finding

active

finding:cross-base-fine-tuning-yields-asymmetric-transfer-b10-transfers-most-robustly-b9-least

Cross-base fine-tuning yields asymmetric transfer: B10 transfers most robustly, B9 least

In-base gains accompanied by uneven OOD drops; higher-density priors more robust.

Source paper

extracted_from

(2025) · Edward Yi Chang · Kaya, Zeyneb N. · Ethan Chang

claim

community

Few-shot anchoring & latent structure
members_of
How minimal examples disambiguate and recruit latent arithmetic/reasoning interpretations in LLMs
Model base robustness and transfer learning asymmetries
members_of
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.
Cross-base fine-tuning transfer asymmetry
members_of
Empirical finding that base-10 representations transfer more robustly out-of-distribution than base-9 after fine-tuning.

question

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Cross-base transfer: B10 transfers most robustlyfinding0.865
B10 fine-tuning yields smallest OOD drops when transferring to other bases
Higher-density priors (B10) are more robust to cross-base OOD drops than lower-density ones (B9) after fine-tuningfinding0.793
E2 asymmetric transfer finding consistent with UCCT's mismatch-driven OOD fragility
Higher-density priors (B10) are more robust to fine-tuning than lower-density ones (B9).claim0.774
Interpretation of cross-base transfer asymmetry.
Cross-base transferconcept0.753
Transfer of arithmetic rules from one numeral base to another after SFT.
An intervention benign at context v4<0.75 produces a class-C behavioral flip at 0.75<v4<1, demonstrating dormant behavioral changes from latent divergencefinding0.746
Synthetic example showing an intervention that appears safe in tested contexts but causes behavior changes in others
Hypothesis: Fine-tuning reduces mismatch dr between prior and targethypothesis0.745
UCCT's theoretical prediction about how fine-tuning maps onto the anchoring score
Fine-Tuning via Reinforcement Learningmethod0.745
Technique used to impose guardrails on base LLMs, analogized to censorship on the simulator's range of simulacra
Fine-tuning reduces dr; retrieval increases effective ρd; few-shot k trades budget against bothhypothesis0.743
UCCT's unified view of adaptation methods