Research map / Multi-agent systems and games / LLM agents
Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Plan-and-Act, a planner/executor split for web agents trained on synthetic plan data generated by reverse-engineering trajectories and LLM-based plan expansion, plus dynamic replanning and CoT, reaching 57.58% on WebArena-Lite.
Problem
LLM web agents struggle with complex, long-horizon tasks because a single model must both plan and execute, and high-quality planning training data is scarce.
Approach
Separate a Planner LLM (generates structured high-level plans) from an Executor LLM (maps plan steps to web actions), and create planner training data by annotating successful trajectories with grounded plans and expanding them synthetically with LLM-generated query-plan pairs, plus dynamic replanning after each step.
Key results
- Plan-and-Act with Llama-3.3-70B reaches 57.58% success on WebArena-Lite, versus 49.1% for prior SOTA WebRL-3.1-70B.
- Dynamic replanning raises the static-planner result from 43.63% to 53.94%.
- QwQ-32B Plan-and-Act reaches 81.36% on WebVoyager text-only, a claimed new text-only SOTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add explicit planning instead of using a ReAct-style executor without a planner (ablation) | Plan-and-Act with synthetic-trajectory-trained executor · WebArena-Lite | success rate: 36.97 -> 43.63 | Helped |
| Use a zero-shot base planner rather than no planner (ablation) | Plan-and-Act · WebArena-Lite | success rate: 36.97 -> 23.63 | Hurt |
| Fine-tune the planner only on WebArena-Lite training plans (ablation) | Plan-and-Act · WebArena-Lite | success rate: 23.63 -> 20.60 | Hurt |
| Add grounded synthetic trajectories for planner training (ablation) | Plan-and-Act · WebArena-Lite | success rate: 20.60 -> 30.30 | Helped |
| Expand planner training with 10,000 synthetic query-plan pairs (ablation) | Plan-and-Act · WebArena-Lite | success rate: 30.30 -> 39.40 | Helped |
| Add 5,000 failure-analysis-guided synthetic plans (ablation) | Plan-and-Act · WebArena-Lite | success rate: 39.40 -> 43.63 | Helped |
| Replan after each executor step instead of using a static planner (ablation) | Plan-and-Act with targeted augmentation · WebArena-Lite | success rate: 43.63 -> 53.94 | Helped |
| Add CoT reasoning to the planner and executor (ablation) | Plan-and-Act with dynamic replanning · WebArena-Lite | success rate: 53.94 -> 57.58 | Helped |
| Add synthetic action trajectories while keeping the executor untrained (ablation) | No Planner · WebArena-Lite | success rate: 36.36 -> 36.97 | No clear effect |
Surprising findings
- Naively finetuning the planner on the 1,113 WebArena-Lite plans did not improve the finetuned executor (16.36% vs 17.16% for planner base), indicating overfitting.
- A base untuned Executor with a trained planner jumps from 9.85% to 44.24% with dynamic replanning.
Limitations
- Synthetic trajectory generation depends on a baseline model that can complete web tasks, limiting use where no training data exists.
- Dynamic replanning after every action is inefficient and slow.
- Evaluation on the full WebArena and WebVoyager uses fewer or different model configurations, and only text (HTML) observations are used.
Experimental setup
Datasets: WebArena-Lite training set WebVoyager (synthetic trajectories)
Benchmarks: WebArena-Lite WebArena WebVoyager
Metrics: success rate
Baselines: GPT-4-Turbo GPT-4o AWM WebPilot WebRL-Llama-3.1-70B NNetNav AutoWebGLM AgentOccam OpenWebVoyager Wilbur Agent-E WebVoyager
Builds on: LLMCompiler ReAct WebRL AgentOccam
Closest papers by meaning
- Large Language Models as Planning Domain Generators (Student Abstract) (AAAI 2024 · similarity 0.90)
- Planning in the Dark: LLM-Symbolic Planning Pipeline Without Experts (AAAI 2025 · similarity 0.90)
- Tree-Planner: Efficient Close-loop Task Planning with Large Language Models (ICLR 2024 · similarity 0.90)
- LLMs Can Plan Only If We Tell Them (ICLR 2025 · similarity 0.90)
- TravelPlanner: A Benchmark for Real-World Planning with Language Agents (ICML 2024 · similarity 0.90)
- Improving Large Language Model Planning with Action Sequence Similarity (ICLR 2025 · similarity 0.89)
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.