Technical Summary: Aligned but Not Partner-Specific

Problem Statement

The paper addresses a critical gap in understanding how Multimodal Large Language Models (MLLMs) achieve coordination in cooperative referential tasks. While prior research establishes that human interlocutors in repeated reference games (e.g., the Tangrams task) develop "conceptual pacts"—short, partner-specific conventions grounded in shared interaction history—it remains unclear whether MLLMs exhibit similar behavior.

Existing studies suggest MLLMs align on labels but fail to compress descriptions over time. However, a methodological limitation persists: observed lexical alignment in MLLMs could stem from partner-specific grounding (human-like adaptation) or merely from shared task vocabulary and pretrained priors (statistical consistency). Without a control that breaks interaction history while preserving the task structure, model, and turn sequence, it is impossible to distinguish between these two mechanisms.

Methodology

The authors propose a novel methodological framework to disentangle partner-specific grounding from global style alignment.

1. Experimental Setup

• Task Adaptation: The study adapts the KTH Tangrams corpus paradigm for MLLM dyads. In this cooperative game, an "informant" describes a target shape to a "guesser" who must identify it.
• Model Selection: After screening, GPT-5 (specifically the gpt-5-2025-08-07 configuration with high reasoning and medium verbosity) was selected. It was the only model to satisfy three criteria: robust visual discrimination of target tags, strict adherence to task rules (no leakage of target labels), and support for multimodal interaction history across rounds.
• Corpus Construction:
  • Real Dyads: 45 MLLM dyads (905 rounds) compared against 42 human dyads (3,288 rounds) from the KTH corpus.
  • Pseudo-Dyads (The Core Contribution): A constrained extraction algorithm generates synthetic dialogues by pairing rounds from different source dyads. This breaks the partner history while maintaining the referential target, round position, turn count, and success outcome.
  • Constraints: Pairs are matched based on: (1) same target shape, (2) comparable trajectory position, (3) similar turn counts (per role), and (4) identical success/failure outcomes.

2. Analytical Framework

The analysis operates across three layers:

1. Task Competence and Effort: Measuring success rates and the number of turns required to reach a correct guess.
2. Description Strategy: Analyzing the compression of descriptions (content-word volume) over time.
3. Alignment Dynamics: Quantifying lexical alignment using lexical cores (clusters of near-equivalent surface forms sharing a content-based referential basis). Alignment is measured by comparing "Real" dyads against their "Pseudo" counterparts.

Key Contributions

1. Constrained Pseudo-Dyad Baseline: The paper introduces a method to generate pseudo-dyads that serve as a counterfactual control. This allows researchers to test if alignment depends on specific partner interaction or is merely a byproduct of the task and model priors.
2. MLLM Tangrams Corpus: The creation of a comparative dataset involving 45 MLLM dyads and 905 rounds, adapted specifically for multimodal agents with automated scoring and role-swapping.
3. Three-Layer Analytic Protocol: A comprehensive framework for comparing dyadic corpora that isolates effort, strategy, and alignment dynamics, applicable to future agent studies.

Results

The study reveals a fundamental divergence between human and MLLM coordination mechanisms:

1. Task Competence and Effort

• Humans: Show a clear trajectory of increasing efficiency. They start with longer descriptions (5 turns) and compress them significantly (2.56 turns) as they form conventions. Success rates remain near-perfect (99.79%) throughout.
• MLLMs: Maintain fixed effort levels. They start with verbose descriptions (~2 turns) and do not compress them over time. While initial success is high (95.03%), performance declines in later rounds, suggesting a failure to leverage accumulated referential history.

2. Description Strategy

• Humans: Exhibit proportional compression, reducing content-word volume to ~30% of initial levels by the end of the session.
• MLLMs: Maintain descriptions at ~90% of their initial volume. They rely on exhaustive, verbose geometric descriptions rather than developing compact, partner-specific labels.

3. Alignment Dynamics

• Humans: Show a robust "Real vs. Pseudo" gap. Real human dyads are significantly more likely to use shared lexical cores than pseudo-paired humans (Odds Ratio = 6.85). This confirms that human alignment is driven by interactive grounding.
• MLLMs: Show no significant difference between Real and Pseudo conditions. Both conditions exhibit near-ceiling lexical overlap (Real mean: 0.979; Pseudo mean: 0.982).
  • Interpretation: MLLM alignment is not driven by partner-specific adaptation but by the sheer volume of verbose descriptions and shared pretrained priors. The "alignment" is a statistical artifact of describing the same visual stimuli with a consistent, instruction-tuned vocabulary.

Significance and Claims

The paper concludes that MLLM agents achieve coordination without convention.

• Distinction from Human Dialogue: Unlike humans, who form compact, history-dependent referring expressions (conceptual pacts), MLLMs succeed through "verbose description." They achieve high task success not by learning a partner's specific shorthand, but by generating exhaustive descriptions that are statistically likely to be understood by any agent with similar pretrained priors.
• Methodological Implication: The study demonstrates that raw performance metrics (e.g., success rate, lexical overlap) are insufficient for evaluating cooperative dialogue competence. Without a pseudo-dyad control, high alignment in agents can be misinterpreted as human-like grounding.
• Theoretical Contribution: The findings suggest that current instruction-tuned MLLMs operate in a regime of global style alignment rather than partner-specific grounding. The formation of true conventions appears to require explicit cost-sensitive training objectives (minimizing message length) rather than standard success-only optimization.

The authors position this work as a necessary correction to the interpretation of agent alignment, providing a reusable framework (pseudo-dyad generation and analysis pipeline) to rigorously test whether future models develop genuine interactive grounding.