AxiomicLabs Releases 'Tiny Theory of Mind' Benchmark for Small Models
AxiomicLabs has launched "Tiny Theory of Mind," a new benchmark hosted on Hugging Face designed to evaluate the social reasoning and cognitive abilities of small language models. As small language models are increasingly deployed in edge devices and conversational tools, assessing their social-cognitive abilities is critical for building nuanced AI interactions. This benchmark provides researchers with a dedicated tool to test fundamental reasoning about mental states without relying on large-scale model requirements. The benchmark contains 2,000 synthetic multiple-choice questions covering 40 distinct constructs, ranging in human difficulty from Pre-K to 6th grade. It is primarily tailored for base-model continuation log-likelihood scoring, meaning it can test raw model capabilities without needing specialized instruction tuning or prompting.
## BACKGROUND
Theory of Mind (ToM) is a concept from cognitive psychology referring to the capacity to attribute mental states—such as beliefs, intents, desires, and emotions—to oneself and others. In artificial intelligence research, evaluating ToM helps determine how well language models can infer human intentions, resolve ambiguities, and participate effectively in complex social interactions.