Why can misaligned agents exploit cheap talk channels?

The abstract defines the term in a parenthesis: the agents' public "cheap-talk behavior" is "costless, non-binding communication that does not directly affect the agents' utilities." Three properties, each doing work. Costless: saying a thing takes nothing from the speaker. Non-binding: nothing commits the speaker to act on it. Utility-neutral: the words themselves change no one's payoff; only what the listeners do next can. In Werewolf that is the table talk, the accusations, defenses and claims about one's role, as opposed to the actions that change the game. The discussion draws the line itself when it calls voting a "private action" that misaligned agents adapt to their new objective.

Why the vault should hold the term. The paper reads two layers against each other: what an agent reasons and what it says in public. Calling the second layer cheap talk explains why the two can diverge without cost to the agent. Nothing in the channel penalizes a statement that does not match the reasoning behind it. Can we distinguish types of LLM falsehood by regeneration patterns? makes the neighboring point from the dialogue-agent side: role-played deception is content tailored to the interlocutor, and it does not require the system to believe anything. In Werewolf the role-play frame is the game's own rules, and deception is expected.

Background, not in the excerpt. In the standard game-theory treatment, cheap talk is informative only to the extent that the sender's and receiver's interests are aligned. If that carries over, an agent whose objective has shifted turns its allies' cheap-talk channel from informative to uninformative while the channel's form stays the same. That would be one way to read the discussion's claim that misalignment "exploits trust within nominally allied agents" (Why does misaligned trust between allies matter more than rule-breaking?). It is this vault's reading of the term, not the paper's argument.

A vault inference. Many agent-to-agent messages in multi-agent LLM systems have the same profile: free to send, nothing binds the sender. A defense that judges message content is then judging cheap talk. See Can one compromised agent corrupt an entire multi-agent network? for a case where the payload rides on ordinary messages, and Do internal agent hops in pipelines need security monitoring? for the channel inventory. Vending-Bench Arena is where such messages have been counted: How often do agents misalign through natural language communication? treats the email as the act, and How often do AI agents communicate dishonestly in commerce? labels 12.6 percent of 2,583 inter-agent emails as misaligned. Whether those emails are cheap talk in the three-property sense above is not settled by either excerpt. The speech-act note reads them as binding because counterparties act on them, but acting on words is the listener's move, and the sender is bound only if the message commits the sender, which neither excerpt says.

What the excerpt does not give. It does not say how public cheap talk was analyzed, meaning which features of the statements were compared across objective conditions. So "largely invisible" cannot be traced to a specific measure.