Skip to content

Latest commit

 

History

History
175 lines (127 loc) · 11.6 KB

README.md

File metadata and controls

175 lines (127 loc) · 11.6 KB
Raw

Agents under Unified Alignment for Agents (UA$^2$) Framework

🚀 This repo implements the preliminary version of agents designed under Unified Alignment for Agents (UA$^2$) framework with results benchmarked on UA$^2$-Webshop.
🚀 The project is a practice of LLM-powered agent framework design under the guidance of Towards Unified Alignment Between Agents, Humans, and Environment.

If you find this repo useful, please cite our project:

@article{yang2024towards,
   title = {Towards Unified Alignment Between Agents, Humans, and Environment},
   author = {Yang, Zonghan and Liu, An and Liu, Zijun and Liu, Kaiming and Xiong, Fangzhou and Wang, Yile and Yang, Zeyuan and Hu, Qingyuan and Chen, Xinrui and Zhang, Zhenhe and Luo, Fuwen and Guo, Zhicheng and Li, Peng and Liu, Yang},
   journal={arXiv preprint arXiv:2402.07744},
   year = {2024}
}

Performance Comparison on UA$^2$-Webshop

Plot

Caption The performance of averaged reward, success rate (SR) (%), alignment gap (%) with human intentions (G_{HI}) and environment dynamics (G_{ED}), time (s) and money ($) cost of all methods tested in our retrofitted WebShop environment. The best result for each metric is in bold. The better performance under each metric is indicated by the darker green shades. *LATS is tested on 1/10 subset of the entire task instructions due to the significant cost.

$$$$

🎡 Methods used for comparison:

  • ReAct: On each step, the agent decides whether to take actions or reason about the query based on the current state.
  • ReACT-SC: Instead of Chain-of-Thought reasoning, we apply self-consistency strategy to ReAct to further improve the performance.
  • Reflexion: Using verbal feedback as the reflecion signal for agents. In this case, we follow the external reward as the feedback directly from the environment, following its original setting.
  • LATS: An advanced method that unifies ReAct, self-reflection, and tree search based planning. We adapt the original implementation in our scenario for comparison.
  • Ours: the UA$^2$-Agent framework, which is introduced in the below section.

Introduction

🌐 UA$^2$-Webshop

To be specific, the key contribution of our UA$^2$ principles to the LLM-powered agent research community is the concept of alignment rooted in the agent-assisting process. In UA$^2$-Webshop, the requirements of alignment originates from different sources. To name a few:

  • Human Intentions (HI): Whether the authentic goals need to be inferred during task execution, or the intentions of humans are precisely conveyed in the descriptions.
  • Environmental Dynamics (ED): Whether the state transitions of the environment are intrinsically endowed with partial observability, temporality, or stochasticity.
  • Self-Constraints: Whether the status of budgetary resources is reflected, including time consumption, the maximum number of actions or reasoning steps, etc.

Kindly refer to the online article for detailed depiction on how we introduce those requirements. The live site demo can be found here, as well as the environment repo for local deployment purpose.

These requirements are already reflected by the task and the website design of UA$^2$-Webshop. In this repo, we wrap the environment with runtime information (time and money) to compute the resource consumptions of an LLM-powered agent.

🤖 UA$^2$-Agent Framework

The key challenge is to build an agent framework that manages to assist the decision process in a realistic environment, in consideration of different principles of alignment. We leverage a structured memory with low-level insights to make better decisions upon ReACT agents.

Profiler

Caption The details of our agent design that follows the principles of UA^{2}. Compared to traditional ReACT agents, we append structured experience as the long-term memory: By filtering and analyzing raw trajectories, we extracted key actions from prior successes as low-level insights in reasoning/action paths. By retrieving reference low-level insights under the same user, we can find the high-level experience under most similar user instructions, expressing similar human intentions. Agents are able to understand human intentions and environment dynamics by extrapolating key actions from a similar, prior task.

$$$$

Low-level action insights are a list of key actions solicited from different runs in the environment under the same task instruction. The key actions are extracted from the high-reward trajectories with an analyzer, with which the contributions of actions are computed in the task-solving process.

💡 Here are the ways we follow the UA$^2$ principles:

  • Human Intentions: quickly adapting to the user's preference by retrieving trajectories with high rewards beforehand as the high-level experience. Note that the structured memory only stores the experience of the same user.
  • Environment Dynamics: directly transfer the key actions as short-cuts in interaction / reasoning process that lead to the success under ever-changing environment.
  • Self-Constraints: retrieving the trajectory of the most similar instruction before as a reference from structured memory. The agent directlly extrapolating the experience, instead of planning from scratch with LLMs or searching in huge memory / experience space.

Two major modules:

  • Analyzer part:

    • Infer the impact of the reference retrieved from the structured Memory to better complete the current task
    • Recognize key actions of the current trajectory via reflection in a single LLM call

  • Memory part:

    • Experimence accumulation with a specific user

    • Structured representation for instruction relations: references points to the actual trajectory of a task

    • Using semantic similarities for inter-task retrieval

    • Direct action extrapolation from the key action list

      (For better efficiency, we only utilize the best matched profile as a reference in decision making)

What is the function of different directories?

  • ./ua2-agent: the core of our UA$^2$-Agent framework

    • Insight.py: the implementation of Analyzer part
    • Profiler.py: the implementation of Memory part
    • react_w_insights_w_profiler_v1benchmark.py: the implementation of our UA$^2$-Agent algorithm on UA$^2$-Webshop benchmark

  • ./environments: running environment of our UA$^2$-Webshop benchmark (the encapsulation of our core environment)

    • env_instr_list_ua2webshop_runtime_session.py: the capsule of run-time environment leveraging cost information for UA$^2$-Webshop benchmark

  • ./baselines: source code of baselines

    • README.md: the instruction of how to run baselines and implementation details

How to replicate experiments of baselines and ours?

Prepare for the conda environment:

conda create -n ua2
pip install -r requirements.txt
conda activate ua2

Add your OpenAI API key to your environment:

# on Linux/Mac
export OPENAI_API_KEY=<YOUR_API_KEY>
# on Windows
set OPENAI_API_KEY=<YOUR_API_KEY>

For Reflexion, ReAct-series and CoT-series baselines, change your working directory and run the corresponding script directly:

cd baselines
python cot_least_to_most.py
python cot_sc.py
python react.py
python react_sc.py

For LATS baselines:

cd baselines/lats
mkdir runtime_logs
./lats.sh

For Reflexion:

cd baselines/reflexion
./reflexion.sh

To test our method:

cd code
python react_w_insights_w_profiler_v1benchmark.py

After running the script, the results can be found in the directory ./runtime_logs. More details can be found in ./baselines/README.md.

Authors

The **UA**$^2$-Agent project is a collaborative effort of Logo AgentForce Team. The method design, implementation and the testing of baseline methods in the context of the **UA**$^2$-Webshop is initiated and co-led by Zijun Liu (liuzijun20@mails.tsinghua.edu.cn) and An Liu (la22@mails.tsinghua.edu.cn). The following members are listed with main contributions:

This project is advised by Peng Li (lipeng@air.tsinghua.edu.cn) and Yang Liu (liuyang2011@tsinghua.edu.cn).

Contributions

We look forward to all kinds of suggestions from anyone interested in our project with whatever backgrounds! Either PRs, issues, or leaving a message is welcomed. We'll be sure to follow up shortly!