Theory of Mind (ToM) can be used to assess the capabilities of Large Language Models (LLMs) in complex scenarios where social reasoning is required. While the research community has proposed many ToM benchmarks, their hardness varies greatly, and their complexity is not well defined. This work proposes a framework inspired by cognitive load theory to measure the complexity of ToM tasks. We quantify a problem’s complexity as the number of states necessary to solve it correctly. Our complexity measure also accounts for spurious states of a ToM problem designed to make it apparently harder. We use our method to assess the complexity of five widely adopted ToM benchmarks. On top of this framework, we design a prompting technique that augments the information available to a model with a description of how the environment changes with the agents’ interactions. We name this technique Discrete World Models (DWM) and show how it elicits superior performance on ToM tasks.

Multimodal Argument Mining (MAM) is a recent area of research aiming to extend argument analysis and improve discourse understanding by incorporating multiple modalities. Initial results confirm the importance of paralinguistic cues in this field. However, the research community still lacks a comprehensive platform where results can be easily reproduced, and methods and models can be stored, compared, and tested against a variety of benchmarks. To address these challenges, we propose MAMKit, an open, publicly available, PyTorch toolkit that consolidates datasets and models, providing a standardized platform for experimentation. MAMKit also includes some new baselines, designed to stimulate research on text and audio encoding and fusion for MAM tasks. Our initial results with MAMKit indicate that advancements in MAM require novel annotation processes to encompass auditory cues effectively.