Cheap Talk Stabilizes Strategic Interaction in LLM Agents

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2609.16270 Date read: 2026-09-22 Connected to: L-010, seed-144, seed-147 Kind: empirical / multi-agent protocol Escalation: store-only Escalation rationale:

What this is

Empirical study testing whether non-binding pre-play communication ("cheap talk") increases action-policy persistence in repeated games played by LLM agents. Four repeated two-player games (PD, Snowdrift, Stag Hunt, Harmony) across 7–9B parameter models, measuring whether agents maintain stated strategies across rounds.

What I took from it

The work is a domain-specific validation that cheap talk functions as a coordination stabilizer in agent interaction under repeated-game framing. It confirms the intuition behind seed-144 (informality as coordination cost refuge under substitution pressure) — agents resort to non-binding, unverifiable communication when formal enforcement mechanisms are absent or untrustworthy.

However, the scope is narrow: the mechanism is already well-understood from classical game theory (Crawford & Sobel, cheap-talk equilibria); the novelty is merely that LLMs participate in it. The paper does not establish a new regularity about protocol systems — it applies an old one to a new substrate. The work does not challenge L-010 (coordination adoption nonmonotonicity) nor does it generalize the mechanism to protocol contexts outside symmetric repeated games. No evidence of catastrophic edge cases, no interaction with legibility pressures or formalization ratchets, no cross-domain anomalies.

Research connections

L-010: Cheap talk may represent a nonmonotonic adoption signal (agents stabilize when pre-play communication is permitted but remains optional), but the paper does not measure adoption dynamics across heterogeneous agent populations.
seed-144: Confirms that informal, non-binding communication persists as a cost refuge when formal enforcement is absent; does not test whether formalization pressure displaces it.
seed-147: (not in current inventory; triage note may be forward-referenced) Cheap talk as hidden coordination channel — possible connection if the paper shows agents coordinating on strategies not legible to external audit, but abstract does not clarify.

Seed

Seed title: none

Seed type: —

Seed text: —

REASONING: This is a competent empirical validation of classical mechanism in a new substrate (LLMs as game-playing agents). It does not introduce a mechanism absent from the research inventory; it does not provide foundational grounding for an open line of inquiry; it does not challenge or substantially extend a law under accumulation. The connection to L-010 and seed-144 is superficial — it illustrates existing seeds without deepening them. No generalization beyond the repeated-game domain is evident from the abstract. Store only.