). A cyclic judge makes stable outcomes impossible even within one slot. The proposal therefore separates the two causes, bundling across claims and cycles within a slot, and treats them separately.
Relation to Prior Work
Three papers motivate the gap, because they document instability without a cause. Kaesberg et al. compare seven ways of turning a debate into a decision. They find that voting beats consensus on reasoning tasks (on MuSR, 54.8% versus 28.4%) and that accuracy drifts as rounds increase, but they leave the drift unexplained. Pitre et al. measure process health, including a post-hoc check of whether the final state is a Nash equilibrium. They do not predict when stable outcomes must fail. Qiao et al. attribute debate stagnation to rising within-agent reasoning noise, which they call the "Aleatoric Barrier" (Epistemic Gain, Aleatoric Cost). That explanation lives at the level of the individual agent. Ours lives at the level of the game structure, and the two can both hold.
The closest existing work is ECON (From Debate to Equilibrium). ECON frames multi-agent reasoning as a Bayesian game, learns temperature and repetition penalty so the agents reach an assumed equilibrium, and reports a sublinear regret bound. ECON assumes a stable outcome exists and tunes the agents toward it. We ask when a stable outcome cannot exist, and we change the task structure rather than the agents' generation settings.
Two papers supply methods we reuse. MACM (MACM) already splits math solving into a Thinker that proposes conditions, a Judge that accepts or rejects them into a shared list, and an Executor that computes. That shared list is almost a slot structure. However, MACM has a single proposer and treats its Judge as an oracle, so it has neither competition nor an analysis of instability. The non-transitivity paper supplies two cycle metrics we reuse: the percentage of non-transitive triplets (PNT) and a divergence from a transitive Bradley-Terry fit (SNTD).
Two papers constrain the design. AgentEval (AgentEval) shows that 63% of step-level failures propagate from upstream steps. It also states that its dependency-graph method degrades on looping, multi-agent traces, which is exactly the regime debates create. Slot structure gives us per-slot histories that can be evaluated even when the debate loops. Finally, the Buy Box analysis (Price Stability and Improved Buyer Utility) shows that making one seller prominent stabilizes prices. It suggests a second lever: the current holder of a slot keeps it unless a challenger beats it by a set margin.
Research Program
The work has three parts.
The first part is a formal model. We write judged debate as a game between agents that post claims, with a judge valuation $v$, and restate the Lev et al. non-existence construction and the CDSP existence result in that language. We add one extension the pricing literature does not need: a judge whose pairwise preferences within a slot are themselves cyclic. In that case we bound how often best responses revisit a previous state as a function of the measured PNT.
The second part is the system. We build slot debate, a protocol with four steps. A decomposer, reused from MACM-style condition mining or from the natural fields of TravelPlanner, defines the slots. Every agent drafts one candidate per slot independently, following the All-Agents Drafting practice of Kaesberg et al. The judge compares candidates within a slot, conditioning on the claims already committed upstream, as AgentEval does. Revision happens slot by slot. An optional incumbency rule means a new candidate displaces the current holder only if the judge prefers it by margin $\delta$.
The third part is measurement. We define a revisit rate: the fraction of debates in which some agent's claim set, or the assembled answer, returns to an earlier state after changing. We track it next to accuracy at every round.
Baselines
The naive reference points are a single agent with chain-of-thought and plain majority voting over independent drafts, with no debate at all. These tell us whether any interaction is worth its tokens.
The flat-debate baseline is standard whole-answer debate with a judge. Each agent defends a complete solution for up to five rounds, and we report both the voting and the consensus decision rules from Kaesberg et al. On MuSR they report 54.8% for voting, the stronger of the two.
The strongest existing methods differ by benchmark. On TravelPlanner, ECON reports a 15.2% pass rate against 7.1% for standard debate. On MATH Level 5, MACM reports 76.73% with GPT-4 Turbo, up from 54.68% with plain prompting.
