Research map / Large language models / Agents, code and math
Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents
Ruihan Yang, Fanghua Ye, Xiang Wei, Ruoqing Zhao, Kang Luo, Xinbo Xu, Bo Zhao, Ruotian Ma, Shanyi Wang, Zhaopeng Tu, Xiaolong Li, Deqing Yang and 1 more
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 |
|---|---|---|---|
| use average log-probability instead of min log-probability as confidence (ablation) | CoPO Qwen2.5-7B · ALFWorld | success rate: 81.5 -> 92.5 | Improved |
| use average log-probability instead of entropy as confidence (ablation) | CoPO Qwen2.5-7B · ALFWorld | success rate: 79.0 -> 92.5 | Improved |
| use average log-probability instead of max log-probability as confidence (ablation) | CoPO Qwen2.5-7B · ALFWorld | success rate: 89.0 -> 92.5 | Improved |
| add CoSFT cold-start initialization before RL (ablation) | CoPO Qwen2.5-7B · ALFWorld | success rate: 86.0 -> 92.5 | Improved |
| add balanced CoSFT initialization instead of expert-selected levels (ablation) | CoPO Qwen2.5-7B · ALFWorld | success rate: 87.5 -> 92.5 | Improved |
| reweight only successful trajectories instead of both (ablation) | CoPO Qwen2.5-7B · ALFWorld | success rate: 83.0 -> 92.5 | Improved |
| add cold start (ablation) | CoPO Qwen2.5-7B · ScienceWorld | success rate: 64.5 -> 72.0 | Improved |
| replace GRPO with CoPO | CoSFT Qwen2.5-7B · ALFWorld+ScienceWorld average | success rate: 68.3 -> 82.3 | Improved |
| replace GiGPO with CoPO | CoSFT Llama3.1-8B · ALFWorld+ScienceWorld average | success rate: 76.8 -> 81.0 | Improved |
| CoPO instead of GRPO token usage | CoSFT Qwen2.5-7B · ALFWorld+ScienceWorld average | tokens per trajectory: 4367.3 -> 1641.4 | Improved |
| CoSFT_exp instead of SFT | Qwen2.5-7B · ALFWorld+ScienceWorld average | success rate: 40.8 -> 54.0; tokens 577.5 -> 3562.4 | Mixed |
| R1-distilled variant instead of vanilla | Qwen2.5-7B ReAct · ALFWorld+ScienceWorld average | success rate: 26.5 -> 2.5 | Hurt |
Closest papers by meaning
- Think Twice, Act Once: A Co-Evolution Framework of LLM and RL for Large-Scale Decision Making (ICML 2025 · similarity 0.88)
- Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models (ICLR 2025 · similarity 0.88)
- Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning (ICML 2025 · similarity 0.88)
- Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.87)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · 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.
