Contribution
The paper turns Betz’s debate dynamics into a two-player game. Debaters care about ending with true beliefs and also about not changing their minds. It shows that this mixed motive makes debate a cooperative dilemma in parts of parameter space: there are profiles that are socially optimal for accuracy but are not equilibria. It also shows that stronger truth motivation does not always raise equilibrium accuracy.
Key insight
In deductive debate, being refuted is costly to the refuted party and useful to both. So each debater would rather the other did the belief-disrupting work. Arguing boldly means choosing premises one accepts for conclusions one rejects, so that every argument eliminates one’s own position. Bold argument is a contribution to a public good, and free-riding on it is individually rational.
Assumptions
- Two debaters and n logically independent atomic propositions, with n = 3 for exact results. Positions are vectors in {0,1}^n, and one random position is the truth.
- Arguments have two premises and one conclusion. An argument eliminates the positions satisfying the premises and the negation of the conclusion. Logic is classical and monotonic. In the baseline, arguments are valid, meaning they never eliminate the truth.
- Minimal revision. A refuted debater moves to the nearest tenable position by Hamming distance, with ties broken at random.
- Strategies are fixed rules on premises × conclusion, each drawn from {S+, S−, O+, O−} (self or other, accept or reject), giving 16 strategies plus Exit. A debater picks uniformly among feasible argument classes and passes if none is feasible. A rule that depends only on current positions is a strong restriction: there is no memory and no planning.
- Turns are random. After each argument the debate continues with probability δ. The SI says “ends with probability δ” at one point, but Eq. 18 and the main text treat δ as the continuation probability.
- Payoff u = w·v − d, where v ∈ [0, n] is the number of true beliefs at the end and d is the total Hamming distance moved. Exit gives w·n/2 to both.
- Error extension. With weight α, invalid arguments (which eliminate the truth) are chosen too.
- Equilibrium selection. Nash equilibria after iterated removal of weakly dominated strategies. Among those left, the one most truth-conducive (an “epistemic planner”) is chosen. A utilitarian criterion is checked in SI D.3.
Key results
- Strategy classes and accuracy (Fig 2, n = 3).
- Profiles with at least one bold strategy (S+S− or O+O−) maximise collective accuracy.
- Profiles where both are conservative or both refusenik do much worse.
- Profiles with a compromising strategy and none bold vary widely.
- Error (Fig 3). At α = 0, (S+S−, S+S−) is truth-maximising for all δ. At α > 0, less argumentative profiles do better in long debates, because bold argument keeps eliminating until something goes, possibly the truth.
- Cooperative dilemmas (§2.4.2, after Peña and Nöldeke). C and D form a cooperate–defect pair iff E(C,C) > E(D,D), E(C,C) ≥ E(C,D) and E(D,C) ≥ E(D,D), with at least one inequality strict. A game is a cooperative dilemma if C is in a social optimum, D is played in a suboptimal equilibrium, and (C, D) is such a pair.
- Table 1 (δ = 0.8).
- Truth weight 2.5: a prisoner’s dilemma, S−O− against S+S+.
- Truth weight 3.5: a stag hunt (S−O− against S+O+) and a snowdrift (bold against S+O+).
- Truth weight 4.5: a snowdrift.
- Truth weight 9.5: no dilemma.
- (Exit, Exit) appears as a pure equilibrium in every row. That is structural: if the other debater exits, no deviation changes your payoff.
- Harmony (Fig 4). The rank correlation of payoffs is always positive, so debate is never zero-sum. Cooperation is easiest in long debates and a narrow band of truth weights.
- Regimes (Fig 5, α = 0).
- Low w: refusenik, “almost invariably” (Exit, Exit), an “agreement to disagree”.
- High w and short debates: all-bold. S+S− weakly dominates, and (S+S−, S+S−) is the unique equilibrium after elimination and truth-optimal.
- Slightly lower w or longer debates: the best reply to S+S− is a compromise such as S+O+, giving an asymmetric bold–compromise equilibrium, a snowdrift with “a modest loss of collective accuracy”.
- A borderline band: conservative strategies appear.
- Error and disengagement (§2.6.4). At α = 0.1 and α = 0.3, refusenik strategies, mostly Exit, take over much of the space.
