干预认证放大（ICA）：σ 加权后训练的有效性门槛

View PDF HTML (experimental)

Abstract:Hybrid post-training often mixes supervised fine-tuning (SFT) and reinforcement learning (RL), using training-side signals to set each sample's balance. Such signals characterize the anchor data and policy, but not whether the emphasized behavior helps in the target task and evaluation setting. This mismatch can reactivate and amplify pretrained behavior even after its literal form is removed from supervision. We demonstrate this with DIAA (Dual-Implicit Adversarial Amplification), a per-sample $\sigma$-weighted SFT/RL hybrid. With all literal Python fences removed from its anchor, DIAA emits unexecuted Tool-Integrated Reasoning (TIR) on $29%$ of MATH-500 prompts, three times vanilla GRPO. Paired interventions show $10.6$ percentage points of accuracy recovery when TIR is suppressed and comparable contributions from SFT imitation and $\sigma$-amplification; a student trained on a $0%$-TIR teacher anchor emits it at $45.8%$. Intervention-Certified Amplification (ICA) gates $\sigma$ with paired-rollout target evidence. Frozen and online ICA reduce TIR to $13.8%$ and $11.2%$, halve residual drag, improve accuracy from $47.5%$ to $48.7%$ and $49.3%$, and transfer to held-out GSM8K without retuning.

Comments:	13 pages, 2 figures, 4 tables. Published in Findings of the Association for Computational Linguistics: EMNLP 2026; see this https URL
Subjects:	Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Cite as:	arXiv:2610.33882 [cs.LG]
	(or arXiv:2610.33882v1 [cs.LG] for this version)
	https://doi.org/10.48550/arXiv.2610.33882 arXiv-issued DOI via DataCite (pending registration)
Journal reference:	Findings of the Association for Computational Linguistics: EMNLP 2026, pages 44510-44522, Budapest, Hungary. Association for Computational Linguistics, 2026

Submission history

From: Leiyue Chen [view email]
[v1] Sat, 24 Oct 2026 19:49:59 UTC (47 KB)