Answers are grounded in this paper and cite their sources.
Tests reasoning transfer in language models, suggesting limitations in capability transfer between architectures.
We test whether a contrastive activation difference computed at a single layer of a reasoning-distilled donor can be injected into a generally instruction-tuned host of the same architecture to transfer mathematical reasoning capability. Donor: DeepSeek-R1-Distill-Llama-8B. Host: Llama-3.1-8B-Instruct. Both share Llama-3 architecture and tokenizer, removing common cross-family confounds. We inject the contrastive vector at layer 12 with alpha in {0, 0.1, 0.3, 0.5} and evaluate on 10 held-out GSM8K problems. Baseline accuracy is 65.0%. Injection produces a monotonically decreasing accuracy: 50.0% at alpha=0.1, 30.0% at alpha=0.3, and 10.0% at alpha=0.5 (Delta = -55 percentage points relative to baseline). No cell achieves raw p<0.05 under paired exact McNemar testing; closest is p=0.0625 at alpha=0.5, which does not survive Bonferroni correction. We conclude that naive last-token contrastive subtraction does not encode a transferable reasoning capability in this single-layer setting. Reasoning-marker rate stays stable (2.00 to 2.30) across all injection cells while accuracy collapses, indicating the contrastive vector captures surface mathematical vocabulary rather than transferable computational structure. We discuss three structural hypotheses (reasoning is not a single direction; wrong layer/geometry; donor-host capability gap too small) and propose four concrete follow-up protocols. All code, calibration prompts, test prompts, per-cell generation outputs, and aggregation results are released open-source for reproduction.
No takes yet. Share an insight, caveat, or question.
Ho Yiing Chen (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: