Communication Enables Cooperation in LLM Agents: A Comparison with Curriculum-Based Approaches

Eliciting cooperation in multi-agent LLM systems is critical for AI alignment. We investigate two approaches: direct communication and curriculum learning. In a 4-player Stag Hunt, a one-word “cheap talk” channel increases cooperation from 0% to 96.7%, demonstrating communication as a robust coordination mechanism. In contrast, we find that curriculum learning is highly sensitive to design choices: our pedagogical curriculum through progressively complex games reduced agent payoffs by 27.4% in an Iterated Public Goods Game with Punishment, demonstrating that optimizing for short-term rationality can actively undermine alignment goals. Qualitative analysis reveals that curricula emphasizing defection-equilibrium games can induce “learned pessimism” in agents. These findings suggest that for coordination problems, simple communication protocols may be more reliable than experience-based training, and that curriculum design for social dilemmas requires careful attention to the strategic lessons embedded in game sequences.

💡 Research Summary

This paper investigates two fundamentally different methods for inducing cooperation among multi‑agent large language model (LLM) systems: a direct “cheap‑talk” communication channel and a curriculum‑learning approach that sequentially exposes agents to increasingly complex games. The authors conduct experiments with four state‑of‑the‑art instruction‑tuned LLMs (Mixtral‑8x22B, Qwen2.5‑72B, Llama‑3.3‑70B, DeepSeek‑V3) across several classic game‑theoretic environments, most notably a four‑player Stag Hunt and an Iterated Public Goods Game with Punishment (IPGG+P).

In the Stag Hunt pilot, agents are either allowed or denied a one‑word, costless communication token before each round. Without communication, heterogeneous groups achieve 0 % cooperation, while the same agents with a single word of “cheap talk” coordinate successfully 96.7 % of the time. Even when agents are paired into same‑family coalitions (which already enjoy a modest baseline cooperation of 52.2 %), the communication channel pushes coordination to a perfect 100 %. These results demonstrate that modern LLMs can understand the strategic value of signaling, converge on a shared protocol, and trust one another’s messages without any explicit training on the task.

The curriculum‑learning arm of the study constructs four conditions, each consisting of 30 trials: (1) Full Curriculum (2‑player Prisoner’s Dilemma → N‑player Prisoner’s Dilemma → 3‑round IPGG → 10‑round IPGG+P), (2) Scrambled order, (3) Direct Precursor (only the two IPGG stages), and (4) Control (only the final IPGG+P). After each stage, Claude Opus 4.1 generates a concise strategic “lesson” from the game logs, which is prepended to the prompts for the next stage. The authors find a monotonic degradation of performance as curriculum complexity increases: the control group attains the highest average payoff (211.7 tokens), while the Full Curriculum reduces it by 27.4 % to 153.6 tokens. The Scrambled and Direct Precursor conditions also underperform relative to the control, indicating that the mere presence of a curriculum—regardless of order—can be harmful if not carefully designed.

A crucial diagnostic experiment replaces the AI‑generated lessons with neutral, non‑strategic prompts (“Consider your options carefully”). This “Neutral Lesson Ablation” restores average payoffs to 251.1 tokens—a 63.5 % improvement over the Full Curriculum with strategic lessons—while cooperation rates remain at 0 % (the task is not a coordination game). The stark contrast isolates the failure mechanism: the content of the early‑stage lessons, which emphasized defection‑dominant equilibria, poisoned agents’ priors, leading to a phenomenon the authors label “learned pessimism.”

Qualitative analysis of chain‑of‑thought traces reveals three failure modes: (1) Learned pessimism—agents extrapolate short‑horizon defection experience to longer games, pre‑emptively defecting in the first round of IPGG+P; (2) Heuristic over‑fitting—agents rigidly apply simple rules (e.g., always punish the lowest contributor) without contextual nuance; (3) Generic role‑play—control agents produce textbook‑style reasoning detached from opponent behavior. Coding of reasoning traces shows learned pessimism in 62 % of Full Curriculum traces versus <5 % in the neutral condition, confirming the causal link between strategic lesson content and performance loss.

The paper also probes whether cheap talk remains beneficial in the more complex IPGG+P environment. Two incentive structures are tested: a standard 1.6× multiplier and a high‑stakes 4.0× multiplier. With the low multiplier, communication raises contribution rates from 48 % to 71 % but paradoxically lowers average welfare (184.4 → 127.5 tokens) because cooperators become easy exploitation targets. Under the high multiplier, communication yields perfect coordination (100 % contribution) and maximizes welfare (480 tokens). This demonstrates that cheap talk is a powerful coordination tool, but its welfare impact depends critically on the underlying payoff structure.

Overall, the study contributes three major insights: (i) Minimal, costless communication can dramatically improve coordination among heterogeneous LLM agents, even in the absence of explicit training; (ii) Curriculum learning for social dilemmas is highly sensitive to the strategic content of early games—front‑loading defection‑equilibrium games can induce pessimistic priors that cripple later cooperation; (iii) The design of lesson generation (AI‑generated versus neutral) is a pivotal factor, suggesting that human‑curated or carefully balanced lesson content may be necessary for successful curriculum‑based alignment.

The authors conclude by recommending that system designers prioritize simple communication protocols for coordination problems, while treating curriculum design as a high‑risk component that must avoid embedding counter‑productive strategic lessons. Future work should explore alternative curricula that begin with coordination games, assess the impact of higher‑quality (human‑written) lesson generation, and investigate whether fine‑tuning rather than in‑context learning can more robustly embed cooperative principles. Limitations include the specific ordering of games, reliance on a single lesson‑generation model, and the fact that experiments were conducted via API calls rather than on‑device fine‑tuning. Nonetheless, the findings provide a clear empirical foundation for the comparative efficacy of communication versus curriculum learning in multi‑agent LLM alignment.