finding
active
finding:slope-0-4757-for-llama-resampling-every-token-on-full-s-200-datasetSlope −0.4757 for Llama resampling every token on full S=200 dataset.
Noise scaling.
Source paper
extracted_from(2026) · Eric Bigelow · Amir Zur · Satchel Grant · Tal Haklay +6
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Sampling noise decreases in proportion to 1/√S, giving a linear relation between log TVD and log S.supportsDerived from multinomial sampling.
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.
- Noise scaling.
- Llama sample-efficiency multiplier at 1-token spacing: 22.1× at S=5, 7.3× at S=30, 1.6× at S=100.finding0.808Efficiency gain.
- Llama sample-efficiency multiplier at 4-token spacing: 5.0× at S=5, 3.3× at S=30, 1.0× at S=100.finding0.802Efficiency gain.
- LLaMA-3.1-8B: Sbmax = -1.896 ± 0.211, AUSN = -2.119 ± 0.198, peak layer ℓ* = 10 (median)finding0.796Seed-pooled geometry-only statistics (per-dev z units).
- Andromeda Galaxy steering demo.
- Only model showing marginal benefit from increased reflection, at substantial token cost
- At Llama layer 26 (~81% depth) the predictive component does not dominate the latent norm.finding0.792Fig. 14.
- Quantitative argument for the richness of quasi-psychological connections enabled by attention streams