{"id":"59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a","title":"Obadinma and Zhu: Symbolic Swapper exposes weaknesses in LLM planning","argument":"Stephen Obadinma and Xiaodan Zhu argue that LLM planning can be fragile when models cannot rely on familiar cues in symbolic vocabulary—the names of actions, predicates and objects used to define planning tasks. They introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies. The authors report that these attacks can decrease planning performance by over 70% for base models and robust planning frameworks. Their finding is not just that unfamiliar vocabulary makes planning harder: they report that models may retain some ability to understand a domain even as their planning performance and apparent reasoning quality degrade. This distinction matters to interpreting a model’s failure: poor planning performance may coexist with some ability to understand the task domain. The authors use adversarial obfuscation to probe weaknesses in joint symbolic interpretation and planning, including in reasoning models that perform better on tasks with obfuscated symbolic vocabularies. The abstract describes autonomous agents as a motivation, but its reported results concern planning tasks; it does not establish how those results translate to deployed agents.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper’s abstract and does not provide the task-level conditions, model-by-model results, or measurements behind the reported performance decrease. The abstract does not establish how the results apply to deployed autonomous agents beyond the planning tasks studied.","speaker":{"name":"Obadinma, Stephen, Zhu, Xiaodan","roleAtSource":"Article author"},"source":{"title":"Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.17910","publishedAt":"2026-11-13","dateBasis":"Date basis not provided."},"evidence":[{"text":"Obadinma and Zhu argue that LLM planning is not robust when models cannot rely on commonsense cues in symbolic vocabulary.","excerpt":"LLMs' planning ability is not robust, particularly when they are not allowed to exploit commonsense cues in the symbolic vocabulary","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"},{"text":"The authors introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies.","excerpt":"devising an LLM-based attack model, which we call Symbolic Swapper, to automatically generate obfuscated planning domains with non-standard symbolic vocabularies","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"},{"text":"The authors report that attacks can decrease planning performance by over 70% for base models and robust planning frameworks.","excerpt":"attacks can decrease the planning performance of base models and even robust planning frameworks by over 70%","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"},{"text":"The authors report a distinction between models' ability to understand a domain and their ability to plan under it.","excerpt":"successful obfuscation causes models to appear to significantly degrade in reasoning quality despite them showing some inherent ability to understand the domain, revealing a gap in the perception of a task domain and being able to plan under it.","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.17910","checkedAt":"2026-11-16T05:50:08.499Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What task conditions or model differences account for the gap between understanding a domain and planning under it?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Obadinma and Zhu: Symbolic Swapper exposes weaknesses in LLM planning","summary":"Stephen Obadinma and Xiaodan Zhu argue that LLM planning can be fragile when models cannot rely on familiar cues in symbolic vocabulary—the names of actions, predicates and objects used to define planning tasks. They introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies. The authors report that these attacks can decrease planning performance by over 70% for base models and robust planning frameworks. Their finding is not just that unfamiliar vocabulary makes planning harder: they report that models may retain some ability to understand a domain even as their planning performance and apparent reasoning quality degrade. This distinction matters to interpreting a model’s failure: poor planning performance may coexist with some ability to understand the task domain. The authors use adversarial obfuscation to probe weaknesses in joint symbolic interpretation and planning, including in reasoning models that perform better on tasks with obfuscated symbolic vocabularies. The abstract describes autonomous agents as a motivation, but its reported results concern planning tasks; it does not establish how those results translate to deployed agents.","question":"What task conditions or model differences account for the gap between understanding a domain and planning under it?","claims":[{"text":"Obadinma and Zhu argue that LLM planning is not robust when models cannot rely on commonsense cues in symbolic vocabulary.","excerpt":"LLMs' planning ability is not robust, particularly when they are not allowed to exploit commonsense cues in the symbolic vocabulary","support":"source-stated"},{"text":"The authors introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies.","excerpt":"devising an LLM-based attack model, which we call Symbolic Swapper, to automatically generate obfuscated planning domains with non-standard symbolic vocabularies","support":"source-stated"},{"text":"The authors report that attacks can decrease planning performance by over 70% for base models and robust planning frameworks.","excerpt":"attacks can decrease the planning performance of base models and even robust planning frameworks by over 70%","support":"source-stated"},{"text":"The authors report a distinction between models' ability to understand a domain and their ability to plan