Those numbers come from different backbone models. We will therefore rerun every method on the same open backbones, Llama 3 70B and one Qwen model, with the same judge model and a matched total token budget per problem. Each method gets the same prompt-tuning budget: a fixed number of prompt variants, tuned on a held-out split.
Beating the baselines means three things together. Slot debate must beat ECON's pass rate and MACM's accuracy under these matched conditions. It must show a lower revisit rate than flat debate. And its accuracy must not decline as rounds increase.
The key ablation keeps everything in slot debate except ownership. Agents still see the slot structure, but each defends a full bundle of slots, and the judge scores whole bundles. This isolates bundled ownership as the cause, separate from the benefit of simply having a decomposition. A second ablation varies the incumbency margin $\delta$ to test the Buy Box lever. By analogy to the Search Cost Paradox, we expect higher margins to stabilize the debate while locking in more early errors.
Why not just…?
Why not just stop after one or two rounds and vote? This is the most serious threat. Kaesberg et al. show that voting over independent drafts is strong on reasoning, and drift is a problem of many rounds. If early stopping with voting matches slot debate at equal tokens, our stabilization buys nothing in practice. Accuracy-versus-rounds curves at matched budgets settle this. Our claim survives only if slot debate keeps improving after round two, where voting has already plateaued.
Why not just use a stronger judge? Stronger judges do show fewer preference cycles (Investigating Non-Transitivity). Our theory, however, predicts cycling from bundling even when the judge is perfectly transitive, because a non-additive valuation is enough. We test this directly with a synthetic control, described below. We also report the revisit rate against the judge's PNT, which lets us estimate how much of the cycling a better judge could remove.
Why not just aggregate with a tournament and Bradley-Terry ranking, as SWIM does? Tournament aggregation fixes how final answers are ranked, not what agents do during revision. It also collapses quality onto a single scale, a limitation its own authors acknowledge. Within each slot we will use tournament aggregation as a component, so the comparison is between flat debate and slot debate with the same aggregation.
Why not just use MACM? MACM is the decomposition half of our proposal with a single proposer. If its lone Thinker matches multi-proposer slot debate, the conclusion is that competition within a slot adds nothing. That would be a useful negative result, and running MACM as a baseline measures it directly.
Proposed Experiments
The benchmarks are TravelPlanner, MATH Level 5, and MuSR. TravelPlanner's fields form categories with one choice each, a near-literal case of the CDSP structure. MATH Level 5 is split into slots through mined conditions. MuSR serves as a stress test, since its slots (suspect, motive, opportunity) interact more strongly.
The decisive check is a synthetic control that tests the mechanism rather than the benchmark. We build arithmetic and scheduling tasks where the scoring function is programmatic and can be switched between two modes. In the additive mode, each claim is scored independently. In the complementary mode, pairs of claims interact, for example a hotel only counts if it matches the flight date. We then use a scripted, perfectly transitive judge. The theory makes four predictions. Flat debate should cycle only in the complementary mode. Slot debate should converge in both modes. When we swap in the LLM judge, any remaining within-slot revisits should concentrate on near-tie candidates, the Similarity Trap. And revisit rate should predict the final error rate.
If flat debate cycles just as often in the additive mode, the bundling explanation is wrong. If slot debate does not beat matched early-stopped voting on the real benchmarks, the practical claim fails, even if the mechanism holds.
References
- Voting or Consensus? Decision-Making in Multi-Agent Debate
- A Diagnostic Study of Multi-Agent LLMs for Real-World Debates
- The Pricing War Continues: On Competitive Multi-Item Pricing
- Investigating Non-Transitivity in LLM-as-a-Judge
- Epistemic Gain, Aleatoric Cost: Uncertainty Decomposition in Multi-Agent Debate for Math Reasoning
- From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium
- MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems
- AgentEval: DAG-Structured Step-Level Evaluation for Agentic Workflows with Error Propagation Tracking
- Price Stability and Improved Buyer Utility with Presentation Design: A Theoretical Study of the Amazon Buy Box