Search and filter through extracted claims from AI researchers.
Showing 161-171 of 171 claims in topic "interpretability"
Pre-training generally improves molecular structure awareness of chemical language models, particularly in the upper layers
Randomly initialized chemical language models already encode ring structures well in the first layer, even without pre-training
Fine-tuning affects task-relevant molecular substructures more than others, indicating that representation changes follow chemical theory
Standard language models fail to provide traceable training data influence because their dense network pathways distribute influence across parameters
Prototype language models remain within 2.5 percentage points on average downstream accuracy compared to dense baselines across 130M to 1.6B parameters
PRISM prototype language models can match or nearly match dense baseline performance while providing interpretable training data influence tracking
RMSNorm models require signed-permutation gauge alignment rather than permutation-only alignment for coordinate-indexed object transport
Sign-marginalized Hungarian matching removes the permutation-accuracy ceiling caused by decorrelated coordinates in gauge alignment
Raw signed-correlation matching has a structural permutation-accuracy ceiling at the positive-sign fraction of the true gauge with decorrelated coordinates
Geometric stability predicts trial-by-trial neural-behavioral coupling (ρ=0.18, p=0.005) while centroid drift does not (ρ=0.002, p=0.976) across 229 area-session observations spanning 68 brain regions
Geometric stability (pairwise distance structure reproducibility) is an independent dimension of neural representation analysis that is dissociable from temporal stability and decoding accuracy
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 8,949 pending.