AWhat 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.
Scientific 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
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.
Contributors
- The named authors in the research record · SCIENCE: 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.
Provisional WRS assessment, edition 2026-10-23.6. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction