Reproduce “From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jike Zhong, Yuxiang Lai, Ming Li, Yuheng Li, Wuao Liu, Behzad Dariush, Konstantinos Psounis, Shao-Yuan Lo
This research uncovers widespread "shortcut" issues in Theory of Mind (ToM) datasets, which hinder genuine mental-state reasoning evaluation. It demonstrates that Reinforcement Fine-Tuning with explicit reasoning (Thinking-RFT) effectively addresses this, improving ToM accuracy and generalization across various complex tasks, particularly for higher-order and multimodal scenarios.
Key insights
- Pervasive "shortcut" issues in popular Theory of Mind (ToM) datasets can mislead evaluations, causing models to achieve high accuracy without genuine mental-state reasoning and hindering robust ToM development.
- Reinforcement Fine-Tuning that emphasizes explicit reasoning chains (Thinking-RFT) effectively fosters genuine ToM, outperforming methods that rely on implicit reasoning or supervised fine-tuning.
- Thinking-RFT enhances models' ability to ground reasoning in causal factors by focusing attention on critical "causal hinge" sentences within narratives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.