-
Notifications
You must be signed in to change notification settings - Fork 8
/
index.html
251 lines (227 loc) · 14.7 KB
/
index.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="DiscoveryWorld: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents">
<meta name="keywords" content="DiscoveryWorld, Discovery World, AI2, Aristo">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>DiscoveryWorld - Allen Institute for AI</title>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/ai2_website_top.png">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-2 publication-title"><strong>DiscoveryWorld:</strong> A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://cognitiveai.org/">Peter Jansen</a><sup>1,2</sup>,
</span>
<span class="author-block">
<a href="https://www.microsoft.com/en-us/research/people/macote/">Marc-Alexandre Côté</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://allenai.org/team/tushark">Tushar Khot</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://allenai.org/team">Erin Bransom</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.majumderb.com/">Bodhisattwa Prasad Majumder</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://bhavanadalvi.github.io/">Bhavana Dalvi Mishra</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://allenai.org/team/oyvindt">Oyvind Tafjord</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://allenai.org/team/peterc">Peter Clark</a><sup>1</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Allen Institute for AI</span>,
<span class="author-block"><sup>2</sup>University of Arizona</span>,
<span class="author-block"><sup>3</sup>Microsoft Research</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2406.06769"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<!-- <span class="link-block">
<a href=""
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span> -->
<!-- Video Link. -->
<!-- <span class="link-block">
<a href=""
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-youtube"></i>
</span>
<span>Video</span>
</a>
</span> -->
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/allenai/discoveryworld"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- Dataset Link. -->
<!-- <span class="link-block">
<a href=""
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="far fa-images"></i>
</span>
<span>Data</span>
</a> -->
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<center>
<iframe width="640" height="385" src="https://www.youtube.com/embed/hKWd-pwF0_E?si=nswvkTZZ1ScIXMa_" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>
</center>
<h2 class="subtitle has-text-centered">
An example playthrough of one of DiscoveryWorld's shortest tasks, the Proteomics task.
</h2>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h3 class="title is-4">Abstract</h3>
<div class="content has-text-justified">
<p>
Automated scientific discovery promises to accelerate progress across scientific domains. However, developing and evaluating an AI agent's capacity for end-to-end scientific reasoning is challenging as running real-world experiments is often prohibitively expensive or infeasible. In this work we introduce DiscoveryWorld, the first virtual environment for developing and benchmarking an agent's ability to perform complete cycles of novel scientific discovery. DiscoveryWorld contains a variety of different challenges, covering topics as diverse as radioisotope dating, rocket science, and proteomics, to encourage development of general discovery skills rather than task-specific solutions. DiscoveryWorld itself is an inexpensive, simulated, text-based environment (with optional 2D visual overlay). It includes 120 different challenge tasks, spanning eight topics each with three levels of difficulty and several parametric variations. Each task requires an agent to form hypotheses, design and run experiments, analyze results, and act on conclusions. DiscoveryWorld further provides three automatic metrics for evaluating performance, based on (a) task completion, (b) task-relevant actions taken, and (c) the discovered explanatory knowledge. We find that strong baseline agents, that perform well in prior published environments, struggle on most DiscoveryWorld tasks, suggesting that DiscoveryWorld captures some of the novel challenges of discovery, and thus that DiscoveryWorld may help accelerate near-term development and assessment of scientific discovery competency in agents.
</p>
<p>
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<!-- <h2 class="title is-3">Abstract</h2> -->
<div class="content has-text-justified">
<h3 class="title is-4">Benchmark Environments</h3>
<img src="./static/images/discoveryworld-logo.png">
<p>
As a benchmark, DiscoveryWorld contains 120 different challenge tasks, organized as parametric variations of 8 discovery topics spanning 3 levels of difficulty. The discovery topics come from diverse areas of science, such as discovering the cause of an illness, validating radioisotope dating, or adapting rocket science to a new planet. Each parametric variation of a given task changes the underlying data, task solution, and often times also reorgnizes the environment, creating new discovery tasks for agents to solve.
</p>
<h3 class="title is-4">End-to-end Discovery</h3>
<p>
<strong>DiscoveryWorld demands all the key facets in end-to-end scientific discovery,</strong> from ideation, forming hypotheses, planning and executing experiments, and drawing conclusions to solve challenge tasks developed for this virutal world. Some hypotheses may lead to dead ends that don't help solve the task, while others uncover helpful knowledge that leads to making descriptive discoveries, or critical insights that lead to making explanatory discoveries.
