concept
active
concept:prelude-and-coda-layersPrelude and Coda Layers
Non-recurrent feedforward layers before and after the recurrent block in a looped transformer
Neighborhood — ranked by edge-count
Papers (1)
paper
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Discussion.
- Layers with weak anchoring due to generic representations.
- Consistent but alternative explanations like amplification exist.
- Layers where anchoring weakens systematically due to representational drift.
- Mask virtual weights during training targeting interference.
- Task-specific peak anchoring score for structured reasoning domains.
- Lenc & Vedaldi finding on layer-wise alignment
- Adaptation of logit attribution to CLT features that decode into all subsequent layers.