Jev-ToM: Evaluating Theory of Mind in Typed Decision Models
We present Jev-ToM, comparing five typed decision models and seven generative baselines on 11,117 English text questions from nine ToM benchmarks across six capability groups.
ProofPaper ↗
Key points
- Dialogue and collaborative assistants repeatedly judge what a partner knows, believes, or intends.
- Typed decision models offer a potential alternative to generative large language models for bounded Theory of Mind (ToM) selection.
- Other typed models show distinct strengths across tasks.
- The findings motivate testing Jev as a ToM selection component in assistants, with task-specific confidence calibration, candidate design, and output validation as integration priorities.
Sources (1)
- [1]Jev-ToM: Evaluating Theory of Mind in Typed Decision ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 03:59 PM
We present Jev-ToM, comparing five typed decision models and seven generative baselines on 11,117 English text questions from nine ToM benchmarks across six capability groups.
Dialogue and collaborative assistants repeatedly judge what a partner knows, believes, or intends.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026PRISM: Calibrating Black-Box Judges with Counterfactual Probes
- Oct 13, 2026Agent-Reflex: Teaching Language Agents to Act Reflectively
- Oct 13, 2026Virtual Head Attention
- Oct 13, 2026Diff Diffusion: Predicting Edit Paths with Diffusion Models for Efficient Code Editing
- Oct 12, 2026Deploy decision models with llama.cpp on Amazon SageMaker AI
- Oct 12, 2026Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards