Research map / Large language models / Agents, code and math
MindZero: Learning Online Mental Reasoning With Zero Annotations
Shunchi Zhang, Jin Lu, Chuanyang Jin, Yichao Zhou, Zhining Zhang, Tianmin Shu
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 |
|---|---|---|---|
| add explicit LLM prior modeling to the reward (ablation) | MindZero Qwen3-4B · Household proactive assistance | speedup (%): 17.0 -> 19.1 | Improved |
| maintain multiple hypotheses instead of a single most likely one (ablation) | MindZero Qwen3-4B · Household proactive assistance | speedup (%): 10.3 -> 19.1 | Improved |
| add entropy bonus to the reward (ablation) | MindZero Qwen3-4B · Household proactive assistance | speedup (%): 5.2 -> 19.1 | Improved |
| add multiple hypotheses (compute cost) (ablation) | MindZero Qwen3-4B · Household proactive assistance | TFLOPs: 132.6 -> 201.2 | Hurt |
| train with MindZero self-supervised RL | Qwen3-VL-4B · GridWorld QA | accuracy: 37.7 -> 95.0 | Improved |
| train with MindZero self-supervised RL | Qwen3-VL-8B · GridWorld proactive assistance | speedup (%): -0.1 -> 24.5 | Improved |
| train with MindZero self-supervised RL | Qwen3-4B · Household proactive assistance | speedup (%): 2.3 -> 19.1 | Improved |
| train Llama-3.2-3B with MindZero after format warm-up | Llama-3.2-3B · Household proactive assistance | speedup (%): 2.3 -> 4.3 | No clear effect |
| train Qwen3-4B with MindZero (human study) | Qwen3-4B · Household human experiment | speedup (%): 2.62 -> 19.70 | Improved |
| use Gemini-3-Flash instead of MindZero Qwen3-4B | helper agent · Household human experiment | speedup (%): 19.70 -> 23.40 (p=0.24) | No clear effect |
| use thinking mode for GPT-5.2 | GPT-5.2 · Household QA | accuracy: 65.0 -> 73.5 | Improved |
| use thinking mode for Gemini/GPT-5.2 on GridWorld | GPT-5.2 · GridWorld QA | accuracy: 50.7 -> 50.7 | No clear effect |
Closest papers by meaning
- AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.90)
- Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner (ICML 2025 · similarity 0.88)
- Evolving Minds: Logic-Informed Inference from Temporal Action Patterns (ICML 2025 · similarity 0.87)
- Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models (ICLR 2025 · similarity 0.87)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL (NeurIPS 2025 · similarity 0.87)
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.
