Technical Summary: Communication Enhances LLMs' Stability in Strategic Thinking

Problem Statement

Large Language Models (LLMs) deployed as autonomous agents in multi-agent systems frequently exhibit pronounced context-dependent variability. This instability undermines predictable strategic behavior, particularly in environments requiring repeated interaction and strategic alignment, such as social dilemmas. While instability can stem from sampling noise or hardware nondeterminism, the authors identify the inherent context-dependence of modern LLMs as a critical failure mode. In strategic settings like the Prisoner's Dilemma, minor stochastic fluctuations can propagate into qualitatively distinct behavioral trajectories, rendering optimization for utility indistinguishable from stochastic drift. The paper posits that behavioral stability is a functional prerequisite for strategic alignment; agents must first be reliable before they can be optimized for "good" outcomes.

Methodology

The study evaluates whether short, costless pre-play messages (emulating the "cheap-talk" paradigm) can stabilize strategic behavior in LLM agents.

• Experimental Design: The authors conducted simulations using the Iterated Prisoner's Dilemma (IPD) over ten rounds. Agents were never informed of the finite time horizon to prevent trivial convergence.
• Models: The analysis focused on the 7B–9B parameter regime, selecting four state-of-the-art open-weights models: Qwen 2.5 7B, Falcon 3 7B, Granite 3.3 8B, and Gemma 2 9B. This range was chosen to test "high-risk" architectures that are computationally lightweight yet exhibit higher susceptibility to noise than larger models.
• Contexts: Experiments were run across six contextual framings (Neutral, Team, Social, Business, Environment, International Relations) plus a neutral baseline.
• Conditions:
  • Messaging: Agents exchanged one-sentence, non-binding messages before each decision.
  • No-Messaging: Agents acted based solely on play history.
  • Robustness Checks: Variations included zero-temperature decoding, variant and robust prompting styles, and networked interactions (Erdős-Rényi, Power Law, and Core-Periphery topologies).
• Metrics and Analysis:
  • Stability Measure: The authors quantified stability using the Root Mean Squared Error (RMSE) of a Locally Weighted Scatterplot Smoothing (LOWESS) regression fitted to the average cooperation trajectory. Lower RMSE indicates a smoother, more predictable trajectory.
  • Statistical Inference: To account for nested data structures (agents within simulations), the study employed simulation-level bootstrap resampling (10,000 iterations for dyadic, 1,000 for networked cases). Significance was assessed using 95% confidence intervals and binomial tests for omnibus patterns.
  • Temperature: Standard experiments used a sampling temperature of 0.8 to probe policy distributions rather than deterministic outputs, with a specific robustness check at temperature 0.

Key Results

1. General Stabilizing Effect

Across the majority of model-context pairings, pre-play communication significantly reduced trajectory noise (lowered RMSE).

• Magnitude: The stabilizing effect was most pronounced in models with high baseline volatility (e.g., Granite 3.3 8B and Qwen 2.5 7B), which saw transformative reductions in RMSE. Models with higher intrinsic stability (Falcon 3 7B and Gemma 2 9B) showed incremental gains.
• Consistency: In the standard dyadic setting, 20 out of 24 model-context pairs showed a positive difference (reduced RMSE), with 17 being statistically significant.

2. Zero-Temperature Findings

When sampling temperature was set to 0 (quasi-deterministic), the results revealed that instability is not solely an artifact of sampling noise.

• Elimination of Instability: For Granite and Qwen, messaging completely eliminated instability (RMSE dropped to 0) in contexts where the no-messaging condition was unstable.
• Contextual Destabilization: Conversely, for Falcon 3 and Gemma 2, the introduction of messaging in the "social" framing caused a sharp spike in instability. The authors attribute this to a semantic conflict where "cheap talk" signals clashed with the model's internal priors regarding the Prisoner's Dilemma payoff structure, causing oscillation between logic paths even under deterministic decoding.

3. Networked Interactions and Bandwidth Constraints

• Full Messaging: In networked settings (ER, Power Law, Core-Periphery), full-sentence messaging generally mirrored dyadic results, stabilizing behavior or having no significant effect.
• Constrained Communication: Limiting agents to "one-word" messages frequently destabilized the system, increasing RMSE below the no-communication baseline for models like Qwen and Gemma. The authors hypothesize that insufficient semantic bandwidth injects noise rather than signal, failing to generate shared conventions. Falcon 3 7b remained an outlier, maintaining stability even with low-fidelity signals.

4. Prompt Sensitivity

Robustness tests with variant and robust prompting confirmed that the general tendency toward reduced instability holds, though the magnitude and statistical significance vary. Granite 3.3 8B remained highly responsive to communication across all prompt styles, while Gemma 2 9b showed increased instability under robust prompting in specific contexts, suggesting sensitivity to lexical variations.

Key Contributions

1. Empirical Evidence for Cheap Talk: The paper demonstrates that non-binding, costless communication acts as a low-cost mechanism to constrain stochastic drift in LLM agents, yielding more consistent behavioral trajectories.
2. Characterization of LLM Strategic Regimes: The study identifies that LLMs occupy a distinct behavioral regime compared to classical game theory predictions (babbling equilibria) and human benchmarks (oversharing). LLMs exhibit a unique sensitivity where communication can either suppress noise or, in specific semantic conflicts, amplify instability.
3. Stability vs. Size: The findings challenge the assumption that stability scales monotonically with model size. Architectural choices and training priors were found to be more determinative of baseline volatility than raw parameter counts within the 7B–9B range.
4. The "Valley of Uncanniness": The paper identifies a critical design constraint in multi-agent systems: constrained communication (low bandwidth) can be more harmful than silence, acting as noise injection rather than a coordination signal.

Significance and Claims

The authors position these findings as a foundational step toward the reliable deployment of multi-agent LLM systems. They argue that stability is the "foundational substrate" upon which utility-maximizing interventions must be built. By demonstrating that cheap talk can dampen the inherent noise of LLM decision-making, the study provides a mechanism to secure the "control loop" of agentic systems.

However, the paper maintains a modest scope regarding its claims:

• It does not claim to solve strategic alignment but rather to reduce trajectory variance to make agents controllable and correctable.
• It acknowledges that communication rarely produces harmful instability but documents specific context-dependent exceptions (e.g., semantic conflicts in "social" framings).
• The study is limited to the 7B–9B parameter range and the Prisoner's Dilemma, leaving the behavior of larger models (70B+) and environments with strictly dominant deceptive incentives as open questions.

Ultimately, the paper concludes that while communication is a practical tool for improving predictability, its efficacy is mediated by model choice, prompt formulation, and contextual framing, requiring careful calibration to avoid introducing new sources of instability.