-
Notifications
You must be signed in to change notification settings - Fork 8
/
index.html
567 lines (486 loc) · 30.6 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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
<!DOCTYPE html>
<html>
<head>
<title>MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback</title>
<link rel="icon" href="website/img/mint-leaf-logo.png" type="image/icon type">
<meta name="viewport" content="width=device-width, initial-scale=1">
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/chart.umd.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/[email protected]"></script>
<script
src="https://cdn.jsdelivr.net/npm/[email protected]/dist/chartjs-plugin-annotation.min.js"></script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="website/css/bulma.min.css">
<link rel="stylesheet" href="website/css/bulma-carousel.min.css">
<link rel="stylesheet" href="website/css/bulma-slider.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="./website/javascript/bulma-carousel.min.js"></script>
<script src="./website/javascript/bulma-slider.min.js"></script>
<link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" rel="stylesheet"
integrity="sha384-1BmE4kWBq78iYhFldvKuhfTAU6auU8tT94WrHftjDbrCEXSU1oBoqyl2QvZ6jIW3" crossorigin="anonymous">
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.bundle.min.js"
integrity="sha384-ka7Sk0Gln4gmtz2MlQnikT1wXgYsOg+OMhuP+IlRH9sENBO0LRn5q+8nbTov4+1p"
crossorigin="anonymous"></script>
<link href="https://unpkg.com/[email protected]/dist/css/tabulator_bootstrap4.min.css" rel="stylesheet">
<script type="text/javascript" src="https://unpkg.com/[email protected]/dist/js/tabulator.min.js"></script>
<!-- <script src="website/javascript/peity-vanilla.js"></script> -->
<script src="website/javascript/benchmark_table.js" type="module"></script>
<script src="website/javascript/success_rate_vs_k_vis.js" type="module"></script>
<script src="website/javascript/feedback_success_rate_vis.js" type="module"></script>
<script src="website/javascript/feedback_provider_efficacy.js" type="module"></script>
<link rel="stylesheet" href="website/css/index.css">
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-C7GJ4FYMY9"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() { dataLayer.push(arguments); }
gtag('js', new Date());
gtag('config', 'G-C7GJ4FYMY9');
</script>
<noscript>
<p><img alt="Clicky" width="1" height="1" src="//in.getclicky.com/101339888ns.gif" /></p>
</noscript>
</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 publication-title">
<img src="website/img/mint-leaf-logo.png" alt="logo" width="40" height="40" />
MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://xingyaoww.github.io">Xingyao Wang</a><sup>1*</sup>,
</span>
<span class="author-block">
<a href="https://zihanwang314.github.io/">Zihan Wang</a><sup>2*</sup>,
</span>
<span class="author-block">
<a href="https://lumos-jiateng.github.io/">Jiateng Liu</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://yangyi-chen.github.io/">Yangyi Chen</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://lifan-yuan.github.io/">Lifan Yuan</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://haopeng-nlp.github.io/">Hao Peng</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://blender.cs.illinois.edu/hengji.html">Heng Ji</a><sup>1</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>University of Illinois Urbana-Champaign,</span>
<span class="author-block"><sup>2</sup>Renmin
University of China</span>
<br><span>To appear at ICLR 2024</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2309.10691" class="btn btn-outline-dark"
role="button">📝
Paper</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/xingyaoww/mint-bench" class="btn btn-outline-dark"
role="button">💻
Code</a>
</span>
<!-- Dataset Link. -->
<span class="link-block">
<a href="https://github.com/xingyaoww/mint-bench/blob/main/docs/DATA.md"
class="btn btn-outline-dark" role="button">📂
Data</a>
</div>
</div>
<!-- <h2 class="subtitle" style="text-align: left;">
<b>MINT benchmark</b> measures LLMs' ability to solve tasks with multi-turn interactions
by
(1) using tools and (2) leveraging natural language feedback.
</h2> -->
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<h2 class="subtitle">
<b>MINT benchmark</b> measures LLMs' ability to solve tasks with multi-turn interactions
by
(1) using tools and (2) leveraging natural language feedback.
