Grokking vs. Learning: Same Features, Different Encodings Article Swipe

PDF

Dmitry Manning-Coe , Jacopo Gliozzi , Alexander G. Stapleton , Edward Hirst , Giuseppe De Tomasi , Barry Bradlyn , David S. Berman ·

YOU? · · 2025 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2502.01739

Grokking typically achieves similar loss to ordinary, "steady", learning. We ask whether these different learning paths - grokking versus ordinary training - lead to fundamental differences in the learned models. To do so we compare the features, compressibility, and learning dynamics of models trained via each path in two tasks. We find that grokked and steadily trained models learn the same features, but there can be large differences in the efficiency with which these features are encoded. In particular, we find a novel "compressive regime" of steady training in which there emerges a linear trade-off between model loss and compressibility, and which is absent in grokking. In this regime, we can achieve compression factors 25x times the base model, and 5x times the compression achieved in grokking. We then track how model features and compressibility develop through training. We show that model development in grokking is task-dependent, and that peak compressibility is achieved immediately after the grokking plateau. Finally, novel information-geometric measures are introduced which demonstrate that models undergoing grokking follow a straight path in information space.

Related Topics

Computer Science

Artificial Intelligence

Concepts

Computer science Artificial intelligence

Metadata

Type: preprint
Language: en
Landing Page: http://arxiv.org/abs/2502.01739
PDF: https://arxiv.org/pdf/2502.01739
OA Status: green
Related Works: 10
OpenAlex ID: https://openalex.org/W4407184786

All OpenAlex metadata

Raw OpenAlex JSON

OpenAlex ID: https://openalex.org/W4407184786

Canonical identifier for this work in OpenAlex
DOI: https://doi.org/10.48550/arxiv.2502.01739

Digital Object Identifier
Title: Grokking vs. Learning: Same Features, Different Encodings

Work title
Type: preprint

OpenAlex work type
Language: en

Primary language
Publication year: 2025

Year of publication
Publication date: 2025-02-03

Full publication date if available
Authors: Dmitry Manning-Coe, Jacopo Gliozzi, Alexander G. Stapleton, Edward Hirst, Giuseppe De Tomasi, Barry Bradlyn, David S. Berman

List of authors in order
Landing page: https://arxiv.org/abs/2502.01739

Publisher landing page
PDF URL: https://arxiv.org/pdf/2502.01739

Direct link to full text PDF
Open access: Yes

Whether a free full text is available
OA status: green

Open access status per OpenAlex
OA URL: https://arxiv.org/pdf/2502.01739

Direct OA link when available
Concepts: Computer science, Artificial intelligence

