Research map / Large language models / Agents, code and math
AutoGuide: Automated Generation and Selection of Context-Aware Guidelines for Large Language Model Agents
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 AutoGuide, which mines contrastive offline trajectory pairs to generate conditional natural-language guidelines keyed by context, and retrieves top-k relevant guidelines at test time to prompt LLM agents.
Problem
LLM agents perform poorly in unfamiliar domains such as web navigation, and demonstration-based in-context learning is limited by context length, prompt sensitivity, and complex reasoning when using offline experiences.
Approach
Contrast pairs of trajectories with different returns to find the deviation timestep, summarize the shared context, and extract a conditional guideline ("when in context X, do Y"); at test time, identify the current context and select top-k matching guidelines to insert into the agent prompt, rather than providing all guidelines (as ExpeL does) or raw demonstrations.
Key results
- On WebShop, AutoGuide reaches 46% SR vs 30% for ReAct and 35% for ExpeL; on WebArena-Reddit it reaches 47.1% SR vs 8.0% for ReAct and 21.8% for ExpeL.
- On ALFWorld, AutoGuide achieves 79.1% SR vs 54.5% for ReAct; with Reflexion it reaches 88.1% vs 67.2%.
- On real websites with GPT-4V, AutoGuide solves 19/30 GitHub tasks vs 2/30 for SoM, and 14/20 Coursera tasks vs 1/20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add context-aware guidelines from offline experiences | ReAct · ALFWorld | success rate: 54.5\text{\%} -> 79.1\text{\%} | Helped |
| Add context-aware guidelines from offline experiences | ReAct · WebShop | success rate: 30\text{\%} -> 46\text{\%} | Helped |
| Add context-aware guidelines from offline experiences | ReAct · WebArena-Reddit | success rate: 8.0\text{\%} -> 47.1\text{\%} | Helped |
| Transfer WebShop guidelines with grounding to WebArena | ReAct · WebArena-Shopping | success rate: 10.2\text{\%} -> 20.4\text{\%} | Helped |
| Add context identification without guidelines (ablation) | ReAct · WebShop | success rate: 30\text{\%} -> 36\text{\%} | Helped |
| Add guideline extraction and selection without contexts (ablation) | ReAct · WebShop | success rate: 30\text{\%} -> 37\text{\%} | Helped |
| Combine context identification with guideline extraction and selection (ablation) | ReAct · WebShop | success rate: 37\text{\%} -> 46\text{\%} | Helped |
| Select top 1 context-aware guideline instead of none (ablation) | AutoGuide · WebShop | success rate: 30\text{\%} -> 42\text{\%} | Helped |
| Use top 3 rather than top 2 guidelines (ablation) | AutoGuide · WebShop | success rate: 46\text{\%} -> 47\text{\%} | Helped |
| Use top 5 rather than top 3 guidelines (ablation) | AutoGuide · WebShop | success rate: 47\text{\%} -> 43\text{\%} | Hurt |
| Add AutoGuide to SoM | SoM · GitHub | success rate: 2/30 -> 19/30 | Helped |
| Add AutoGuide to SoM | SoM · Google Flights | success rate: 5/20 -> 9/20 | Helped |
Surprising findings
- Simply adding more ReAct in-context demonstrations plateaus (WebShop SR 30% to 38% at 6-shot), and ExpeL combined with Reflexion underperforms ExpeL-free Reflexion pairing in some settings.
Limitations
- Performance depends on the diversity of offline experiences.
- No standardized metric for quality of generated contexts and guidelines.
- Evaluation on real websites relies on human experts and small task sets (20-30 tasks per domain).
Experimental setup
Datasets: ALFWorld WebShop WebArena Real-world multi-modal websites (GitHub, Google Flights, Coursera)
Benchmarks: ALFWorld WebShop WebArena-Reddit WebArena-Shopping GitHub Google Flights Coursera
Metrics: success rate reward
Baselines: ReAct ExpeL Reflexion SoM ReAct few-shot (1/2/4/6-shot)
Builds on: ReAct ExpeL Reflexion
Closest papers by meaning
- Self-Generated In-Context Examples Improve LLM Agents for Sequential Decision-Making Tasks (NeurIPS 2025 · similarity 0.89)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.88)
- Learning to Contextualize Web Pages for Enhanced Decision Making by LLM Agents (ICLR 2025 · similarity 0.88)
- AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning (NeurIPS 2024 · similarity 0.88)
- Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models (ICML 2024 · similarity 0.88)
- Can large language models explore in-context? (NeurIPS 2024 · 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.