Research map / Large language models / Instruction and fine-tuning
Language-based Trial and Error Falls Behind in the Era of Experience
Haoyu Wang, Guozheng Ma, Shugang Cui, Yilun Kong, Haotian Luo, Li Shen, Mengya Gao, Yichao Wu, Xiaogang Wang, Dacheng Tao
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 multi-turn PPO Evolving Stage after scout distillation (ablation) | SCOUT (Qwen2.5-0.5B-It) · Average over 6 unseen tasks | average success rate: 0.69 -> 0.81 | Improved |
| add multi-turn PPO Evolving Stage after scout distillation (ablation) | SCOUT (Qwen2.5-1.5B-It) · Average over 6 unseen tasks | average success rate: 0.71 -> 0.85 | Improved |
| add multi-turn PPO Evolving Stage after scout distillation (ablation) | SCOUT (Qwen2.5-3B-It) · Average over 6 unseen tasks | average success rate: 0.73 -> 0.86 | Improved |
| add multi-turn PPO Evolving Stage after scout distillation (ablation) | SCOUT (Qwen2.5-7B-It) · Average over 6 unseen tasks | average success rate: 0.78 -> 0.88 | Improved |
| add multi-turn PPO Evolving Stage after scout distillation (ablation) | SCOUT (LLaMA3.1-1B-It) · Average over 6 unseen tasks | average success rate: 0.63 -> 0.83 | Improved |
| add scout-trajectory Exploration & Distillation warm-up before multi-turn PPO | Qwen2.5-3B-It · Average over 6 unseen tasks | average success rate: 0.38 -> 0.86 | Improved |
| add Evolving Stage RL after distillation on Sudoku (ablation) | Qwen2.5-3B-It · Sudoku | success rate: 0.29 -> 0.97 | Improved |
| use DQN instead of PPO to train the scout | Scouts · Average over 6 unseen tasks | average success rate: 0.79 -> 0.83 | Mixed |
| add scout warm-up with a sub-optimal scout instead of direct RL | Qwen2.5-3B-It · Rubik's Cube Rotation3 | success rate: 0.14 -> 0.60 | Improved |
| add scout warm-up with a sub-optimal scout instead of direct RL | Qwen2.5-3B-It · Sudoku | success rate: 0.06 -> 0.98 | Improved |
| use SCOUT instead of direct RL on vision inputs | Qwen2.5-VL-3B-It · Visual-Sokoban | success rate: 0.57 -> 0.95 | Improved |
| use SCOUT instead of direct PPO for training cost | Qwen2.5-3B-It · Rubiks' Cube Rotation3 training cost | GPU hours: 24.0 -> 9.6 | Improved |
Closest papers by meaning
- Expanding LLM Agent Boundaries with Strategy-Guided Exploration (NeurIPS 2026 · similarity 0.91)
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.90)
- Can large language models explore in-context? (NeurIPS 2024 · similarity 0.89)
- Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.89)
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.89)
- WALL-E: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents (NeurIPS 2025 · 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.