under it.","excerpt":"successful obfuscation causes models to appear to significantly degrade in reasoning quality despite them showing some inherent ability to understand the domain, revealing a gap in the perception of a task domain and being able to plan under it.","support":"source-stated"}],"limitations":["The supplied evidence is the paper’s abstract and does not provide the task-level conditions, model-by-model results, or measurements behind the reported performance decrease.","The abstract does not establish how the results apply to deployed autonomous agents beyond the planning tasks studied."],"mediaId":"b00f578f43b58a1c49d22039ffead22d7c33f3236ec5ce5ee75cc8da88f13e99","correctionReason":null,"id":"59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a","publicationId":"51c3a11f-80c7-4403-828d-742ab42d4e1f","version":1,"sourceUrl":"https://arxiv.org/abs/2611.17910","sourceTitle":"Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-13","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-16T05:03:55.991Z","createdAt":"2026-11-16T05:35:05.346Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b00f578f43b58a1c49d22039ffead22d7c33f3236ec5ce5ee75cc8da88f13e99","kind":"article","url":"https://arxiv.org/abs/2611.17910","sourceUrl":"https://arxiv.org/abs/2611.17910","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-16T05:34:16.189Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.17910","checkedAt":"2026-11-16T05:50:08.499Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"b00f578f43b58a1c49d22039ffead22d7c33f3236ec5ce5ee75cc8da88f13e99","confidence":0.64,"provider":"typesafe","model":"jev-1.13.0","inputHash":"eea53524ec02794af8670ca5a42cd83f00cc281deba68e05e0a15a9d8abf377f","decidedAt":"2026-11-16T05:34:18.014Z"},"provenance":{"textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observationId":"df7b836959b5dbc8675e41978deb0c3d8ad8e43ae9b9cce7843ef633994bbd2e","retrievedAt":"2026-11-16T05:34:15.914Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"8608c3f2-d7f8-4694-8e84-6653321ca585"},"reviewer":{"model":"gpt-6-luna","runId":"bb73ed12-860b-4d81-8320-e002e0d012ad"}},"audit":{"status":"passed","reviewedAt":"2026-11-16T05:35:05.346Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Obadinma, Stephen","url":null,"kind":"unknown"},{"name":"Zhu, Xiaodan","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Obadinma, Stephen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Obadinma, Stephen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.17910","locator":"head/meta[name=author|citation_author]"},"assertionId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-16T05:50:08.499Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d","sourceUrl":"https://arxiv.org/abs/2611.17910","contentSha256":"429db3a76dd41f61e4cd82f12bc32e3ed9a8eb71fec9109ccffd4e170a4eddf0","textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observedAt":"2026-11-16T05:50:08.499Z"}},{"actor":{"name":"Zhu, Xiaodan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhu, Xiaodan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.17910","locator":"head/meta[name=author|citation_author]"},"assertionId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-16T05:50:08.499Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d","sourceUrl":"https://arxiv.org/abs/2611.17910","contentSha256":"429db3a76dd41f61e4cd82f12bc32e3ed9a8eb71fec9109ccffd4e170a4eddf0","textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observedAt":"2026-11-16T05:50:08.499Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.17910","snapshotId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d","contentSha256":"429db3a76dd41f61e4cd82f12bc32e3ed9a8eb71fec9109ccffd4e170a4eddf0","textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observedAt":"2026-11-16T05:50:08.499Z","basis":"publisher-page-metadata","sourceLinks":40,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a/source"}},"canonicalUrl":"https://escapemeridian.com/readings/obadinma-and-zhu-symbolic-swapper-exposes-weaknesses-in-llm-59efa372","formats":{"html":"/readings/obadinma-and-zhu-symbolic-swapper-exposes-weaknesses-in-llm-59efa372","markdown":"/readings/obadinma-and-zhu-symbolic-swapper-exposes-weaknesses-in-llm-59efa372.md","json":"/readings/obadinma-and-zhu-symbolic-swapper-exposes-weaknesses-in-llm-59efa372.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-llm","name":"Large language model","explanation":"A language model with many learned numerical settings, used to process and generate language.","referenceUrl":"https://developers.google.com/machine-learning/glossary#large-language-model"},{"id":"term-reasoning-model","name":"Reasoning model","explanation":"A language model trained to work through a problem in written steps before it answers.","referenceUrl":"https://en.wikipedia.org/wiki/Reasoning_model"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a","revision":"51c3a11f-80c7-4403-828d-742ab42d4e1f","revisionKind":"publication-id","title":"Obadinma and Zhu: Symbolic Swapper exposes weaknesses in LLM planning","url":"https://escapemeridian.com/readings/obadinma-and-zhu-symbolic-swapper-exposes-weaknesses-in-llm-59efa372"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.17910","revision":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","revisionKind":"sha256-source-text","title":"Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners","url":"https://arxiv.org/abs/2611.17910"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a","title":"Obadinma and Zhu: Symbolic Swapper exposes weaknesses in LLM planning","argument":"Stephen Obadinma and Xiaodan Zhu argue that LLM planning can be fragile when models cannot rely on familiar cues in symbolic vocabulary—the names of actions, predicates and objects used to define planning tasks. They introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies. The authors report that these attacks can decrease planning performance by over 70% for base models and robust planning frameworks. Their finding is not just that unfamiliar vocabulary makes planning harder: they report that models may retain some ability to understand a domain even as their planning performance and apparent reasoning quality degrade. This distinction matters to interpreting a model’s failure: poor planning performance may coexist with some ability to understand the task domain. The authors use adversarial obfuscation to probe weaknesses in joint symbolic interpretation and planning, including in reasoning models that perform better on tasks with obfuscated symbolic vocabularies. The abstract describes autonomous agents as a motivation, but its reported results concern planning tasks; it does not establish how those results translate to deployed agents.