finding
active
finding:lda-barely-outperforms-random-features-across-all-pythia-model-sizes-in-causalgymLDA barely outperforms random features across all pythia model sizes in CausalGym
Surprising negative result for LDA despite being a supervised method
Source paper
extracted_from(2024) · Aryaman Arora · Dan Jurafsky · Christopher Potts
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.
- DAS consistently finds the most causally-efficacious features across all pythia model sizes in CausalGymfinding0.894Main benchmark result showing DAS superiority over probing, diff-in-means, PCA, k-means, LDA, and random
- CausalGym results may differ on models trained on different data or in different orders beyond the pythia seriesquestion0.813Identified limitation about generalizability across model training regimes
- Attributed to model anisotropy from saturation making hidden states harder to access
- DAS achieves overall odds-ratio of 10.24 on pythia-410m averaged across all CausalGym tasksfinding0.783Numerical result for pythia-410m
- Supervised metrics discriminate.
- Smaller models' weaker concept representations explain why supervised metrics differ on Pythia-160Mclaim0.772Interpretation of model-scale effects.
- Task accuracy on CausalGym increases consistently with model scale from 0.62 (14M) to 0.89 (6.9B)finding0.755Scaling result showing larger pythia models perform better on CausalGym linguistic tasks
- SAEs on random Pythia-1B achieve higher TPP than trained-model SAEs, attributed to lower probe baselinefinding0.751Caveated result.