A rhetorical capability was present in every base model and only became steerable with instruction tuning
AspireAI community · arXiv
Language models absorb rhetorical strategies from pretraining data, including reasoning by analogy, and this paper asks whether that capability is geometrically accessible before instruction tuning and what determines whether it can be activated. The authors extract an Analogy Vector from the residual stream by difference-in-means and show it enables steerable explanation style across seven instruction-tuned models (Gemma-2, Llama-3.1, Qwen-2.5; 2B to 32B), improving harmonic mean from 3.22 to 3.94 on 300 held-out prompts for Gemma-2-9B-IT, with a small human study confirming the judge. The central finding comes from seven matched base-instruct pairs: the analogy direction exists in every base model, but its behavioural accessibility under linear intervention scales continuously with instruction exposure across the training pipeline. Gemma-2 base models show flat steering landscapes - present but behaviourally inert - while Qwen-2.5 base models, whose instruction data was woven into pretraining, are already partially accessible, with base-instruct cosine similarity of 0.94-0.97 for Qwen against 0.37-0.66 for Gemma. That challenges the binary base-versus-instruct framing of what steering can reach.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.