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper’s abstract and does not provide the task-level conditions, model-by-model results, or measurements behind the reported performance decrease. The abstract does not establish how the results apply to deployed autonomous agents beyond the planning tasks studied.","speaker":{"name":"Obadinma, Stephen, Zhu, Xiaodan","roleAtSource":"Article author"},"source":{"title":"Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.17910","publishedAt":"2026-11-13","dateBasis":"Date basis not provided."},"evidence":[{"text":"Obadinma and Zhu argue that LLM planning is not robust when models cannot rely on commonsense cues in symbolic vocabulary.","excerpt":"LLMs' planning ability is not robust, particularly when they are not allowed to exploit commonsense cues in the symbolic vocabulary","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"},{"text":"The authors introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies.","excerpt":"devising an LLM-based attack model, which we call Symbolic Swapper, to automatically generate obfuscated planning domains with non-standard symbolic vocabularies","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"},{"text":"The authors report that attacks can decrease planning performance by over 70% for base models and robust planning frameworks.","excerpt":"attacks can decrease the planning performance of base models and even robust planning frameworks by over 70%","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"},{"text":"The authors report a distinction between models' ability to understand a domain and their ability to plan under it.","excerpt":"successful obfuscation causes models to appear to significantly degrade in reasoning quality despite them showing some inherent ability to understand the domain, revealing a gap in the perception of a task domain and being able to plan under it.","url":"https://arxiv.org/abs/2611.17910","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.17910","checkedAt":"2026-11-16T05:50:08.499Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What task conditions or model differences account for the gap between understanding a domain and planning under it?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Obadinma and Zhu: Symbolic Swapper exposes weaknesses in LLM planning","summary":"Stephen Obadinma and Xiaodan Zhu argue that LLM planning can be fragile when models cannot rely on familiar cues in symbolic vocabulary—the names of actions, predicates and objects used to define planning tasks. They introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies. The authors report that these attacks can decrease planning performance by over 70% for base models and robust planning frameworks. Their finding is not just that unfamiliar vocabulary makes planning harder: they report that models may retain some ability to understand a domain even as their planning performance and apparent reasoning quality degrade. This distinction matters to interpreting a model’s failure: poor planning performance may coexist with some ability to understand the task domain. The authors use adversarial obfuscation to probe weaknesses in joint symbolic interpretation and planning, including in reasoning models that perform better on tasks with obfuscated symbolic vocabularies. The abstract describes autonomous agents as a motivation, but its reported results concern planning tasks; it does not establish how those results translate to deployed agents.","question":"What task conditions or model differences account for the gap between understanding a domain and planning under it?","claims":[{"text":"Obadinma and Zhu argue that LLM planning is not robust when models cannot rely on commonsense cues in symbolic vocabulary.","excerpt":"LLMs' planning ability is not robust, particularly when they are not allowed to exploit commonsense cues in the symbolic vocabulary","support":"source-stated"},{"text":"The authors introduce Symbolic Swapper, an LLM-based attack model that automatically generates planning domains with non-standard symbolic vocabularies.","excerpt":"devising an LLM-based attack model, which we call Symbolic Swapper, to automatically generate obfuscated planning domains with non-standard symbolic vocabularies","support":"source-stated"},{"text":"The authors report that attacks can decrease planning performance by over 70% for base models and robust planning frameworks.","excerpt":"attacks can decrease the planning performance of base models and even robust planning frameworks by over 70%","support":"source-stated"},{"text":"The authors report a distinction between models' ability to understand a domain and their ability to plan under it.","excerpt":"successful obfuscation causes models to appear to significantly degrade in reasoning quality despite them showing some inherent ability to understand the domain, revealing a gap in the perception of a task domain and being able to plan under it.","support":"source-stated"}],"limitations":["The supplied evidence is the paper’s abstract and does not