</p>
<img src="./static/images/discovery-figure.png" display="block" style="max-width: 100%; align-items: center; margin: auto;">
<h3 class="title is-4">Detailed Discovery Scorecards</h3>
<p>
Measuring progress on scientific discovery is challenging. DiscoveryWorld allows measuring three kinds of discovery progress automatically: (1) Task completion, or whether the challenge task was solved correctly, (2) Procedural progress, or whether an agent took actions that look like what a scientist might perform (like exploring locations, taking samples, measuring those samples, etc.), and (3) Explanatory knowledge, or whether the agent uncovered critical insights -- like the cause of an illness, and how to cure it, or that to get a quantum reactor to work correctly, there is a specific mathematical relationship between physical measurements of its components that needs to be tuned.
</p>
<img src="./static/images/discovery-scorecard.png" display="block" style="max-width: 80%; align-items: center; margin: auto;">
<h3 class="title is-4">DiscoveryWorld is hard for language models, but solvable for professional human scientists</h3>
<p>
<strong>Professional human scientists with Masters or PhD degrees in the natural sciences (such as physics, biology, chemistry, and geology) do well at DiscoveryWorld.</strong> All tasks are solvable by scientists with specific backgrounds, and some scientists even solved nearly all the DiscoveryWorld tasks!
</p>
<img src="./static/images/discoveryworld-human-performance.png" display="block" style="max-width: 100%; align-items: center; margin: auto;">
<br>
<br>
<p>
Conversely, <strong>DiscoveryWorld is very challenging for language models.</strong> The best baseline models successfully complete less than 20% of the discovery tasks on Normal and Challenge difficulty, and frequently don't discover critical knowledge that would explain the discovery.
</p>
<img src="./static/images/discoveryworld-agent-performance.png" display="block" style="max-width: 100%; align-items: center; margin: auto;">
<br>
<br>
<h3 class="title is-4">You can play DiscoveryWorld, too!</h3>
<p>
<strong>DiscoveryWorld is designed for both humans and agents to play!</strong> Humans can make use of the graphical user interface, while agents can make use of text (JSON), visual (2D) renderings, or both. It generally takes only a few mintutes to install DiscoveryWorld, where istructions can be found in the <a href="https://github.com/allenai/discoveryworld">DiscoveryWorld GitHub Repository</a>. If you play yourself, it's recommended you not read about the tasks in the paper before hand, as the paper contains spoilers!
</p>
<h3 class="title is-4">Related Readings</h3>
<p>
DiscoveryWorld joins an increasing number of virtual environments that help develop and benchmark an agent's reasoning abilities. While several real-world discovery systems have shown success in areas such as Chemistry (e.g. <a href="https://www.nature.com/articles/s41586-023-06792-0">CoScientist</a>) and Protenomics (e.g. <a href="https://deepmind.google/technologies/alphafold/">AlphaFold</a>), DiscoveryWorld is the first virtual environment with highly varied discovery tasks across a variety of different scientific disciplines, and designed to test <strong>general discovery skills</strong>.
</p>
<img src="./static/images/discoveryworld-comparison.png" display="block" style="max-width: 100%; align-items: center; margin: auto;">
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{jansen2024discoveryworld,
author = "Peter Jansen and Marc-Alexandre Côté and Tushar Khot and Erin Bransom and Bhavana Dalvi Mishra and Bodhisattwa Prasad Majumder and Oyvind Tafjord and Peter Clark",
title = "DISCOVERYWORLD: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents",
journal = "arXiv",
year = "2024",
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="publication-links">
<!-- <div class="column is-8"> -->
<span class="link-block">
<img src="./static/images/ai2-logo-header.png" display="block" style="max-width: 20%; align-items: center; margin: auto;">
</span>
<br>
<span class="link-block">
<img src="./static/images/aristo-logo-header.png" display="block" style="max-width: 15%; align-items: center; margin: auto;">
</span>
</div>
<!-- </div> -->
</div>
<br>
<p><center><a href="https://allenai.org/">Allen Institute for AI</a> - all rights reserved.<br>
The website template is borrowed from <a href="https://nerfies.github.io/">here</a>.</center></p>
</div>
</div>
</footer>
</body>
</html>