Top concepts (fields/topics) attached by OpenAlex
Cited by: 0

Total citation count in OpenAlex
Related works (count): 10

Other works algorithmically related by OpenAlex

Full payload

id	https://openalex.org/W4407184786
doi	https://doi.org/10.48550/arxiv.2502.01739
ids.doi	https://doi.org/10.48550/arxiv.2502.01739
ids.openalex	https://openalex.org/W4407184786
fwci
type	preprint
title	Grokking vs. Learning: Same Features, Different Encodings
biblio.issue
biblio.volume
biblio.last_page
biblio.first_page
is_xpac	False
apc_list
apc_paid
concepts[0].id	https://openalex.org/C41008148
concepts[0].level	0
concepts[0].score	0.5496460199356079
concepts[0].wikidata	https://www.wikidata.org/wiki/Q21198
concepts[0].display_name	Computer science
concepts[1].id	https://openalex.org/C154945302
concepts[1].level	1
concepts[1].score	0.4064268469810486
concepts[1].wikidata	https://www.wikidata.org/wiki/Q11660
concepts[1].display_name	Artificial intelligence
keywords[0].id	https://openalex.org/keywords/computer-science
keywords[0].score	0.5496460199356079
keywords[0].display_name	Computer science
keywords[1].id	https://openalex.org/keywords/artificial-intelligence
keywords[1].score	0.4064268469810486
keywords[1].display_name	Artificial intelligence
language	en
locations[0].id	pmh:oai:arXiv.org:2502.01739
locations[0].is_oa	True
locations[0].source.id	https://openalex.org/S4306400194
locations[0].source.issn
locations[0].source.type	repository
locations[0].source.is_oa	True
locations[0].source.issn_l
locations[0].source.is_core	False
locations[0].source.is_in_doaj	False
locations[0].source.display_name	arXiv (Cornell University)
locations[0].source.host_organization	https://openalex.org/I205783295
locations[0].source.host_organization_name	Cornell University
locations[0].source.host_organization_lineage	https://openalex.org/I205783295
locations[0].license
locations[0].pdf_url	https://arxiv.org/pdf/2502.01739
locations[0].version	submittedVersion
locations[0].raw_type	text
locations[0].license_id
locations[0].is_accepted	False
locations[0].is_published	False
locations[0].raw_source_name
locations[0].landing_page_url	http://arxiv.org/abs/2502.01739
locations[1].id	doi:10.48550/arxiv.2502.01739
locations[1].is_oa	True
locations[1].source.id	https://openalex.org/S4306400194
locations[1].source.issn
locations[1].source.type	repository
locations[1].source.is_oa	True
locations[1].source.issn_l
locations[1].source.is_core	False
locations[1].source.is_in_doaj	False
locations[1].source.display_name	arXiv (Cornell University)
locations[1].source.host_organization	https://openalex.org/I205783295
locations[1].source.host_organization_name	Cornell University
locations[1].source.host_organization_lineage	https://openalex.org/I205783295
locations[1].license	cc-by
locations[1].pdf_url
locations[1].version
locations[1].raw_type	article
locations[1].license_id	https://openalex.org/licenses/cc-by
locations[1].is_accepted	False
locations[1].is_published
locations[1].raw_source_name
locations[1].landing_page_url	https://doi.org/10.48550/arxiv.2502.01739
indexed_in	arxiv, datacite
authorships[0].author.id	https://openalex.org/A5067312862
authorships[0].author.orcid
authorships[0].author.display_name	Dmitry Manning-Coe
authorships[0].author_position	first
authorships[0].raw_author_name	Manning-Coe, Dmitry
authorships[0].is_corresponding	False
authorships[1].author.id	https://openalex.org/A5068083761
authorships[1].author.orcid
authorships[1].author.display_name	Jacopo Gliozzi
authorships[1].author_position	middle
authorships[1].raw_author_name	Gliozzi, Jacopo
authorships[1].is_corresponding	False
authorships[2].author.id	https://openalex.org/A5091983021
authorships[2].author.orcid	https://orcid.org/0009-0009-6784-7779
authorships[2].author.display_name	Alexander G. Stapleton
authorships[2].author_position	middle
authorships[2].raw_author_name	Stapleton, Alexander G.
authorships[2].is_corresponding	False
authorships[3].author.id	https://openalex.org/A5021626559
authorships[3].author.orcid	https://orcid.org/0000-0003-1699-4399
