Cross-architecture transferability of stability-filtered steering vectors

Determine whether steering vectors constructed using stability filtering and content-subspace projection, extracted from a Qwen-architecture 1.5B model, transfer to language models from more distant architecture families without re-extraction.

Background

The paper extracts steering vectors from DeepSeek-R1-Distill-Qwen-1.5B and demonstrates that these vectors improve two additional Qwen-architecture 1.5B models (Nemotron-Research-Reasoning-1.5B and DeepScaleR-1.5B-Preview) without re-extraction.

Because the tested models share an architecture family and similar training data, the authors caution that it remains unknown whether such transferability holds for more distant model families.

References

Whether this extends to more distant model families remains an open question.

— Reliable Control-Point Selection for Steering Reasoning in Large Language Models  (2604.02113 - Zhuang et al., 2 Apr 2026) in Section 4 (Experiments), Main Results — Cross-model transfer paragraph

Our current implementation uses a separate calibration table, readout layer, and intervention strength for each backbone. We evaluate 11 models from six architecture families, but we calibrate and tune each model separately. Our experiments therefore do not establish whether these components transfer across backbones without additional calibration or development data.

— ReWEIGH the Evidence: Calibrating Token-Level Ordinal Visual Evidence to Mitigate Hallucinations in Large Vision-Language Models  (2608.19075 - Jeong et al., 19 Aug 2026) in Limitations section