</h2>
<ul class="nav nav-tabs" id="myTab" role="tablist">
<li class="nav-item" role="presentation">
<button class="nav-link active" id="main-results-tab" data-bs-toggle="tab"
data-bs-target="#benchmark-table-content" type="button" role="tab"
aria-controls="main-results-tab" aria-selected="true">Micro Average</button>
</li>
<li class="nav-item" role="presentation">
<button class="nav-link" id="eurus-code-table-tab" data-bs-toggle="tab"
data-bs-target="#eurus-code-table-content" type="button" role="tab"
aria-controls="eurus-code-table-tab" aria-selected="false">Code (Eurus subset)</button>
</li>
<li class="nav-item" role="presentation">
<button class="nav-link" id="eurus-math-table-tab" data-bs-toggle="tab"
data-bs-target="#eurus-math-table-content" type="button" role="tab"
aria-controls="eurus-math-table-tab" aria-selected="false">
Math (Eurus subset)</button>
</li>
</ul>
<div class="tab-content" id="myTabContent">
<div class="tab-pane fade show active" id="benchmark-table-content" role="tabpanel"
aria-labelledby="benchmark-table-content">
<p class="mt-2 px-2">
This table contains the micro average across all task instances originally featured in the
<a href="https://arxiv.org/abs/2309.10691">MINT paper</a>. It includes test instances from
several sources: HumanEval, MBPP, GSM8K, HotpotQA, MATH, MMLU, TheoremQA, and AlfWorld.
</p>
<div id="benchmark-table"></div>
</div>
<div class="tab-pane fade" id="eurus-code-table-content" role="tabpanel"
aria-labelledby="eurus-code-table-content">
<p class="mt-2 px-2">
This code subset follows the <a href="https://arxiv.org/abs/2404.02078">Eurus
paper</a> and contains MBPP and HumanEval.
</p>
<div id="eurus-code-table"></div>
</div>
<div class="tab-pane fade" id="eurus-math-table-content" role="tabpanel"
aria-labelledby="eurus-math-table-content">
<p class="mt-2 px-2">
This math subset follows the <a href="https://arxiv.org/abs/2404.02078">Eurus
paper</a> and contains TheoremQA, MATH and MMLU.
</p>
<div id="eurus-math-table"></div>
</div>
</div>
<br>
<h2 class="subtitle">
<b>MINT</b> can measure different LLMs' ability to provide natural language feedback by measuring
the benefit of their feedback (Δ Success Rate) to a fixed LLM (gpt-3.5-turbo-0613).
</h2>
<div id="benchmark-feedback-efficancy-table"></div>
<br>
<h2 class="subtitle">
Please refer to our <a href="https://github.com/xingyaoww/mint-bench">GitHub repo</a> to add your
model to the leaderboard.
</h2>
</div>
</div>
</section>
<section class="section" id="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">
<p>
To solve complex tasks, large language models (LLMs) often require multiple rounds of
interactions with the user, sometimes assisted by external tools.
However, current evaluation protocols often emphasize benchmark performance with single-turn
exchanges, neglecting the nuanced interactions among the user, LLMs, and external tools,
while also underestimating the importance of natural language feedback from users. These
oversights contribute to discrepancies between research benchmark evaluations and real-world
use cases.
We introduce MINT, a benchmark that evaluates LLMs' ability to solve tasks with multi-turn
interactions by (1) using tools and (2) leveraging natural language feedback.
To ensure reproducibility, we provide an evaluation framework where LLMs can access tools by
executing Python code and receive users' natural language feedback simulated by GPT-4.
We repurpose a diverse set of established evaluation datasets focusing on reasoning, coding,
and decision-making and carefully curate them into a compact subset for efficient
evaluation.
<br>
Our analysis of 20 open- and closed-source LLMs offers intriguing findings.
</p>
<ul>
<li>(a) LLMs generally benefit from tools and language feedback, with performance gains
(absolute, same below) of 1-8% for each turn of tool use and 2-17% with natural language
feedback.</li>
<li>(b) Better single-turn performance does not guarantee better multi-turn performance.
