Rethinking Deep Alignment Through The Lens Of Incomplete Learning Article Swipe
Thong Bach
,
Truyen Tran
·
YOU?
·
· 2025
· Open Access
·
· DOI: https://doi.org/10.48550/arxiv.2511.12155
YOU?
·
· 2025
· Open Access
·
· DOI: https://doi.org/10.48550/arxiv.2511.12155
Large language models exhibit systematic vulnerabilities to adversarial attacks despite extensive safety alignment. We provide a mechanistic analysis revealing that position-dependent gradient weakening during autoregressive training creates signal decay, leading to incomplete safety learning where safety training fails to transform model preferences in later response regions fully. We introduce base-favored tokens -- vocabulary elements where base models assign higher probability than aligned models -- as computational indicators of incomplete safety learning and develop a targeted completion method that addresses undertrained regions through adaptive penalties and hybrid teacher distillation. Experimental evaluation across Llama and Qwen model families demonstrates dramatic improvements in adversarial robustness, with 48--98% reductions in attack success rates while preserving general capabilities. These results establish both a mechanistic understanding and practical solutions for fundamental limitations in safety alignment methodologies.
Related Topics
Concepts
No concepts available.
Metadata
- Type
- preprint
- Landing Page
- http://arxiv.org/abs/2511.12155
- https://arxiv.org/pdf/2511.12155
- OA Status
- green
- OpenAlex ID
- https://openalex.org/W4416353948
All OpenAlex metadata
Raw OpenAlex JSON
- OpenAlex ID
-
https://openalex.org/W4416353948Canonical identifier for this work in OpenAlex
- DOI
-
https://doi.org/10.48550/arxiv.2511.12155Digital Object Identifier
- Title
-
Rethinking Deep Alignment Through The Lens Of Incomplete LearningWork title
- Type
-
preprintOpenAlex work type
- Publication year
-
2025Year of publication
- Publication date
-
2025-11-15Full publication date if available
- Authors
-
Thong Bach, Truyen TranList of authors in order
- Landing page
-
https://arxiv.org/abs/2511.12155Publisher landing page
- PDF URL
-
https://arxiv.org/pdf/2511.12155Direct link to full text PDF
- Open access
-
YesWhether a free full text is available
- OA status
-
greenOpen access status per OpenAlex
- OA URL
-
https://arxiv.org/pdf/2511.12155Direct OA link when available
- Cited by
-
0Total citation count in OpenAlex
Full payload
| id | https://openalex.org/W4416353948 |
|---|---|
| doi | https://doi.org/10.48550/arxiv.2511.12155 |
| ids.doi | https://doi.org/10.48550/arxiv.2511.12155 |
| ids.openalex | https://openalex.org/W4416353948 |
| fwci | |
| type | preprint |
| title | Rethinking Deep Alignment Through The Lens Of Incomplete Learning |
| biblio.issue | |
| biblio.volume | |
| biblio.last_page | |
| biblio.first_page | |
| is_xpac | False |
| apc_list | |
| apc_paid | |
| language | |
| locations[0].id | pmh:oai:arXiv.org:2511.12155 |
| 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/2511.12155 |
| 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/2511.12155 |
| locations[1].id | doi:10.48550/arxiv.2511.12155 |
| 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.2511.12155 |
| indexed_in | arxiv, datacite |
| authorships[0].author.id | https://openalex.org/A5065706587 |
| authorships[0].author.orcid | |
| authorships[0].author.display_name | Thong Bach |
| authorships[0].author_position | first |
| authorships[0].raw_author_name | Bach, Thong |
| authorships[0].is_corresponding | False |
| authorships[1].author.id | https://openalex.org/A5085471517 |
| authorships[1].author.orcid | https://orcid.org/0000-0001-6531-8907 |
| authorships[1].author.display_name | Truyen Tran |
| authorships[1].author_position | middle |
| authorships[1].raw_author_name | Tran, Truyen |
| authorships[1].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/2511.12155 |
| open_access.oa_status | green |
| open_access.any_repository_has_fulltext | False |
| created_date | 2025-11-19T00:00:00 |
| display_name | Rethinking Deep Alignment Through The Lens Of Incomplete Learning |
| has_fulltext | False |
| is_retracted | False |
| updated_date | 2025-11-28T11:58:38.604648 |
| primary_topic | |
| cited_by_count | 0 |
| locations_count | 2 |
| best_oa_location.id | pmh:oai:arXiv.org:2511.12155 |
| 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/2511.12155 |
| 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/2511.12155 |
| primary_location.id | pmh:oai:arXiv.org:2511.12155 |
| 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/2511.12155 |
| 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/2511.12155 |
