finding
active
finding:at-s-30-fork-region-tvd-rises-from-0-0669-to-0-0936-to-0-1587-as-spacing-goes-n-4-to-8-to-16-tokens-saturating-near-0-16At S=30 fork-region TVD rises from 0.0669 to 0.0936 to 0.1587 as spacing goes N=4 to 8 to 16 tokens, saturating near 0.16.
Forks missed with coarse spacing.
Source paper
extracted_from(2026) · Eric Bigelow · Amir Zur · Satchel Grant · Tal Haklay +6
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Ablation interpretation.
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.
- Ablation.
- Limitation.
- On recorded DeepSeek example, Kernel Pooling alone has TVD 0.2720 versus 0.0088 for Full Model.finding0.787Blurred forks.
- Model excess TVD above noise floor falls monotonically from 0.0265 (S=5) to 0.0056 (S=100).finding0.786Accuracy gain.
- At S=200 variation across runs (TVD 0.0451) nearly equals sampling variation within a run (TVD 0.0442).finding0.785Supports multinomial model.
- Replacing kernel pooling with flat per-segment pooling raises pooled TVD by 0.0021 for Llama at S=30.finding0.778Ablation.
- Noise matches null.
- Log-log slope of −0.4903 of pooled replicate TVD vs S through S=200 on two development questions, near −1/2.finding0.767Noise scaling.