Saved in:
| Hovedforfatter: | |
|---|---|
| Format: | Recurso digital |
| Sprog: | |
| Udgivet: |
Zenodo
2025
|
| Online adgang: | https://doi.org/10.57967/hf/7066 |
| Tags: |
Tilføj Tag
Ingen Tags, Vær først til at tagge denne postø!
|
Indholdsfortegnelse:
- <p>We present LeftAndRight, a diagnostic framework using four algorithmic primitives (>>,<br><<, 1, 0) to reveal a fundamental property of transformer representations: they geometrically<br>collapse backward operations, regardless of attention architecture.<br>The counterintuitive discovery: We initially hypothesized that causal attention masks<br>cause this collapse. Through systematic validation across three levels—attention patterns, to-<br>ken embeddings, and sentence embeddings—we discovered that even bidirectional models<br>collapse backward operations. DistilBERT, which can attend to future tokens (36.2% future<br>attention), shows zero backward primitives (<< = 0%) at both token and sentence levels.<br>This reveals that the collapse is not caused by attention masks, but by representation<br>geometry itself. Our experiments on 25 boundary problems (OpenXOR, TSP, SAT) and<br>three model architectures (MiniLM, Pythia, DistilBERT) show universal collapse (A = 1.000<br>across all tests). We demonstrate that learned representations encode inherent temporal<br>directionality—possibly from positional encodings, training data ordering, or fundamental<br>properties of sequential modeling—that prevents encoding of backward operations even when<br>attention is bidirectional.<br>This is not about causal attention. This is about how representations form. The<br>4 atoms revealed a deeper geometric truth than expected: transformers fail at backtracking not<br>because of attention architecture, but because their representation space is geometrically<br>unidirectional</p>