Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Agent Learning via Early Experience
Kai Zhang, Xiangchao Chen, Bo Liu, Tianci Xue, Zeyi Liao, Zhihan Liu, Xiyao Wang, Yuting Ning, Zhaorun Chen, Xiaohan Fu, Jian Xie, Yuxuan Sun and 18 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 |
|---|---|---|---|
| add implicit world modeling stage before imitation learning | Llama-3.2-3B agent · WebShop | success rate (%): 41.8 -> 60.2 | Improved |
| add self-reflection data to imitation learning | Llama-3.2-3B agent · TravelPlanner | success rate (%): 19.4 -> 32.2 | Improved |
| add self-reflection data to imitation learning | Llama-3.1-8B agent · ScienceWorld | success rate (%): 54.7 -> 68.0 | Improved |
| add early experience (IWM or SR) to imitation learning | Llama-3.2-3B agent · WebArena-Lite | success rate (%): 6.1 -> 8.5 | Improved |
| add early experience (IWM or SR) to imitation learning | Llama-3.2-3B agent · SearchQA | F1: 38.0 -> 39.0 | No clear effect |
| add early experience to imitation learning, evaluated out-of-domain | Llama-3.1-8B agent · ALFWorld (out-of-domain split) | success rate (%): 63.3 -> 78.1 | Improved |
| initialize GRPO RL from early-experience checkpoints instead of imitation learning (ablation) | WebShop, ALFWorld, SearchQA agents · post-RL performance | success rate / F1 | Improved |
| run GRPO directly from raw pretrained model with no supervised stage (ablation) | GRPO baseline · WebShop, ALFWorld, SearchQA | post-RL performance | Hurt |
| use only 1/8 of expert demonstrations with early experience (ablation) | Llama-3.1-8B agent · WebShop | success rate (%) | Improved |
| increase branching factor K for implicit world modeling (ablation) | IWM · WebShop and ALFWorld | success rate (%) | Improved |
| increase branching factor K for self-reflection to large values (ablation) | SR · WebShop and ALFWorld | success rate (%) | Mixed |
| add test-time long chain-of-thought to imitation-learned model (ablation) | Llama-3.1-8B IL · WebShop | success rate (%): 47.3 -> 0.0 | Hurt |
Closest papers by meaning
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.89)
- Large Language Model Agents Are Not Always Faithful Self-Evolvers (ICML 2026 · similarity 0.89)
- Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.89)
- From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills (NeurIPS 2026 · similarity 0.88)
- Don't Just Fine-tune the Agent, Tune the Environment (ICLR 2026 · similarity 0.88)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 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.
