L2 regularisation with bias term delivers best probe performance; L2 regularisation increases probe selectivity

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

• LLaMA-2-70B and 13B probes generalize better across datasets than 7B probes across all training sets and probe typesfinding0.770
  Larger models linearly represent more general concepts including truth
• The L_retain regularization objective is empirically effective at preserving unrelated model capabilities during cone trainingclaim0.759
  Interpretation of low KL divergence results as validation of the training objective
• Larger models should amplify bias less than smaller models, with model biases more accurately reflecting data biases rather than exacerbating themclaim0.755
  Implication of PRH for AI fairness and bias
• A probe may achieve high performance even on representations that are not causally relevant for the taskclaim0.746
  Key interpretive claim from Case Study II distinguishing probe accuracy from causal relevance
• The target vs. off-target probe area metric quantifies steering selectivity and distinguishes selectively steerable from entangled interventions.claim0.746
  Justification for the novel metric introduced in the paper
• We hypothesize that degraded generalization on benchmarks like MMLU may reflect the computational demands of the tasks.hypothesis0.745
  Connecting the paper's task-difficulty findings to prior observations of weak generalization on complex QA benchmarks.
• Probe-based ranking outperforms gradient-based and LLM-judge methods at lower costclaim0.744
  Authors' claim that their approach is both more effective in reduction and cheaper than prior methods.
• Probes trained on h_b activations achieve perfect test accuracy in every case; h_s probes achieve perfect accuracy in only 0.60% of casesfinding0.744
  Justifies restricting probe-based vector derivation to h_b activations; attributed to Yes/No semantics