EMNLP Demo Experiments

Of course, the question remains: is using recursive multi-agent systems actually better than just improving a single model? To test this, we ran three different benchmarks and compared the performance of various configurations of ReDel systems. The benchmarks we ran were:

• FanOutQA, a multi-hop, multi-document information seeking benchmark with open-domain search

• TravelPlanner, a real-world planning benchmark for language agents

• WebArena, an autonomous agent benchmark with diverse tasks in a realistic web environment

All of our experiment code is open-source on the demo/emnlp branch of the ReDel repository: https://github.com/zhudotexe/redel/tree/demo/emnlp

System Configurations

We tested the following system configurations:

System ID	Root Model	Delegate Model	Root Functions?	Delegation?	Root Context	Delegate Context
full	gpt-4o	gpt-4o	no	yes	128000	128000
root-fc	gpt-4o	gpt-4o	yes	yes	128000	128000
baseline	gpt-4o	N/A	yes	no	128000	N/A
small-leaf	gpt-4o	gpt-3.5-turbo	no	yes	128000	16385
small-all	gpt-3.5-turbo	gpt-3.5-turbo	no	yes	16385	16385
small-baseline	gpt-3.5-turbo	N/A	yes	no	16385	N/A
short-context	gpt-4o	gpt-4o	no	yes	8192	8192
short-baseline	gpt-4o	N/A	yes	no	8192	N/A

For each benchmark, we also provided the system with benchmark-specific tools (e.g. Wikipedia search for FanOutQA, database search for TravelPlanner, and browser click/type actions for WebArena).

Results

System ID	FanOutQA (GPTScore)	TravelPlanner (SR)	WebArena (SR)
full	0.494	2.778	0.203
root-fc	0.429	0.000	0.188
baseline	0.394	0.000	0.162
small-leaf	0.255	0.556	0.122
small-all	0.087	0.000	0.092
small-baseline	0.077	0.000	0.085
short-context	0.426	—	0.129
short-baseline	0.361	—	0.122

We find that overall, recursive delegation systems perform better than their baseline systems in every single benchmark.

From these high-level results, we can conclude that:

• Adding recursive delegation on top of a model improves its performance as tasks get harder

• To fully take advantage of delegation, it is often important to remove tools from the root node to avoid the root node overcommitting

• Recursive delegation allows you to use a “strong” model as the root and “weaker” models as delegates to gain a major performance boost over a system comprised entirely of a weaker model

• Recursive delegation can help in long-context tasks by removing the requirement for a single model to process everything in serial

We will release a deeper dive into the results in a future paper.

Reproducing Experiments

In the demo/emnlp branch of the ReDel repository, we include the logs of every single experiment run in the experiments/ directory. You can load any of these runs in the visualization to view what the ReDel system did!

The experiments directory is broken down into the following structure: experiments/BENCHMARK_NAME/BENCHMARK_SPLIT/[RUN_ID]/SYSTEM_ID/QUERY_ID, where:

• BENCHMARK_NAME is the name of the benchmark (fanoutqa, travelplanner, or webarena)

• BENCHMARK_SPLIT is the split of the benchmark we ran (usually the dev/validation split)

• RUN_ID is an internal split in the FanOutQA experiment to analyze an edge-case behaviour wrt parallel function calling and long contexts

• SYSTEM_ID is the system under test, configured as in the table below

• QUERY_ID is the benchmark-specific ID of a single run (loadable in the visualizer).

To reproduce the experiments included in this repository, we include scripts to run each benchmark.

Follow these steps to setup the environment, then follow the instructions in each benchmark. We recommend setting up a virtual environment for this project.

1. First, you’ll need to clone this repository and check out the demo/emnlp branch: git clone -b demo/emnlp https://github.com/zhudotexe/redel

2. Install the necessary dependencies: pip install -r requirements.txt

FanOutQA

output path: experiments/fanoutqa/dev/trial2/SYSTEM_ID

Run

python bench_fanoutqa.py <full|root-fc|baseline|small-leaf|small-all|small-baseline|short-context|short-baseline>

This will run the given system on the FanOutQA dev set in the Open Book setting.

Evaluate

Set the FANOUTQA_OPENAI_API_KEY environment variable to a valid OpenAI API key. You can use export FANOUTQA_OPENAI_API_KEY=$OPENAI_API_KEY to copy an existing API key from environment variables.

python score_fanoutqa.py experiments/fanoutqa/**/results.jsonl

This will output a score.json file in the output path with the final scores.

TravelPlanner

output path: experiments/travelplanner/validation/SYSTEM_ID

Setup

1. Install the TravelPlanner database:

   1. Download the database from this link

   2. Extract the zip file in redel/tools/travelplanner. This should create a directory named db.

2. In another directory, clone our fork of the TravelPlanner repository. This will be used for scoring, and includes the fixes discussed in our paper.

   1. git clone https://github.com/zhudotexe/TravelPlanner

Run

python bench_travelplanner.py <full|root-fc|baseline|small-leaf|small-all|small-baseline>

Note: This benchmark does not test the short-ctx systems since this benchmark doesn’t have a long-context requirement.

Evaluate

python score_travelplanner.py experiments/travelplanner/**/results.jsonl

This script will write files in the correct format for the TravelPlanner evaluation in the output path, and print the command to run to score the results.

You should now switch to the TravelPlanner repository you cloned in the setup step and run the commands output by this script.

WebArena

output path: experiments/webarena/test/SYSTEM_ID

Setup

We reproduce some of the scripts and data contained in the WebArena repository in this repo under the terms of the Apache-2.0 license, contained in experiments/webarena/vendor/LICENSE.

First, you’ll need to set up your own WebArena environment. See https://github.com/web-arena-x/webarena/blob/main/environment_docker/README.md for instructions.

Next, run the following to setup the webarena configuration:

# setup env vars (see https://github.com/web-arena-x/webarena/blob/main/environment_docker/README.md for env setup)
export SHOPPING="<your_shopping_site_domain>:7770"
export SHOPPING_ADMIN="<your_e_commerce_cms_domain>:7780/admin"
export REDDIT="<your_reddit_domain>:9999"
export GITLAB="<your_gitlab_domain>:8023"
export MAP="<your_map_domain>:3000"
export WIKIPEDIA="<your_wikipedia_domain>:8888/wikipedia_en_all_maxi_2022-05/A/User:The_other_Kiwix_guy/Landing"
export HOMEPAGE="<your_homepage_domain>:4399"
# generate config files
python experiments/webarena/generate_test_data.py

You’ll also need to ensure Playwright is installed:

playwright install chromium

Run

First, make sure you have reset your WebArena environment (see https://github.com/web-arena-x/webarena/blob/main/environment_docker/README.md#environment-reset).

Then, launch the WebArena environment.

As the default WebArena script is incompatible with asyncio, ReDel launches a separate process to handle the WebArena environment, which it communicates with over a pipe. This is done automatically.

Finally, run the bench script:

python bench_webarena.py <full|root-fc|baseline|small-leaf|small-all|small-baseline|short-context|short-baseline>