Successful Misunderstandings

Definition

A successful misunderstanding is a multi-agent communication episode where agents receive positive task reward even though the communicated signal is not interpreted in the same way by the interacting agents.

Why It Matters

This failure mode breaks a common evaluation shortcut: using coordination success as evidence of shared understanding. It is dangerous because the system can look solved until a new partner, changed population structure, or diagnostic interaction reveals that the learned semantics were not shared.

Formalism / Key Objects

• In Kondylidis2025 - Successful Misunderstandings: Learning to Coordinate Without Being Understood, the successful misunderstanding ratio counts episodes where reward is positive while the Sender’s intent is not met.
• Sender intent is met when the Sender, if placed in the Receiver role for the emitted signal, would choose the same action the Receiver actually chose.
• Semantic alignment is measured by the fraction of the population sharing the dominant interpretation for each signal-role pair:

$$\frac{|\text{interpretation majority}|-1}{|\text{population}|-1}.$$

• A robustness test introduces new interaction partners or lifts pairing restrictions after convergence to see whether the protocol generalizes beyond the pairings that produced reward.

Connections

• Emergent Communication is the broader mechanism; successful misunderstandings are one brittle equilibrium of that mechanism.
• Explainable AI is relevant because performance alone does not explain what an agent’s internal signal use means.
• Multi-Agent Non-Stationarity appears when population membership or interaction graph changes after a protocol has converged.
• Safe Multi-Agent Reinforcement Learning should not treat high coordination reward as sufficient evidence that agents share task semantics.

Common Confusions

• A successful misunderstanding is not ordinary random noise; it can be a stable convention within a restricted pairing.
• It is not detected by reward alone.
• Adding more agents is not sufficient unless the interaction graph lets all relevant pairs calibrate.

Key Sources

• Kondylidis2025 - Successful Misunderstandings: Learning to Coordinate Without Being Understood