arXiv：cs.AI· Xupeng Chen, Chenqian Le, Jiayi Cheng, Runhao Li, Michael Yue·· 6 小时前
Qwen3-1.7B 研究：SFT 在激活几何中留下可复现签名，提示模型 diff 需校准运行间噪声
SFT Leaves a Reproducible Signature in Activation Geometry: Calibrating Post-Training Interpretability Against Run-to-Run Noise
AI 导读
在 Qwen3-1.7B 的稀疏共激活图上，两次 LoRA-SFT 重跑之间的距离超过其与基座距离的一半，说明模型 diff 需按各自配方的运行间波动带校准。base→SFT 的图位移达四个独立 LoRA-SFT 运行间平均距离的 17.8 个成对距离标准差，该预注册估计器在 Llama-3.1-8B 上借助 Llama-Scope 复现。
正文
Abstract:A model diff credits post-training with whatever separates a base checkpoint from one post-trained descendant, presuming a rerun would land close by. 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. 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. 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. At 29 steps each, four SFT and four DPO runs occupy disjoint displacement ranges, a gap $5.0\times$ and $5.6\times$ the spreads of the two arms, without any band. All bands use graphs built on one GPU architecture. Three other readings stop short of a verdict: DPO band counts (3/4, 1/4, 3/4 contrasts outside at 12, 29, 60 steps) show no consistent dose ordering while its displacement stays flat; light GRPO is a secondary non-detection, not an absence, and we run no equivalence test; and the graph concentration read as an RL signature is already in the zero-RL base. 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. Scope: one backbone, one SFT recipe and general-text probe states for the full protocol.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.17093 [cs.AI] |
| (or arXiv:2610.17093v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.17093 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Chenqian Le [view email]
[v1]
Mon, 12 Oct 2026 16:50:45 UTC (592 KB)
来源：arXiv：cs.AI · arxiv.org