</li>
<li>(c) Surprisingly, on the LLMs evaluated, supervised instruction-finetuning (SIFT) and
reinforcement learning from human feedback (RLHF) generally hurt multi-turn
capabilities.</li>
</ul>
<p>
We expect MINT can help measure progress and incentivize research in improving LLMs'
capabilities in multi-turn interactions, especially for open-source communities where
multi-turn human evaluation can be less accessible compared to commercial LLMs with a larger
user base.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<section class="section" id="interaction-framework">
<div class="container is-max-desktop">
<div class="columns is-full-width">
<!-- Visual Effects. -->
<div class="column">
<div class="content">
<h2 class="title is-3">Interaction Framework</h2>
<p>
MINT mirrors the real-world User-LLM-Tool collaborative problem-solving setting. To solve a
problem,
the
LLM can (1) use external tools by generating and executing Python programs and/or (2)
collecting
natural
language feedback to refine its solutions; the feedback is provided by GPT-4, aiming to
simulate
human
users in a reproducible and scalable way.
</p>
<ul>
<li>We measure LLMs' <b>tool-augmented task-solving capability</b> by analyzing its
performance gain
with increased numbers of turns without language feedback (i.e., no red dotted box in
the figure
below).
</li>
<li>
We quantify LLMs' <b>ability to leverage natural language feedback</b> with the
performance gain
upon receiving GPT-4 generated feedback (i.e., performance without and with red dotted
box in
the
figure below).
</li>
</ul>
<div style="text-align:center;">
<img src="website/img/illustrative-example.jpg" alt="illustrative-example"
style="margin: 0 auto; display: block; max-width: 1000px; width: 100%; height: auto;" />
<br>
</div>
</div>
</div>
<!--/ Visual Effects. -->
</div>
</section>
<section class="section" id="evaluation">
<div class="container is-max-desktop">
<div class="columns is-full-width">
<!-- Visual Effects. -->
<div class="column">
<div class="content">
<h2 class="title is-3">Evaluation</h2>
<p>
We evaluate 20 LLMs where 4 are closed- and 16 are open-source.
We cover different sizes and training techniques to better understand how they affect LLMs'
multi-turn
interaction capability. We consider three variants of training techniques:
</p>
<ul>
<li>Base: Pre-trained model</li>
<li>SIFT: Supervised Instruction-Finetuning</li>
<li>RLHF: Reinforcement Learning from Human Feedback</li>
</ul>
<h3>Tool-augmented Task-Solving capabilities of LLMs</h3>
<div class="text-justify" id="tool-augmented">
<ul>
<li>
We find all open-source models fall behind most commercial closed-source models in
both success
rate
at k=5 and improvement rate (slope).
<br>
<button class="btn btn-outline-secondary btn-sm"
id="visualize-sr-vs-k-open-behind-close">Visualize
This</button>
</li>
<li>
Absolute performance and improvement-per-turn (e.g., slope) scale with model size.
<br>
<div class="btn-group" role="group">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-scale-with-model-size-llama2-base">Visualize: LLaMA-2
Base</button>
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-scale-with-model-size-llama2-rlhf">LLaMA-2
RLHF</button>
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-scale-with-model-size-codellama-base">CodeLLaMA
Base</button>
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-scale-with-model-size-codellama-sift">CodeLLaMA
SIFT</button>
</div>
</li>
<li>
SIFT on multi-turn data can potentially be helpful. <a
href="https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md">Vicuna-v1.5
(7B)</a>, which is a SIFT variant of LLaMA2 trained on ShareGPT conversations
(most are multi-turn), exhibit stronger performance compared to LLaMA-2 (Base and
RLHF)<sup><a href="#footnote-1" id="ref-footnote-1">1</a></sup>.
We observe similar trend for <a
href="https://github.com/OpenLemur/Lemur">Lemur-70b-chat-v1</a>, which continue
pre-train LLaMA-2 (70B) on code intensive data followed by SIFT on multi-turn data.
<br>
<div class="btn-group" role="group">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-vicuna-better-than-llama">Visualize: Vicuna-v1.5
(7B)</button>
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-lemur-better-than-llama">Lemur-v1 (70B)</button>
</div>
</li>
<li>
We find RLHF hurt LLM-tool multi-turn interaction on LLaMA-2 series. However, it's
unclear if RLHF is problematic overall, or if the issue only arise when RLHF is
primarily applied to
single-turn data.