authorships[3].author.display_name	Edward Hirst
authorships[3].author_position	middle
authorships[3].raw_author_name	Hirst, Edward
authorships[3].is_corresponding	False
authorships[4].author.id	https://openalex.org/A5008914714
authorships[4].author.orcid
authorships[4].author.display_name	Giuseppe De Tomasi
authorships[4].author_position	middle
authorships[4].raw_author_name	De Tomasi, Giuseppe
authorships[4].is_corresponding	False
authorships[5].author.id	https://openalex.org/A5064275485
authorships[5].author.orcid	https://orcid.org/0000-0001-6327-1076
authorships[5].author.display_name	Barry Bradlyn
authorships[5].author_position	middle
authorships[5].raw_author_name	Bradlyn, Barry
authorships[5].is_corresponding	False
authorships[6].author.id	https://openalex.org/A5009524117
authorships[6].author.orcid	https://orcid.org/0000-0002-5382-1668
authorships[6].author.display_name	David S. Berman
authorships[6].author_position	last
authorships[6].raw_author_name	Berman, David S.
authorships[6].is_corresponding	False
has_content.pdf	False
has_content.grobid_xml	False
is_paratext	False
open_access.is_oa	True
open_access.oa_url	https://arxiv.org/pdf/2502.01739
open_access.oa_status	green
open_access.any_repository_has_fulltext	False
created_date	2025-10-10T00:00:00
display_name	Grokking vs. Learning: Same Features, Different Encodings
has_fulltext	False
is_retracted	False
updated_date	2025-11-06T06:51:31.235846
primary_topic
related_works	https://openalex.org/W4391375266, https://openalex.org/W2899084033, https://openalex.org/W2748952813, https://openalex.org/W2390279801, https://openalex.org/W4391913857, https://openalex.org/W2358668433, https://openalex.org/W4396701345, https://openalex.org/W2376932109, https://openalex.org/W2001405890, https://openalex.org/W4396696052
cited_by_count	0
locations_count	2
best_oa_location.id	pmh:oai:arXiv.org:2502.01739
best_oa_location.is_oa	True
best_oa_location.source.id	https://openalex.org/S4306400194
best_oa_location.source.issn
best_oa_location.source.type	repository
best_oa_location.source.is_oa	True
best_oa_location.source.issn_l
best_oa_location.source.is_core	False
best_oa_location.source.is_in_doaj	False
best_oa_location.source.display_name	arXiv (Cornell University)
best_oa_location.source.host_organization	https://openalex.org/I205783295
best_oa_location.source.host_organization_name	Cornell University
best_oa_location.source.host_organization_lineage	https://openalex.org/I205783295
best_oa_location.license
best_oa_location.pdf_url	https://arxiv.org/pdf/2502.01739
best_oa_location.version	submittedVersion
best_oa_location.raw_type	text
best_oa_location.license_id
best_oa_location.is_accepted	False
best_oa_location.is_published	False
best_oa_location.raw_source_name
best_oa_location.landing_page_url	http://arxiv.org/abs/2502.01739
primary_location.id	pmh:oai:arXiv.org:2502.01739
primary_location.is_oa	True
primary_location.source.id	https://openalex.org/S4306400194
primary_location.source.issn
primary_location.source.type	repository
primary_location.source.is_oa	True
primary_location.source.issn_l
primary_location.source.is_core	False
primary_location.source.is_in_doaj	False
primary_location.source.display_name	arXiv (Cornell University)
primary_location.source.host_organization	https://openalex.org/I205783295
primary_location.source.host_organization_name	Cornell University
primary_location.source.host_organization_lineage	https://openalex.org/I205783295
primary_location.license
primary_location.pdf_url	https://arxiv.org/pdf/2502.01739
primary_location.version	submittedVersion
primary_location.raw_type	text
primary_location.license_id
primary_location.is_accepted	False
primary_location.is_published	False
primary_location.raw_source_name
primary_location.landing_page_url	http://arxiv.org/abs/2502.01739
publication_date	2025-02-03
publication_year	2025
referenced_works_count	0
abstract_inverted_index.-	16, 21
abstract_inverted_index.a	81, 92, 171
abstract_inverted_index.5x	120
abstract_inverted_index.In	77, 106
abstract_inverted_index.To	30
abstract_inverted_index.We	9, 50, 127, 138
