claim
active
claim:segmentation-is-needed-to-preserve-forks-kernel-pooling-alone-blurs-themSegmentation is needed to preserve forks; kernel pooling alone blurs them.
Ablation interpretation.
Source paper
extracted_from(2026) · Eric Bigelow · Amir Zur · Satchel Grant · Tal Haklay +6
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Forks missed with coarse spacing.
- On recorded DeepSeek example, Kernel Pooling alone has TVD 0.2720 versus 0.0088 for Full Model.supportsBlurred forks.
- Ablation.
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.
- Discovery story.
- Replacing kernel pooling with flat per-segment pooling raises pooled TVD by 0.0021 for Llama at S=30.finding0.743Ablation.
- Interesting special case of copying behavior related to tokenization artifacts; primitive precursor to induction heads
- Combining spaced resampling with smoothing cuts total budget by 1/8 with small error increase.finding0.739Main result 3.
- Observed across SAE scales, e.g., 'San Francisco' split into 11 features.
- Discussion of cortical analogy.
- With raw data, wider spacing has little effect on accuracy except at high S and near forking points.finding0.732Raw data behavior.