Successful Misunderstandings: Learning to Coordinate Without Being Understood

Summary

This partial ingest is based on the extracted arXiv PDF text. Kondylidis et al. study Emergent Communication in populations of agents that learn a signaling system while lacking shared observations of how signals are used. The paper’s central warning is that coordination success is not enough evidence for shared interpretation: agents can learn Successful Misunderstandings, where reward is high because paired conventions compensate for each other, but the semantics fail when new interaction partners are introduced.

Key Claims

• Alignment is not always a by-product of coordination. In two-agent populations, agents can achieve optimal reward while using opposite interpretations of the same signals.
• Successful misunderstandings are hard to detect from reward alone because they do not immediately cause coordination failure.
• Misaligned interpretations become brittle under partner changes. Populations that coordinated through successful misunderstandings often fail the robustness phase when a new agent is added or previously separated groups are allowed to interact.
• The paper identifies two minimum conditions for shared interpretations in its setting: at least three agents, and unrestricted interaction among all agents during convergence.
• A three-agent “misunderstanding detection” group can function as a calibration device: once that group has robust shared interpretations, other agents can be checked against one member.

Methods / Formalism

• The experiments use Lewis-style signaling games. A Sender observes a state and emits a signal; a Receiver chooses an action; both receive reward from the hidden state-action reward function.
• Agents do not observe each other’s state/action use directly, only the communication signal and the reward outcome.
• Sender and Receiver roles maintain value tables such as V(c,s) and V(c,a) for signal-state and signal-action associations; softmax policies with decaying exploration choose signals and actions.
• The paper separates task performance from semantic alignment using metrics for reward, vocabulary size, alignment, Sender intent met, and successful misunderstanding ratio.
• A successful misunderstanding occurs when reward is positive even though the Receiver’s action does not match the Sender’s own interpretation of the communicated signal.

Evidence / Experiments

• In the two-agent convergence phase, reward reaches 1.00, while alignment and intent-met metrics remain around 0.5; roughly half the runs converge to opposite signal interpretations that still coordinate successfully.
• When a third agent is introduced after two-agent convergence, the population does not reliably recover optimal coordination; reported final reward is about 0.80, with persistent successful misunderstandings.
• In the three-agent unrestricted experiment, populations converge to nearly optimal reward and near-shared interpretations, and they recover after adding a fourth agent.
• In the three-agent restricted experiment, where some pairs cannot interact during convergence, the population again supports successful misunderstandings and fails the later unrestricted robustness test.
• Additional asymmetric and 3x3 reward-function experiments support the same qualitative conditions.

Connections

• Extends the Multi-Agent Non-Stationarity and learned-communication thread by showing that high coordination performance can mask incompatible internal signal meanings.
• Supports Explainable AI because observable success is not a faithful explanation of what agents’ representations mean.
• Connects to Strategic Reasoning by treating semantic compatibility as a population-level property, not merely an individual learner property.
• Complements Agrawal2026 - The Multi-Agent Off-Switch Game: both papers show a non-compositionality pattern where individually or pairwise successful behavior fails under changed collective structure.
• Gives a concrete warning for Safe Multi-Agent Reinforcement Learning benchmarks that rely only on task reward or coordination rate.

Open Questions

• Do the same successful-misunderstanding dynamics arise with humans or LLM agents in hybrid populations?
• How large must the fully interacting calibration group be in richer state/action spaces or asymmetric populations?
• Which diagnostics can detect semantic misalignment before new partners expose it?
• Can learned communication protocols be regularized toward shared interpretation without destroying useful task specialization?

Citation

Kondylidis, Nikolaos, Anil Yaman, Frank van Harmelen, Erman Acar, and Annette ten Teije. 2025. “Successful Misunderstandings: Learning to Coordinate Without Being Understood.” arXiv:2509.24660.