Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper disentangles positional and semantic information within Transformer encoders by creating separate processing streams for semantic content, absolute position (AP), and relative position (RP), while restricting the masked language modeling (MLM) objective to the semantic stream. This disentanglement reveals that the AP subspace collapses into a low-frequency 2D manifold capturing document structure, and that attention heads specialize into structure and semantic-oriented groups. The approach improves linguistic representation on the Flash-Holmes probing benchmark, suggesting that explicitly disentangling positional information can enhance language understanding.
Transformers struggle to retain macroscopic document structure because standard positional encodings like RoPE and entangled AP representations lose this information under MLM pressure.
Positional encoding (PE) underpins how permutation-invariant Transformers represent sequence order, yet how positional information is processed and stored remains poorly understood. Modern PE methods such as RoPE still struggle on tasks such as long-context understanding or retrieval \cite{chen-etal-2025-hope}. Hence, a better understanding of the internal positional mechanism could help design better PE. Building on evidence that positional and semantic signals occupy nearly orthogonal subspaces in trained Transformers, we modify an encoder Transformer to process three explicitly disentangled streams: semantic, absolute positional (AP) and relative positional (RP), and confine the masked-language-modeling (MLM) objective to the semantic stream. This decoupling enables a clean mechanistic study and yields three take-aways. (1) The isolated AP subspace spontaneously collapses into a low-frequency two-dimensional manifold that captures the structure of the document; (2) Attention heads specialize into structure and semantic-oriented groups, with RP exclusively supporting the latter; (3) Standard positional encodings do not robustly retain macroscopic structure: RoPE and RP only weakly encode it, and entangled AP loses it in the final layers under MLM pressure. The disentangled approach preserves positional encoding, which improves linguistic representation on 49 of the 65 linguistic phenomena of the Flash-Holmes probing benchmark.