{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "title": "Universal Transformers: repeatedly refining an attention-based state", "summary": "A standard Transformer applies a sequence of distinct layers. A Universal Transformer instead reuses a transformation across processing steps, allowing its representation to be refined repeatedly. Self-attention lets positions exchange information during that refinement. An adaptive-computation variant can allocate different numbers of steps to different positions. Dehghani and colleagues investigate this combination on language, reasoning and algorithmic tasks. The contribution is the recurrent-depth architecture and its tested variants, rather than the invention of attention, Transformers or adaptive computation themselves.", "boundary": "Theoretical computational universality requires sufficient memory and steps that may grow with the input; it does not make a fixed deployed network a universal solver. Longer-sequence algorithmic tests remain imperfect. Adaptive halting and fixed-depth versions have different results and costs, and the strongest translation result is not evidence that adaptive halting always saves computation. Benchmark selection, pre-existing components and finite training settings remain part of the comparison.", "area": "Adaptive-depth sequence architectures", "topic": "Artificial intelligence", "year": 2018, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "target": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "humans": [ "Mostafa Dehghani", "Stephan Gouws", "Oriol Vinyals", "Jakob Uszkoreit", "Łukasz Kaiser" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [ "transformer-architecture" ], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-1807-03819v1", "title": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1", "date": "2018-07-10" } ], "points": "120000", "published_grade": { "id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "entry_id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "band": "A", "modifier": "+", "target": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "what_changed": "A standard Transformer applies a sequence of distinct layers. A Universal Transformer instead reuses a transformation across processing steps, allowing its representation to be refined repeatedly. Self-attention lets positions exchange information during that refinement. An adaptive-computation variant can allocate different numbers of steps to different positions. Dehghani and colleagues investigate this combination on language, reasoning and algorithmic tasks. The contribution is the recurrent-depth architecture and its tested variants, rather than the invention of attention, Transformers or adaptive computation themselves.", "scope": "A standard Transformer applies a sequence of distinct layers. A Universal Transformer instead reuses a transformation across processing steps, allowing its representation to be refined repeatedly. Self-attention lets positions exchange information during that refinement. An adaptive-computation variant can allocate different numbers of steps to different positions. Dehghani and colleagues investigate this combination on language, reasoning and algorithmic tasks. The contribution is the recurrent-depth architecture and its tested variants, rather than the invention of attention, Transformers or adaptive computation themselves. Theoretical computational universality requires sufficient memory and steps that may grow with the input; it does not make a fixed deployed network a universal solver. Longer-sequence algorithmic tests remain imperfect. Adaptive halting and fixed-depth versions have different results and costs, and the strongest translation result is not evidence that adaptive halting always saves computation. Benchmark selection, pre-existing components and finite training settings remain part of the comparison.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Mostafa Dehghani", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Stephan Gouws", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Oriol Vinyals", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jakob Uszkoreit", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Łukasz Kaiser", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "Theoretical computational universality requires sufficient memory and steps that may grow with the input; it does not make a fixed deployed network a universal solver. Longer-sequence algorithmic tests remain imperfect. Adaptive halting and fixed-depth versions have different results and costs, and the strongest translation result is not evidence that adaptive halting always saves computation. Benchmark selection, pre-existing components and finite training settings remain part of the comparison.", "domain": "Adaptive-depth sequence architectures" }, "sources": [ { "citation": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1" } ], "statement": "# Universal Transformers: repeatedly refining an attention-based state\n\n## What if a layer could take another pass?\n\nWhen solving a puzzle, you might revisit the same facts several times rather than use an entirely different rule at every step. A Universal Transformer explores a related design for a neural network. It repeatedly refines a representation using a shared transformation. Self-attention lets different positions exchange information during each pass. The analogy explains the architecture; it is not a claim that the network experiences thought or understands a puzzle as a person does.\n\n## Separate positions from processing depth\n\nA text has positions for its words or tokens. Processing depth is the sequence of transformations applied to their representations. An ordinary Transformer has multiple layers with distinct parameters. Recurrent depth reuses parameters across steps. The adaptive-computation variant also learns when a position should stop updating, within the implementation’s limits. Shared parameters, the number of steps and the number of tokens are different quantities; changing one does not automatically reduce the others.\n\n## Reusing a rule can still require work\n\nFor a classroom example, apply the rule “replace a number by half its value plus one” starting from zero. The first result is one, the second is one and a half, and the third is one and three-quarters. One rule produced three successive states. Reusing its description did not make the three arithmetic operations disappear. This example is not the Transformer’s update equation; it only separates sharing a rule from executing that rule repeatedly.\n\n## Universality has assumptions\n\nThe paper discusses computational universality when enough memory and processing steps are available and steps can grow with the input. That does not promise that a fixed network, fixed context and fixed time budget solve every problem. Its longer-sequence algorithmic experiments still make errors. Adaptive halting and fixed-step variants also have different results. A better translation score from one variant cannot be used to claim that another variant always saves computation or generalizes perfectly.\n\n## Ask for a transparent comparison\n\nHave your agent tabulate the classroom update for one through six passes, showing each intermediate number. Then ask it to identify parameter sharing, processing steps and measured performance as separate concepts in the paper. It can compare two architecture diagrams or descriptions without training a model. A useful result is a small evidence table that states which variant was tested and under which conditions, rather than a verdict that recurrent processing must always be smarter.\n\n## Scientific scope\n\nTheoretical computational universality requires sufficient memory and steps that may grow with the input; it does not make a fixed deployed network a universal solver. Longer-sequence algorithmic tests remain imperfect. Adaptive halting and fixed-depth versions have different results and costs, and the strongest translation result is not evidence that adaptive halting always saves computation. Benchmark selection, pre-existing components and finite training settings remain part of the comparison.\n\n## Sources for this explanation\n\n- [Universal Transformers](https://arxiv.org/abs/1807.03819v1): Original 2018 main text: recurrent-depth architecture, adaptive computation, universality assumptions and finite language/algorithmic evaluations.\n", "summary": "A standard Transformer applies a sequence of distinct layers. A Universal Transformer instead reuses a transformation across processing steps, allowing its representation to be refined repeatedly. Self-attention lets positions exchange information during that refinement. An adaptive-computation variant can allocate different numbers of steps to different positions. Dehghani and colleagues investigate this combination on language, reasoning and algorithmic tasks. The contribution is the recurrent-depth architecture and its tested variants, rather than the invention of attention, Transformers or adaptive computation themselves.", "title": "Universal Transformers: repeatedly refining an attention-based state" }, "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "source_ref": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:ffa51c9c787297fa9b18fced709e852d7dd217cb5bb27bdcfe553d7082408d92" }, "target": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "entry_id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "revision": 1, "title": "Universal Transformers: repeatedly refining an attention-based state", "context": "Adaptive-depth sequence architectures", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "A standard Transformer applies a sequence of distinct layers. A Universal Transformer instead reuses a transformation across processing steps, allowing its representation to be refined repeatedly. Self-attention lets positions exchange information during that refinement. An adaptive-computation variant can allocate different numbers of steps to different positions. Dehghani and colleagues investigate this combination on language, reasoning and algorithmic tasks. The contribution is the recurrent-depth architecture and its tested variants, rather than the invention of attention, Transformers or adaptive computation themselves.", "scope": "A standard Transformer applies a sequence of distinct layers. A Universal Transformer instead reuses a transformation across processing steps, allowing its representation to be refined repeatedly. Self-attention lets positions exchange information during that refinement. An adaptive-computation variant can allocate different numbers of steps to different positions. Dehghani and colleagues investigate this combination on language, reasoning and algorithmic tasks. The contribution is the recurrent-depth architecture and its tested variants, rather than the invention of attention, Transformers or adaptive computation themselves. Theoretical computational universality requires sufficient memory and steps that may grow with the input; it does not make a fixed deployed network a universal solver. Longer-sequence algorithmic tests remain imperfect. Adaptive halting and fixed-depth versions have different results and costs, and the strongest translation result is not evidence that adaptive halting always saves computation. Benchmark selection, pre-existing components and finite training settings remain part of the comparison.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. Theoretical computational universality requires sufficient memory and steps that may grow with the input; it does not make a fixed deployed network a universal solver. Longer-sequence algorithmic tests remain imperfect. Adaptive halting and fixed-depth versions have different results and costs, and the strongest translation result is not evidence that adaptive halting always saves computation. Benchmark selection, pre-existing components and finite training settings remain part of the comparison.", "references": [ { "title": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "120000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "dehghani-et-al-universal-transformer-depth-recurrence-2018", "target": { "record_id": "urn:uuid:0bb9b83e-813e-4a8f-a28a-fa0913c3cf92", "version": 1, "content_hash": "sha256:c9a8f6450edc557201ffefcde324153691c2168efdbc0a1784c05bd066e26528" }, "sections": [ { "title": "What if a layer could take another pass?", "text": "When solving a puzzle, you might revisit the same facts several times rather than use an entirely different rule at every step. A Universal Transformer explores a related design for a neural network. It repeatedly refines a representation using a shared transformation. Self-attention lets different positions exchange information during each pass. The analogy explains the architecture; it is not a claim that the network experiences thought or understands a puzzle as a person does." }, { "title": "Separate positions from processing depth", "text": "A text has positions for its words or tokens. Processing depth is the sequence of transformations applied to their representations. An ordinary Transformer has multiple layers with distinct parameters. Recurrent depth reuses parameters across steps. The adaptive-computation variant also learns when a position should stop updating, within the implementation’s limits. Shared parameters, the number of steps and the number of tokens are different quantities; changing one does not automatically reduce the others." }, { "title": "Reusing a rule can still require work", "text": "For a classroom example, apply the rule “replace a number by half its value plus one” starting from zero. The first result is one, the second is one and a half, and the third is one and three-quarters. One rule produced three successive states. Reusing its description did not make the three arithmetic operations disappear. This example is not the Transformer’s update equation; it only separates sharing a rule from executing that rule repeatedly." }, { "title": "Universality has assumptions", "text": "The paper discusses computational universality when enough memory and processing steps are available and steps can grow with the input. That does not promise that a fixed network, fixed context and fixed time budget solve every problem. Its longer-sequence algorithmic experiments still make errors. Adaptive halting and fixed-step variants also have different results. A better translation score from one variant cannot be used to claim that another variant always saves computation or generalizes perfectly." }, { "title": "Ask for a transparent comparison", "text": "Have your agent tabulate the classroom update for one through six passes, showing each intermediate number. Then ask it to identify parameter sharing, processing steps and measured performance as separate concepts in the paper. It can compare two architecture diagrams or descriptions without training a model. A useful result is a small evidence table that states which variant was tested and under which conditions, rather than a verdict that recurrent processing must always be smarter." } ], "references": [ { "title": "Universal Transformers", "url": "https://arxiv.org/abs/1807.03819v1", "locator": "Original 2018 main text: recurrent-depth architecture, adaptive computation, universality assumptions and finite language/algorithmic evaluations." } ] }, "reading_status": "EXPLANATORY_GUIDE" }