DixitWorld: Evaluating Multimodal Abductive Reasoning in Vision-Language Models with Multi-Agent Dixit Gameplay

TL;DR

DixitWorld introduces DixitArena and DixitBench to evaluate multimodal abductive reasoning in vision-language models within dynamic, multi-agent settings. The framework reveals a pronounced storyteller–listener asymmetry: large models excel as listeners but struggle to generate balanced, creative clues, while smaller models tend to be more imaginative but less discriminative. DixitBench correlates strongly with listener performance in DixitArena, validating a lightweight proxy for hypothesis selection. The work highlights a central trade-off between generative creativity and discriminative understanding, underscoring the need for models that explicitly reason about ambiguity and communicative intent. Together, the suite provides a fine-grained lens on pragmatic reasoning in vision-language systems and guides future directions toward balanced, interactive intelligence.

Abstract

Multimodal abductive reasoning--the generation and selection of explanatory hypotheses from partial observations--is a cornerstone of intelligence. Current evaluations of this ability in vision-language models (VLMs) are largely confined to static, single-agent tasks. Inspired by Dixit, we introduce DixitWorld, a comprehensive evaluation suite designed to deconstruct this challenge. DIXITWORLD features two core components: DixitArena, a dynamic, multi-agent environment that evaluates both hypothesis generation (a "storyteller" crafting cryptic clues) and hypothesis selection ("listeners" choosing the target image from decoys) under imperfect information; and DixitBench, a static QA benchmark that isolates the listener's task for efficient, controlled evaluation. Results from DixitArena reveal distinct, role-dependent behaviors: smaller open-source models often excel as creative storytellers, producing imaginative yet less discriminative clues, whereas larger proprietary models demonstrate superior overall performance, particularly as listeners. Performance on DixitBench strongly correlates with listener results in DixitArena, validating it as a reliable proxy for hypothesis selection. Our findings reveal a key trade-off between generative creativity and discriminative understanding in multimodal abductive reasoning, a central challenge for developing more balanced and capable vision-language agents.

Figures (4)

• Figure 1: An illustration of Dixit gameplay.
• Figure 2: Head-to-head game scores for all model pairings in the round-robin tournament, demonstrating consistent performance hierarchies.
• Figure 3: Distribution of storyteller round outcomes. Only the "Partial-Correct" outcome yields points for the storyteller, making it the desired result.
• Figure 4: Comparison between direct selection and entailment scoring strategies in DixitBench.