<br>
<button class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-sr-vs-k-rlhf">Visualize This</button>
</li>
</ul>
<ol>
<li style="font-size: 0.8rem;" id="footnote-1">We find some performance degradation in
Vicuna-v1.5
(especially for the 13B one), potential due to training artifacts. We refer to paper
Section 3.5
for
more details.</li>
</ol>
</div>
<button class="btn btn-outline-secondary btn-sm" id="visualize-sr-vs-k-all">Visualize All
Models</button>
<div class="chart-container" id="chart-k" style="display:block;margin:0 auto;">
<canvas id="chart-sr-vs-k"></canvas>
</div>
<h3>LLMs' Ability to Leverage Natural Language Feedback</h3>
<ul>
<li>
We find no significant difference between open- and closed-source models in terms of
Δfeedback.
<br>
<button class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-feedback-sr-no-diff-open-close">Visualize
This</button>
</li>
<li>
Similar to previous findings, we find that SIFT and RLHF hurt models' ability to
leverage feedback on CodeLLama (except 7B) and LLaMA-2, as they all have lower
Δfeedback and Success Rate (with feedback) compared to their base variants.
Another two exceptions are Vicuna and Lemur-v1; We speculate using multi-turn
conversations (ShareGPT) for SIFT contributes to these two exceptions.
<br>
<button class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-feedback-sr-sift-rlhf">Visualize
This</button>
</li>
<li>
Models hardly benefit from self-feedback. We find GPT-4-0613 using self-generated
feedback has
limited benefit: only decision-making has improved slightly.
<br>
<button class="btn btn-outline-secondary btn-sm inline-vis-button"
id="visualize-feedback-sr-gpt-4-self">Visualize
This</button>
</li>
</ul>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm" disabled>Choose task type
to
visualize:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="avg_micro">Micro
Average</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="reasoning">Reasoning</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="decision_making">Decision-Making</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="code_generation">Code</button>
</div>
<div class="btn-group btn-group-toggle text-center sort-by-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm" disabled>Sort
by:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="sort-by-feedbacksr">Success
Rate with GPT-4 Feedback</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="sort-by-nofeedbacksr">Without
Feedback</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="sort-by-feedbackdelta">Δ
Feedback</button>
</div>
</div>
<div class="chart-container" id="chart-feedback" style="position:relative;margin:0 auto;">
<canvas id="chart-sr-w-feedback" style="max-height: 100%;"></canvas>
</div>
<h3>LLMs' Ability to Provide Natural Language Feedback</h3>
<p>
In this section, we fixed the evaluated LLM (gpt-3.5-turbo-0613) and use different
LLMs to
<b>provide</b> language feedback.
This allows us to measure different LLMs' effectiveness in providing feedback.
<br>
We find that task-solving ability could be orthogonal to feedback-providing ability: LLM's
higher task-solving performance does not necessarily translate to better feedback-providing
capability and vice versa.
For example, despite performing the worst in solving tasks, CodeLLaMA (34B, SIFT) can
provide feedback that improves the stronger GPT-3.5.
</p>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center feedback-provider-sort-by-selector"
data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm" disabled>Sort
by:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="sort-by-feedback-gain">
Success Rate with Feedback
</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="sort-by-feedback-provider-perf">
Feedback Provider's Performance
</button>
</div>
</div>
<div class="chart-container" id="chart-feedback-p" style="display:block;margin:0 auto;">
<canvas id="chart-feedback-provider"></canvas>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@misc{wang2023mint,
title={MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback},
author={Xingyao Wang and Zihan Wang and Jiateng Liu and Yangyi Chen and Lifan Yuan and Hao Peng and Heng Ji},
year={2023},
eprint={2309.10691},
archivePrefix={arXiv},
primaryClass={cs.CL}
}</code></pre>
</div>
</section>
<footer class="footer">
<div align="center" class="container">
<div class="columns is-centered">
<div class="content is-small">
This website templated is borrowed from <a
href="https://github.com/nerfies/nerfies.github.io">nerfies</a>.
</div>
</div>
</div>
</footer>
</body>
</html>