provide the task-level conditions, model-by-model results, or measurements behind the reported performance decrease.","The abstract does not establish how the results apply to deployed autonomous agents beyond the planning tasks studied."],"mediaId":"b00f578f43b58a1c49d22039ffead22d7c33f3236ec5ce5ee75cc8da88f13e99","correctionReason":null,"id":"59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a","publicationId":"51c3a11f-80c7-4403-828d-742ab42d4e1f","version":1,"sourceUrl":"https://arxiv.org/abs/2611.17910","sourceTitle":"Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-13","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-16T05:03:55.991Z","createdAt":"2026-11-16T05:35:05.346Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b00f578f43b58a1c49d22039ffead22d7c33f3236ec5ce5ee75cc8da88f13e99","kind":"article","url":"https://arxiv.org/abs/2611.17910","sourceUrl":"https://arxiv.org/abs/2611.17910","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-16T05:34:16.189Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.17910","checkedAt":"2026-11-16T05:50:08.499Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"b00f578f43b58a1c49d22039ffead22d7c33f3236ec5ce5ee75cc8da88f13e99","confidence":0.64,"provider":"typesafe","model":"jev-1.13.0","inputHash":"eea53524ec02794af8670ca5a42cd83f00cc281deba68e05e0a15a9d8abf377f","decidedAt":"2026-11-16T05:34:18.014Z"},"provenance":{"textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observationId":"df7b836959b5dbc8675e41978deb0c3d8ad8e43ae9b9cce7843ef633994bbd2e","retrievedAt":"2026-11-16T05:34:15.914Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"8608c3f2-d7f8-4694-8e84-6653321ca585"},"reviewer":{"model":"gpt-6-luna","runId":"bb73ed12-860b-4d81-8320-e002e0d012ad"}},"audit":{"status":"passed","reviewedAt":"2026-11-16T05:35:05.346Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Obadinma, Stephen","url":null,"kind":"unknown"},{"name":"Zhu, Xiaodan","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Obadinma, Stephen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Obadinma, Stephen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.17910","locator":"head/meta[name=author|citation_author]"},"assertionId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-16T05:50:08.499Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d","sourceUrl":"https://arxiv.org/abs/2611.17910","contentSha256":"429db3a76dd41f61e4cd82f12bc32e3ed9a8eb71fec9109ccffd4e170a4eddf0","textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observedAt":"2026-11-16T05:50:08.499Z"}},{"actor":{"name":"Zhu, Xiaodan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhu, Xiaodan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.17910","locator":"head/meta[name=author|citation_author]"},"assertionId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-16T05:50:08.499Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d","sourceUrl":"https://arxiv.org/abs/2611.17910","contentSha256":"429db3a76dd41f61e4cd82f12bc32e3ed9a8eb71fec9109ccffd4e170a4eddf0","textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observedAt":"2026-11-16T05:50:08.499Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.17910","snapshotId":"f31d346ece8945302bea271cd2feccbf1bc6e1b0972568ffdcbc96271d7e893d","contentSha256":"429db3a76dd41f61e4cd82f12bc32e3ed9a8eb71fec9109ccffd4e170a4eddf0","textHash":"2fa01ffe0b109f7ec2eccdf782fef77ac225c396de407e5d23f6b133d4735e8c","observedAt":"2026-11-16T05:50:08.499Z","basis":"publisher-page-metadata","sourceLinks":40,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/59efa372c56d743de50290a42f5d251ecfe8f074b34cbd7eb933547704be892a/source"}}}],"audits":[{"publicationId":"51c3a11f-80c7-4403-828d-742ab42d4e1f","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The paper, its arguments, the Symbolic Swapper attack model and the reported results are attributed to authors Stephen Obadinma and Xiaodan Zhu, matching the arXiv listing. Claims stay framed as the authors' arguments and findings, not as established fact.","excerpt":"Authors:Stephen Obadinma, Xiaodan Zhu"},{"dimension":"summary","status":"pass","explanation":"Title, summary and question match the abstract. The summary leads with the actor and argument, defines symbolic vocabulary, and gives the over-70% result and the understanding-versus-planning gap. It keeps the 'appear to' hedge and notes the agent-deployment gap. Minor optional repairs: expand LLM (large language model) and say plainly that obfuscation means renaming actions and objects with unfamiliar names. Optionally restore the abstract's contrast with symbolic planners and its 'strong reasoning models' force. The question follows from the evidence.","excerpt":"systematically break strong reasoning models' ability to plan"},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-11-13 matches the arXiv v1 submission date. The feed date of 2026-11-16 is correctly not used. The NeurIPS 2026 comment does not change the first-publication date.","excerpt":"[Submitted on 13 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Each claim is directly supported by a verbatim abstract excerpt. The over-70% figure keeps the abstract's 'can decrease' wording. The understanding-versus-planning distinction keeps the source's 'appear to' hedge. The limitations accurately note the missing task-level and model-by-model details and the unestablished link to deployed agents. No unsupported benefits, costs or forecasts are added.","excerpt":"successful obfuscation causes models to appear to significantly degrade in reasoning quality despite them showing some inherent ability to understand the domain"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper. It matches the source URL and is the most direct way to reach the original. No better source media was found.","excerpt":null}],"createdAt":"2026-11-16T05:35:05.346Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}