| publication_date | 2025-11-15 |
| publication_year | 2025 |
| referenced_works_count | 0 |
| abstract_inverted_index.a | 15, 73, 117 |
| abstract_inverted_index.-- | 51, 63 |
| abstract_inverted_index.We | 13, 47 |
| abstract_inverted_index.as | 64 |
| abstract_inverted_index.in | 42, 99, 105, 126 |
| abstract_inverted_index.of | 67 |
| abstract_inverted_index.to | 6, 30, 38 |
| abstract_inverted_index.and | 71, 84, 92, 120 |
| abstract_inverted_index.for | 123 |
| abstract_inverted_index.Qwen | 93 |
| abstract_inverted_index.base | 55 |
| abstract_inverted_index.both | 116 |
| abstract_inverted_index.than | 60 |
| abstract_inverted_index.that | 19, 77 |
| abstract_inverted_index.with | 102 |
| abstract_inverted_index.Large | 0 |
| abstract_inverted_index.Llama | 91 |
| abstract_inverted_index.These | 113 |
| abstract_inverted_index.fails | 37 |
| abstract_inverted_index.later | 43 |
| abstract_inverted_index.model | 40, 94 |
| abstract_inverted_index.rates | 108 |
| abstract_inverted_index.where | 34, 54 |
| abstract_inverted_index.while | 109 |
| abstract_inverted_index.across | 90 |
| abstract_inverted_index.assign | 57 |
| abstract_inverted_index.attack | 106 |
| abstract_inverted_index.decay, | 28 |
| abstract_inverted_index.during | 23 |
| abstract_inverted_index.fully. | 46 |
| abstract_inverted_index.higher | 58 |
| abstract_inverted_index.hybrid | 85 |
| abstract_inverted_index.method | 76 |
| abstract_inverted_index.models | 2, 56, 62 |
| abstract_inverted_index.safety | 11, 32, 35, 69, 127 |
| abstract_inverted_index.signal | 27 |
| abstract_inverted_index.tokens | 50 |
| abstract_inverted_index.48--98% | 103 |
| abstract_inverted_index.aligned | 61 |
| abstract_inverted_index.attacks | 8 |
| abstract_inverted_index.creates | 26 |
| abstract_inverted_index.despite | 9 |
| abstract_inverted_index.develop | 72 |
| abstract_inverted_index.exhibit | 3 |
| abstract_inverted_index.general | 111 |
| abstract_inverted_index.leading | 29 |
| abstract_inverted_index.provide | 14 |
| abstract_inverted_index.regions | 45, 80 |
| abstract_inverted_index.results | 114 |
| abstract_inverted_index.success | 107 |
| abstract_inverted_index.teacher | 86 |
| abstract_inverted_index.through | 81 |
| abstract_inverted_index.adaptive | 82 |
| abstract_inverted_index.analysis | 17 |
| abstract_inverted_index.dramatic | 97 |
| abstract_inverted_index.elements | 53 |
| abstract_inverted_index.families | 95 |
| abstract_inverted_index.gradient | 21 |
| abstract_inverted_index.language | 1 |
| abstract_inverted_index.learning | 33, 70 |
| abstract_inverted_index.response | 44 |
| abstract_inverted_index.targeted | 74 |
| abstract_inverted_index.training | 25, 36 |
| abstract_inverted_index.addresses | 78 |
| abstract_inverted_index.alignment | 128 |
| abstract_inverted_index.establish | 115 |
| abstract_inverted_index.extensive | 10 |
| abstract_inverted_index.introduce | 48 |
| abstract_inverted_index.penalties | 83 |
| abstract_inverted_index.practical | 121 |
| abstract_inverted_index.revealing | 18 |
| abstract_inverted_index.solutions | 122 |
| abstract_inverted_index.transform | 39 |
| abstract_inverted_index.weakening | 22 |
| abstract_inverted_index.alignment. | 12 |
| abstract_inverted_index.completion | 75 |
| abstract_inverted_index.evaluation | 89 |
| abstract_inverted_index.incomplete | 31, 68 |
| abstract_inverted_index.indicators | 66 |
| abstract_inverted_index.preserving | 110 |
| abstract_inverted_index.reductions | 104 |
| abstract_inverted_index.systematic | 4 |
| abstract_inverted_index.vocabulary | 52 |
| abstract_inverted_index.adversarial | 7, 100 |
| abstract_inverted_index.fundamental | 124 |
| abstract_inverted_index.limitations | 125 |
| abstract_inverted_index.mechanistic | 16, 118 |
| abstract_inverted_index.preferences | 41 |
| abstract_inverted_index.probability | 59 |
| abstract_inverted_index.robustness, | 101 |
| abstract_inverted_index.Experimental | 88 |
| abstract_inverted_index.base-favored | 49 |
| abstract_inverted_index.demonstrates | 96 |
| abstract_inverted_index.improvements | 98 |
| abstract_inverted_index.undertrained | 79 |
| abstract_inverted_index.capabilities. | 112 |
| abstract_inverted_index.computational | 65 |
| abstract_inverted_index.distillation. | 87 |
| abstract_inverted_index.understanding | 119 |
| abstract_inverted_index.autoregressive | 24 |
| abstract_inverted_index.methodologies. | 129 |
| abstract_inverted_index.vulnerabilities | 5 |
| abstract_inverted_index.position-dependent | 20 |
| cited_by_percentile_year | |
| countries_distinct_count | 0 |
| institutions_distinct_count | 2 |
| citation_normalized_percentile |