abstract_inverted_index.be	65
abstract_inverted_index.do	31
abstract_inverted_index.in	26, 47, 68, 88, 104, 125, 143, 174
abstract_inverted_index.is	102, 145, 151
abstract_inverted_index.of	41, 85
abstract_inverted_index.so	32
abstract_inverted_index.to	5, 23
abstract_inverted_index.we	33, 79, 109
abstract_inverted_index.25x	114
abstract_inverted_index.and	38, 54, 98, 100, 119, 133, 147
abstract_inverted_index.are	75, 162
abstract_inverted_index.ask	10
abstract_inverted_index.but	62
abstract_inverted_index.can	64, 110
abstract_inverted_index.how	130
abstract_inverted_index.the	27, 35, 59, 69, 116, 122, 155
abstract_inverted_index.two	48
abstract_inverted_index.via	44
abstract_inverted_index.base	117
abstract_inverted_index.each	45
abstract_inverted_index.find	51, 80
abstract_inverted_index.lead	22
abstract_inverted_index.loss	4, 97
abstract_inverted_index.path	46, 173
abstract_inverted_index.peak	149
abstract_inverted_index.same	60
abstract_inverted_index.show	139
abstract_inverted_index.that	52, 140, 148, 166
abstract_inverted_index.then	128
abstract_inverted_index.this	107
abstract_inverted_index.with	71
abstract_inverted_index.after	154
abstract_inverted_index.large	66
abstract_inverted_index.learn	58
abstract_inverted_index.model	96, 131, 141
abstract_inverted_index.novel	82, 159
abstract_inverted_index.paths	15
abstract_inverted_index.there	63, 90
abstract_inverted_index.these	12, 73
abstract_inverted_index.times	115, 121
abstract_inverted_index.track	129
abstract_inverted_index.which	72, 89, 101, 164
abstract_inverted_index.absent	103
abstract_inverted_index.follow	170
abstract_inverted_index.linear	93
abstract_inverted_index.model,	118
abstract_inverted_index.models	42, 57, 167
abstract_inverted_index.space.	176
abstract_inverted_index.steady	86
abstract_inverted_index.tasks.	49
abstract_inverted_index.versus	18
abstract_inverted_index.achieve	111
abstract_inverted_index.between	95
abstract_inverted_index.compare	34
abstract_inverted_index.develop	135
abstract_inverted_index.emerges	91
abstract_inverted_index.factors	113
abstract_inverted_index.grokked	53
abstract_inverted_index.learned	28
abstract_inverted_index.models.	29
abstract_inverted_index.regime"	84
abstract_inverted_index.regime,	108
abstract_inverted_index.similar	3
abstract_inverted_index.through	136
abstract_inverted_index.trained	43, 56
abstract_inverted_index.whether	11
abstract_inverted_index.Finally,	158
abstract_inverted_index.Grokking	0
abstract_inverted_index.achieved	124, 152
abstract_inverted_index.achieves	2
abstract_inverted_index.dynamics	40
abstract_inverted_index.encoded.	76
abstract_inverted_index.features	74, 132
abstract_inverted_index.grokking	17, 144, 156, 169
abstract_inverted_index.learning	14, 39
abstract_inverted_index.measures	161
abstract_inverted_index.ordinary	19
abstract_inverted_index.plateau.	157
abstract_inverted_index.steadily	55
abstract_inverted_index.straight	172
abstract_inverted_index.training	20, 87
abstract_inverted_index."steady",	7
abstract_inverted_index.different	13
abstract_inverted_index.features,	36, 61
abstract_inverted_index.grokking.	105, 126
abstract_inverted_index.learning.	8
abstract_inverted_index.ordinary,	6
abstract_inverted_index.trade-off	94
abstract_inverted_index.training.	137
abstract_inverted_index.typically	1
abstract_inverted_index.efficiency	70
abstract_inverted_index.introduced	163
abstract_inverted_index.undergoing	168
abstract_inverted_index.compression	112, 123
abstract_inverted_index.demonstrate	165
abstract_inverted_index.development	142
abstract_inverted_index.differences	25, 67
abstract_inverted_index.fundamental	24
abstract_inverted_index.immediately	153
abstract_inverted_index.information	175
abstract_inverted_index.particular,	78
abstract_inverted_index."compressive	83
abstract_inverted_index.compressibility	134, 150
abstract_inverted_index.task-dependent,	146
abstract_inverted_index.compressibility,	37, 99
abstract_inverted_index.information-geometric	160
cited_by_percentile_year
countries_distinct_count	0
institutions_distinct_count	7
citation_normalized_percentile