Toby Handfield, Julián Garcia, Christian Hilbe and Shang Long Yeo (2025), arXiv:2504.06485.
Key takeaways
From the reading, NOTE-795.
- The model. Positions are truth assignments, and valid arguments eliminate them. A refuted debater moves minimally. Each debater picks one of 16 premise × conclusion rules (self or other, accept or reject), or Exit. The payoff is u = w·v − d.
- Bold argument is the public good. Strategies like S+S− (premises I accept, conclusions I reject) refute their user’s own position and are truth-maximising when no argument is invalid. Their cost falls on whoever is refuted.
- Free-riding. At moderate truth weight or in longer debates, the best reply to a bold partner is a compromise strategy that forces revision only where the two disagree. This gives an asymmetric snowdrift equilibrium with a modest loss of collective accuracy.
- Exit. (Exit, Exit) is an equilibrium everywhere and is selected at low truth weights, an “agreement to disagree”. With invalid arguments allowed (α = 0.1, 0.3), Exit spreads over much of parameter space.
- Non-monotonicity. Where only mixed equilibria exist, a debater who cares more about truth faces an opponent who must frustrate it more often, so collective accuracy dips as truth weight rises. The dip is small and local, and survives a utilitarian selection rule.
- Scope. Two players, n = 3, memoryless strategies, independent atomic propositions. The authors’ links to confirmation bias, polarisation and mistaken expert consensus are suggestions, not models.
Standing in the record
Filed on 2026-10-12 for QUESTION-047 and read on 2026-10-14 in NOTE-795. What it shows is held as THEORY-290: a mechanism, the private cost and shared benefit of being refuted, by which individually rational reasoners produce a worse collective epistemic state. It does not show a group settling on a belief no member reasoned to. The seeded body said it did, and that was the record’s reading, not the paper’s.