claim
active
claim:higher-density-priors-b10-are-more-robust-to-fine-tuning-than-lower-density-ones-b9Higher-density priors (B10) are more robust to fine-tuning than lower-density ones (B9).
Interpretation of cross-base transfer asymmetry.
Source paper
extracted_from(2025) · Edward Yi Chang · Kaya, Zeyneb N. · Ethan Chang
Neighborhood — ranked by edge-count
Communities (3)
community
- Few-shot anchoring & latent structuremembers_ofHow minimal examples disambiguate and recruit latent arithmetic/reasoning interpretations in LLMs
- Investigates how neural network bases (B10 vs B9) differ in fine-tuning stability and cross-base transfer, with density as a key predictor of robustness.
- Empirical finding that base-10 representations transfer more robustly out-of-distribution than base-9 after fine-tuning.
Concepts (1)
concept
- cross-base interferenceassociated_withAsymmetric transfer after fine-tuning: high-density bases (B10) are more robust.
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.
- E2 asymmetric transfer finding consistent with UCCT's mismatch-driven OOD fragility
- UCCT's theoretical prediction about how fine-tuning maps onto the anchoring score
- Strong priors require higher-cohesion anchors to overcome, manifesting as delayed thresholds or reduced transferhypothesis0.775Prediction for Experiment 1 cross-domain anchoring
- Concurrent work result showing emergent misalignment occurs in small models
- Cross-base fine-tuning yields asymmetric transfer: B10 transfers most robustly, B9 leastfinding0.774In-base gains accompanied by uneven OOD drops; higher-density priors more robust.
- Main result from Experiment 3 on effect of fine-tuning on HH-intent.
- Structural consequence of the supervenience chain illustrated by the fly example.
- Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training