SFT Leaves a Reproducible Signature in Activation Geometry: Calibrating Post-Training Interpretability Against Run-to-Run Noise

Key points

• A model diff credits post-training with whatever separates a base checkpoint from one post-trained descendant, presuming a rerun would land close by.1
• On the sparse co-activation graph of Qwen3-1.7B it does not: two LoRA-SFT reruns sit over half as far apart as either sits from the base.1
• We score each base-to-post contrast against the run-to-run band of its own recipe, and show that $k$ runs certify no better than $p=1/(k+1)$ under run exchangeability, however large the effect, so the band $z$-score is a descriptive scale.1
• A pre-registered causal sweep at three edge budgets on 4,816 held-out prompts found the displaced sources load-bearing ($12\times$ a matched random control in mean, $3.5\times$ in median at 10,000 edges), but their stage-specificity interaction reversed at two budgets and the third failed its faithfulness gate, so we claim calibrated structural displacement, not a changed mechanism.1

Sources (1)

• [1]SFT Leaves a Reproducible Signature in Activation Geometry: Calibrating Post-Training Interpretability Against Run-to-Run Noise
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 04:50 PM

  SFT leaves a large, reproducible signature: base $\to$ SFT displaces the graph by 17.8 pairwise-distance SDs beyond the mean distance between four independent LoRA-SFT runs, and the pre-registered estimator replicates on Llama-3.1-8B with Llama-Scope.

  A model diff credits post-training with whatever separates a base checkpoint from one post-trained descendant, presuming a rerun would land close by.

Extractive summary: sentences quoted from the sources.