- Non-monotonicity (§2.6.5, Fig 5 lower panels, Fig 9). At α = 0 and α = 0.1, collective accuracy dips as w rises within the borderline band. Here there is no pure equilibrium, and in the mixed equilibrium each player randomises to make the other indifferent. So as a player cares more about truth, “their opponent must therefore more frequently use strategies to frustrate that outcome”. The rendered figure shows the dip as local and small, most visible for δ = 0.9. Fig 9 reports that it survives utilitarian equilibrium selection.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Bold argument is truth-optimal for error-free debaters | moderate | exact Markov computation, n = 3, Figs 2–3 |
| C2 | Mixed truth-and-obstinacy motives make debate a cooperative dilemma (PD, stag hunt or snowdrift) at low and moderate truth weights | moderate | Table 1 at δ = 0.8; Fig 5 across δ |
| C3 | The typical dilemma is snowdrift: one debater argues boldly and the other free-rides with a compromise strategy | moderate | §2.6.3, Fig 5 |
| C4 | Raising the truth weight can lower equilibrium collective accuracy | weak | a narrow mixed-equilibrium band, a small dip, Fig 5 and Fig 9 |
| C5 | Higher error rates drive debaters to exit | moderate | Fig 5, α = 0.1 and 0.3 |
| C6 | Obstinacy is a rational response to deductive refutation, since revision may move one further from the truth | weak | informal argument, §3 |
| C7 | The framework could explain confirmation bias, mistaken expert consensus and polarisation | weak | stated as a possible generalisation, §3, not modelled |
Method
For n = 3, the debate is a Markov chain on states (T, p′, p″): 2,816 states, and over 4,000 with error. Expected end-states are computed as (1 − δ)·y(0)(I − δM)^−1, and expected revisions as d = δ Σ_{i,j} z_i M_ij Q_ij with z = y(0)(I − δM)^−1. For n ≥ 4, the chain (about 2.5 million states at n = 4) is replaced by simulation, with 200,000 or more debates per profile (Fig 8). Nash equilibria come from the 17 × 17 payoff table.
Concepts
- Tenable position. A truth assignment not yet eliminated by any argument.
- Bold, conservative, compromising, refusenik. Bold strategies force a move unless the target already holds the truth. Conservative ones never force at least one debater to move. Refusenik ones never force anyone to move. Compromising ones are the remaining dyadic strategies.
- Collective accuracy. The mean of the two debaters’ fractions of true beliefs, which is what “a hypothetical impartial observer” cares about.
- Consensus. The share of propositions on which the two agree. It is tracked separately, since consensus need not be on the truth (SI D.1).
- Obstinacy. The aversion to revision, priced at one unit per proposition flipped.
Connections
The model builds on Betz’s Debate Dynamics (2013). It differs in having random turn order, a continuous end probability and no fallback to random arguments (SI B.1). It takes its definition of cooperation from Peña and Nöldeke (2023). It places itself beside Mayo-Wilson, Zollman and Danks’s independence thesis, List and Pettit’s and Dunn’s epistemic free-riding, and Zollman’s network models (LIT-735, which the record holds), and it cites Zollman’s transient diversity for “why expert communities sometimes maintain mistaken consensus”.
Bearing on the record
- QUESTION-047. The paper supplies a named mechanism by which
individually rational inquirers produce a collectively worse epistemic
state: free-riding on the cost of being refuted. That part is filed as
THEORY-290. Three cautions about using it for the question:
- Two people, not a community. No structural variable of the kind QUESTION-047 lists (communication density, influence weighting) is present. The variables are debate length, truth weight and error rate.
- No settled group belief. The outcome measured is collective accuracy, the mean of two individual accuracies. The paper does not show the pair settling on a shared belief that neither reasoned to. Each debater ends where minimal revision under the other’s arguments put it, which is in a sense where neither reasoned to independently, but the paper does not frame or measure it that way.
- “Sound” needs a qualifier. The debaters are logically flawless at α = 0, but they are not pure truth-seekers: obstinacy is in their utility. QUESTION-047’s “individually sound inquirers” fits only if a preference against revising beliefs counts as soundness. The paper argues it can (§3), informally.
- The LIT’s seeded claim, checked. LIT-966’s body says this “is a named mechanism by which a group settles on what no member reasoned to”. The first half is fair. The second is the record’s reading, not the paper’s: the paper shows equilibrium strategy profiles with lower collective accuracy than the social optimum, not a group settling on a belief. It also says “strategies that are optimal for each debater give a collectively worse belief outcome”, which holds only in the dilemma regions. At high truth weight with short debates the equilibrium is the truth-optimal one (Table 1, “No dilemmas”; §2.6.2).
- The LIT summary’s “raising the debaters’ motivation to seek truth can sometimes give equilibria with worse collective outcomes”. Accurate, and “sometimes” is apt: the effect is confined to a narrow band of mixed equilibria and is small.
- CLAIM-192. The debaters’ strategies are fixed rules over current positions, so the model is no direct test of a multi-session argument. It does suggest one design for one: whether participants who bear the cost of revision argue less boldly.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- n = 3 for the exact results. Larger n is shown only for equilibrium types (Fig 8), from simulation.
- The strategy space is coarse and memoryless. The “ideal cooperative debater” of §2.4.1 lies outside it, as the authors say.
- Propositions are atomic and independent, and logic is classical, which the authors acknowledge.
- Equilibrium selection by a truth-favouring planner is an idealisation. (Exit, Exit) is always an equilibrium, so which equilibrium the results describe depends on selection.
- The broader explanatory claims of the Discussion are not modelled.
Open questions
- Does the snowdrift structure survive in groups larger than two, where free-riding on others’ refutations becomes a many-player public-goods problem?
- Would strategies with memory or adaptive targeting, nearer the ideal cooperative debater, remove the dilemma?
- Is the non-monotonicity robust beyond n = 3 and the borderline band?