Explicit Trait Inference for Multi-Agent Coordination¶

Conference: ACL 2026 arXiv: 2604.19278 Code: None Area: LLM Multi-Agent Systems / Social Reasoning Keywords: Multi-agent coordination, trait inference, warmth-competence dimensions, social cognition, game theory

TL;DR¶

This paper proposes Explicit Trait Inference (ETI), a method that enables LLM agents to reason about and track partners' behavioral traits along the psychological dimensions of warmth and competence. ETI reduces payoff loss by 45–77% in economic games and improves task performance by 3–29% on MultiAgentBench.

Background & Motivation¶

State of the Field: LLM-based multi-agent systems (MAS) have demonstrated promise on complex tasks, yet remain prone to coordination failures—goal drift, error cascades, insufficient information sharing, and behavioral misalignment—limiting their reliability and scalability.

Limitations of Prior Work: (1) Structured approaches (e.g., CAMEL, ChatDev) organize collaboration through fixed roles but do not address how agents reason about and adapt to one another; (2) Theory-of-Mind (ToM) methods primarily model transient mental states (beliefs, intentions) rather than stable behavioral traits (reliability, willingness to cooperate); (3) reputation systems track only task-level metrics (success rates) without capturing the why and how of behavioral patterns.

Root Cause: The core challenge is not whether agents can execute individual actions, but whether they can coordinate effectively with one another—which requires forming stable, actionable cognitive representations of partners.

Paper Goals: To provide a lightweight, psychologically grounded mechanism that enables agents to infer partners' traits from interaction history and adjust behavior accordingly.

Starting Point: The paper draws on the warmth-competence dual-dimension model from social psychology (Fiske et al., 2007), mapping social evaluation to actionable coordination signals.

Core Idea: Agents explicitly infer and maintain trait profiles of their partners along warmth (trust/cooperation) and competence (skill/reliability) dimensions, using these profiles to guide delegation, communication, and strategy adjustment.

Method¶

Overall Architecture¶

ETI is a prompt- and context-management-based framework. After each interaction, an agent receives a structured summary of task objectives, actions, communications, and outcomes, and is prompted to reason about the partner's traits. The agent produces (a) a 1–7 Likert rating for each trait and (b) brief supporting evidence. These profiles are appended to the context for use in subsequent planning and execution.

Key Designs¶

Warmth-Competence Trait Framework:
- Function: Provides agents with a structured representation of partners.
- Mechanism: Eight behaviorally anchored traits are organized into two dimensions—warmth (goal alignment, collaborativeness, trustworthiness, maliciousness) and competence (execution ability, reliability, adaptability, efficiency). Trait definitions explicitly separate warmth from competence to prevent conflation common in everyday language (e.g., misattributing uncooperativeness as incompetence).
- Design Motivation: The warmth dimension addresses goal drift and unreliable collaboration, prompting agents to clarify intentions or discount unreliable outputs when warmth is low; the competence dimension addresses execution errors and error cascades, prompting task reallocation or additional verification when competence is low.
Reason–Plan–Execute Loop:
- Function: Seamlessly integrates trait inference into the multi-agent pipeline.
- Mechanism: After each iteration, an agent (1) infers partner traits from action and outcome history; (2) incorporates structured trait profiles into its context; and (3) plans and executes the next step using the enriched context. Prompts instruct the model to focus on dominant behavioral patterns rather than isolated events, remaining domain-agnostic.
- Design Motivation: The pure-prompt approach requires no fine-tuning or additional data, incurs minimal overhead, and is applicable to arbitrary MAS architectures.
Competence Parameterization in Economic Games:
- Function: Provides a controlled environment with ground truth for evaluating trait inference accuracy.
- Mechanism: Standard Prisoner's Dilemma and Stag Hunt games are augmented with a competence parameter—an agent's intended action succeeds with probability \(p_i\). This allows agents to infer intent (cooperative vs. selfish) from actions and competence (success rate) from outcomes. Agents play 50 rounds against parameterized rule-based opponents.
- Design Motivation: Economic games offer simple yet adaptive decision-making problems that enable precise evaluation of trait inference accuracy.

Loss & Training¶

ETI is a pure-prompt method involving no training. Qwen3-8B serves as the agent backbone, with 25 independent repetitions across all configurations.

Key Experimental Results¶

Main Results¶

In economic games (Qwen3-8B vs. rule-based opponents):

Game	Method	Payoff Deviation ↓	Notes
Prisoner's Dilemma	CoT Baseline	High	No opponent modeling
Prisoner's Dilemma	ETI	45–77% reduction	Trait-aware decision-making
Stag Hunt	CoT Baseline	High	Defaults to conservative strategy
Stag Hunt	ETI	Significant improvement	Accurately estimates cooperation likelihood

On MultiAgentBench:

Scenario Type	ETI Task Gain	Coordination Gain
Cooperative	3–29%	6–42%
Competitive	Improvement observed	Significant

Ablation Study¶

Configuration	Effect	Notes
ETI (informative profiles)	Best	Diverse trait judgments drive improvement
ETI (generic profiles)	Marginal	Non-discriminative profiles are ineffective
No trait inference	Baseline	CoT focuses only on task-level reasoning
Trait-to-behavior prediction	Accurate	ETI profiles reliably predict agent actions

Key Findings¶

ETI's gains derive not from more reasoning but from more targeted reasoning—generic profiles are nearly ineffective, and only highly informative profiles yield benefits.
Trait inference capability is validated: ETI-generated profiles reliably predict subsequent agent behavior, demonstrating that models can infer stable traits from interaction history.
On complex MultiAgentBench scenarios, ETI achieves up to 29% improvement, confirming generalization from controlled settings to realistic MAS.
The warmth dimension proves more important in cooperative scenarios (detecting unreliable partners), while the competence dimension is more critical in complex task scenarios (enabling task reallocation).

Highlights & Insights¶

Importing the warmth-competence model from social psychology into MAS represents an elegant interdisciplinary innovation: human social trust and coordination operate along precisely these two dimensions, making their formalization as an inter-agent reasoning framework highly natural.
The behaviorally anchored trait definition design is worth emulating: explicit behavioral descriptions (rather than abstract concepts) prevent LLMs from conflating dimensions during inference, a principle applicable to any scenario requiring structured LLM judgment.
The pure-prompt implementation entails zero additional training cost and plug-and-play deployment, making it highly practical for real-world MAS applications.

Limitations & Future Work¶

The accuracy of trait inference depends on the LLM's social reasoning capacity; weaker models may produce inaccurate profiles.
The current framework assumes traits are relatively stable, limiting its ability to detect strategic deception (e.g., initial cooperation followed by defection).
The choice of eight traits, though grounded in psychology, is not necessarily optimal for MAS design—task-specific trait dimensions may prove more effective.
In large-scale MAS (>10 agents), maintaining trait profiles for all partners may incur prohibitive context costs.

vs. ToM methods (Li et al., 2023): These model transient beliefs and intentions but do not track stable traits; ETI provides persistent representations across interactions.
vs. Reputation systems (Lou et al., 2026): These track metrics such as success rates without capturing behavioral motivations; ETI offers richer representations (the why and how).
vs. CoT/Reflexion: These structure only task-level reasoning without reasoning about others; ETI extends reasoning into the social domain.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ — First systematic integration of psychological trait theory into MAS
Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Controlled games + realistic MAS, validating accuracy through causality
Writing Quality: ⭐⭐⭐⭐⭐ — Clear motivation, excellent interdisciplinary integration
Value: ⭐⭐⭐⭐⭐ — Introduces a lightweight yet effective new paradigm for LLM multi-agent coordination