Tag
DOG-DPO is a training-free data selection framework that treats preference pairs as structured geometric signals, decomposing multi-dataset preference geometry into anchor and residual subspaces to select diverse subsets for safety alignment. It achieves strong utility-robustness trade-offs using only 11% of preference pairs across six safety benchmarks.
This post extends E8 lattice geometric activation injection to supervised LLM safety routing, using STE-snapped E8 policy heads. While achieving near-perfect routing on clean data, the approach catastrophically fails under adversarial stress, requiring a hybrid symbolic-geometric architecture with audited deterministic rules.
This paper proposes a Polar Probe that linearly recovers semantic structures from LLM activations by representing entity relations through distance and direction in a learned subspace. Testing across arithmetic, visual scenes, family trees, metro maps, and social interactions shows the code emerges in middle layers, generalizes to new entities, and causally influences model predictions.