M3-BENCH: Process-Aware Evaluation of LLM Agents Social Behaviors in Mixed-Motive Games

TL;DR

M3-Bench tackles the challenge of evaluating advanced social behaviors of LLM agents in mixed-motive games by shifting from outcome-only benchmarks to a process-aware framework that analyzes action, reasoning, and communication. It introduces a four-level task design and three analytical modules—Behavioral Trajectory Analysis, Reasoning Process Analysis, and Communication Content Analysis—whose outputs are integrated into portrait-style characterizations anchored by the Big Five and Social Exchange Theory. The framework enables cross-task portrait generation and audits for alignment between what agents do, think, and say, uncovering inconsistencies such as superficially cooperative behavior paired with opportunistic reasoning. Experimental results show that M3-Bench can reliably distinguish social capabilities across models and reveal risks associated with communication as a double-edged tool, underscoring the value of cross-auditing for safety governance and capability improvement in LLM agents.

Abstract

As the capabilities of large language model (LLM) agents continue to advance, their advanced social behaviors, such as cooperation, deception, and collusion, call for systematic evaluation. However, existing benchmarks often emphasize a single capability dimension or rely solely on behavioral outcomes, overlooking rich process information from agents' decision reasoning and communicative interactions. To address this gap, we propose M3-Bench, a multi-stage benchmark for mixed-motive games, together with a process-aware evaluation framework that conducts synergistic analysis across three modules: BTA (Behavioral Trajectory Analysis), RPA (Reasoning Process Analysis), and CCA (Communication Content Analysis). Furthermore, we integrate the Big Five personality model and Social Exchange Theory to aggregate multi-dimensional evidence into interpretable social behavior portraits, thereby characterizing agents' personality traits and capability profiles beyond simple task scores or outcome-based metrics. Experimental results show that M3-Bench can reliably distinguish diverse social behavior competencies across models, and it reveals that some models achieve seemingly reasonable behavioral outcomes while exhibiting pronounced inconsistencies in their reasoning and communication.

Figures (5)

• Figure 1: Iceberg metaphor for evaluating agents' social behavior: observable actions only capture surface-level performance, while key factors such as internal motives, beliefs about others, strategic reasoning, and communication tactics remain beneath the surface. The Before view represents outcome-oriented evaluation based solely on behavioral results, which can overlook process-level differences; the After view illustrates a process-aware evaluation that jointly analyzes what an agent does, thinks, and says, yielding a more comprehensive and diagnosable social behavior portrait.
• Figure 2: M3-Bench process-aware evaluation framework. A four level progressive suite of mixed-motive tasks gradually increases interaction difficulty while synchronously logging three complementary signals: behavioral trajectories, decision reasoning, and communicative dialogue. BTA/RPA/CCA perform quantification and analysis over these signals to characterize agents' action patterns, cognitive attributes, and linguistic strategies. Finally, the outputs of the three modules are mapped into interpretable social personality portraits under the constraints of the Big Five and Social Exchange Theory, producing portrait visualizations and risk diagnostic reports.
• Figure 3: A three-view illustration of key rounds in the repeated Prisoner’s Dilemma, highlighting a latent mismatch between behavior and internal reasoning.
• Figure 4: Portrait visualizations for Agent-LLaMa3.1-70B: (a) Big Five radar (0--100); (b) CCA radar (0--1); (c) BTA metrics by level (0--1); (d) RPA metrics under Silent vs. Comm (0--1).
• Figure 5: Full prompt used for L2 Repeated Prisoner’s Dilemma (RPD-10) in M3-Bench.