Research map / Large language models / Instruction and fine-tuning
From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning
Jike Zhong, Yuxiang Lai, Ming Li, Yuheng Li, Wuao Liu, Behzad Dariush, Konstantinos Psounis, Shao-Yuan Lo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SFT with Thinking-RFT | Qwen2.5-7B-Instruct · OpenToM overall | accuracy (%): 83.14 -> 89.14 | Improved |
| add explicit thinking to RFT (ablation) | No-Thinking-RFT · OpenToM overall (7B) | accuracy (%): 78.71 -> 89.14 | Improved |
| replace SFT with Thinking-RFT | Qwen2.5-VL-7B-Instruct · MMToM / MuMA-ToM average | accuracy (%): 74.75 -> 82.20 | Improved |
| replace SFT with Thinking-RFT on shortcut-prone ExploreToM | Qwen2.5-7B-Instruct · ExploreToM in-domain | accuracy (%): 95.8 -> 94.3 | Hurt |
| train on shortcut-prone ExploreToM | Qwen2.5-7B-Instruct · Hi-ToM OOD | accuracy (%): 43.6 -> 35.3 | Hurt |
| scale model from 3B to 7B on shortcut-prone ExploreToM | SFT · ExploreToM in-domain | accuracy (%): 96.4 -> 95.8 | No clear effect |
| scale model from 3B to 7B on shortcut-free OpenToM | Thinking-RFT · OpenToM overall | accuracy (%): 83.0 -> 89.1 | Improved |
| replace SFT with Thinking-RFT for second-order generalization (train on first order) | Qwen2.5-7B-Instruct · OpenToM second-order | accuracy (%): 65.33 -> 74.33 | Improved |
| replace SFT with Thinking-RFT for unseen environments | Qwen2.5-VL-7B-Instruct · MMToM unseen environments | accuracy drop: -7.11 -> -2.4 | Improved |
| replace SFT with Thinking-RFT for cross-dataset transfer | Qwen2.5-7B-Instruct · OpenToM->ToMATO | accuracy (%): 56.8 -> 70.4 | Improved |
| replace SFT with Thinking-RFT under counterfactual rewrites | Qwen2.5-7B-Instruct · OpenToM counterfactual | accuracy (%): 66.00 -> 81.24 | Improved |
| replace SFT with Thinking-RFT on first-order OpenToM | Qwen2.5-7B-Instruct · OpenToM first-order | accuracy (%): 93.00 -> 95.00 | No clear effect |
Closest papers by meaning
- Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner (ICML 2025 · similarity 0.90)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.89)
- Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (ICLR 2026 · similarity 0.89)
- Unveiling the Cognitive Compass: Theory-of-Mind–Guided Multimodal Emotion Reasoning (ICLR 2026 · similarity 0.88)
- RL of Thoughts: Navigating LLM Reasoning with Inference-time Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
